Drug discovery has a strange money problem. Scientists can spend years studying a disease, testing thousands of possible treatments, and building a drug that looks promising in the laboratory. Yet after hundreds of millions of dollars and years of work, that drug can still fail when it reaches patients.
Artificial intelligence promises to change this expensive process. Instead of testing every idea in a laboratory, researchers can use AI to predict which ideas deserve attention. Instead of making hundreds of molecules, they may be able to make a smaller number of better ones. And instead of discovering problems late in development, companies hope to identify them before they become expensive failures.
Boston is becoming one of the most important places to test whether those promises translate into real savings.
Across Cambridge, Somerville, Boston, and the wider Massachusetts life sciences community, companies are combining AI with protein design, drug research, laboratory automation, and clinical trials. They are asking computers to help solve problems that once required far more trial and error.
But there is a critical difference between making one part of drug discovery cheaper and making an entire new medicine cheaper to develop.
A company might cut the number of laboratory experiments by 80% while spending more money on computing power, specialized scientists, and clinical studies. Another company might produce drug candidates twice as quickly but see little improvement in the chances that those candidates eventually reach patients.
These possibilities matter because the largest financial burden in drug development often comes after a promising molecule has already been discovered.
A 2024 study published in JAMA Network Open found that clinical development accounted for approximately 68% of the average direct development cost per drug candidate in its analysis. When researchers included the cost of unsuccessful projects and the cost of capital, the share attributable to clinical development changed significantly.
JAMA Network
That finding creates an important question for Boston’s AI-biotech companies: Can AI reduce the total cost of producing successful medicines, or is it mainly making the earliest stages of research more efficient?
To explore this, BostonBizScene examined publicly available research, clinical success rates, corporate financial reports, and company disclosures. We also developed original cost models to estimate how different improvements in research efficiency could affect overall development spending.
Our analysis points toward a more useful way to understand the AI-biotech opportunity. The greatest financial value may not come from generating more drug candidates. It may come from using better information to avoid investing in candidates that are unlikely to work.
The Short Version: AI Can Make Drug Discovery Cheaper, but the Biggest Savings Depend on What Happens After Discovery
AI is already showing signs that it can reduce certain kinds of research work. Computers can examine large collections of biological information, suggest new molecules, help predict their properties, and guide scientists toward experiments that are more likely to provide useful answers.
Some companies have reported substantial reductions in the number of molecules they need to make before selecting a development candidate. Others have shown that AI-designed drugs can progress into human studies.
However, the available evidence does not yet establish that AI-driven drug discovery consistently reduces the full cost of bringing an approved medicine to market.
A 2024 analysis in Drug Discovery Today reported that AI-discovered molecules had Phase 1 clinical success rates of roughly 80% to 90% in the programs examined. Phase 2 success was approximately 40%, based on a limited sample. The researchers described these findings as early evidence, rather than proof that AI programs would ultimately achieve better approval rates.
PubMed
That distinction is especially important because Phase 2 studies begin testing whether a treatment actually helps patients with a disease. An AI system may be excellent at designing a molecule that enters the human body safely without being equally good at predicting whether the drug will improve patient health.
BostonBizScene Original Analysis: How Much Could AI Really Save?
Illustrative cost model based on the 68% clinical spending share reported in JAMA Network Open, with the remaining 32% grouped as nonclinical spending. Savings assumptions are hypothetical, not observed AI outcomes.
Savings are percentages of original total direct development cost, not guaranteed reductions.
Source: 2024 JAMA Network Open cost study. Calculations and scenario definitions: BostonBizScene, October 2026.
Our model suggests that even a dramatic reduction in nonclinical development spending would have a limited effect on total direct costs if clinical spending remained unchanged. This is not a prediction of AI’s actual savings. It is a way to show why the location of those savings matters.
For Boston’s drug developers, the lesson is practical. A faster discovery engine can be valuable, but its financial impact becomes far more meaningful when it also produces better clinical decisions, fewer failed trials, or lower spending on work that does not help patients.
Why Drug Discovery Is So Expensive in the First Place
Before examining what AI can change, we need to understand where pharmaceutical companies spend their money.
Drug development is not one activity. It is a long series of decisions, experiments, safety checks, manufacturing challenges, and human studies. Each stage requires different skills, equipment, and funding.
The process starts with understanding a disease and identifying a possible treatment. Scientists then search for or design molecules that could affect the disease, test whether those molecules behave as expected, and determine whether they appear safe enough for further development.
Only a small portion of the original ideas make it into clinical trials.
Discovery Is About Finding Something Worth Testing
In early discovery, researchers need to answer two basic questions. First, what biological process is contributing to the disease? Second, can a medicine change that process in a useful way?
The answers are rarely obvious. A disease may involve many genes, proteins, cell types, and chemical signals. Even when scientists identify something important, they still need to find a molecule that can interact with it safely.
This is where AI can help.
An AI system can compare large amounts of research data, identify patterns, and rank possible drug candidates. Scientists may use those rankings to choose a smaller group of experiments instead of examining every possibility through physical testing.
The potential saving comes from reducing unnecessary work. A company may still need sophisticated laboratories, but it could use their time more efficiently.
Preclinical Development Is About Finding Problems Before Patients Are Involved
After scientists find a promising molecule, they must test it more carefully.
They need to understand how it behaves in biological systems, whether it causes harmful effects, how much of it may be needed, and whether it can be manufactured with consistent quality.
A molecule can look excellent in an early test and still have problems that make it unsuitable as a medicine.
AI may help predict some of these problems before expensive experiments begin. However, a prediction is not a replacement for the physical evidence needed to support human testing.
The financial value comes from finding unsuitable candidates earlier, before additional money is committed.
Clinical Development Is Where Costs Can Escalate
Human testing introduces expenses that cannot be solved simply by making molecules faster.
Companies must recruit patients, organize clinical sites, monitor safety, collect medical data, manufacture trial supplies, manage outside research partners, and satisfy regulatory requirements.
Later trials can be particularly expensive because they often involve more patients, more locations, and longer periods of observation.
Research published in 2025, based on 631 clinical trial projects conducted by Japanese pharmaceutical companies, found median study costs of 859 million yen for Phase 1, 1.739 billion yen for Phase 2, and 4.183 billion yen for Phase 3. These figures came from trials conducted during 2012–2022 and show how spending can grow as programs progress.
PMC
Those amounts are not universal prices for clinical trials, and they should not be treated as direct estimates for a new Boston startup. They show the size and shape of the problem that developers need to manage.
Drug Failure Makes Successful Medicines Much More Expensive
A company does not pay only for the medicine that succeeds.
It also pays for research programs that fail, clinical studies that deliver disappointing results, and years of work on products that never reach the market. Those unsuccessful investments are part of the economics of producing successful medicines.
This is why a relatively small improvement in the likelihood of success could be worth more than a major improvement in laboratory speed.
Imagine two biotech companies that each produce ten clinical drug candidates. One company spends less on discovering those candidates but sees most fail. The other spends slightly more on discovery but identifies problems early enough to avoid several expensive clinical failures.
The second company could end up producing successful medicines at a lower overall cost, even though its initial research was more expensive.
That is the economic problem AI needs to solve.
What the Public Research Actually Says About Drug Development Costs
Drug development cost estimates vary widely, and for good reasons. Different studies examine different sets of drugs, use different time periods, and define costs differently.
Some report direct spending. Others include the money spent on failed projects. Still others add the cost of capital, which represents the financial cost of tying up investment for many years.
These figures cannot be treated as interchangeable.
The Difference Between Cash Spending and the Full Economic Cost
A startup’s actual spending is the money leaving its bank account.
The cost of producing a successful drug is a broader measure. It can include spending on failed programs, investments that were abandoned, and financial costs related to time and uncertainty.
Consider a company that spends $40 million developing one candidate. If that candidate fails, the $40 million is still a real cost, even though it produces no approved treatment.
If several candidates fail before one succeeds, the economics of the successful medicine depend on the whole group of investments.
This creates three useful ways to examine AI’s impact.
| Cost measure | What it tells us | Main limitation |
|---|---|---|
| Cost per experiment | Whether research work becomes more efficient | Does not show total program savings |
| Cost per clinical candidate | How efficiently a company reaches human testing | Does not account for later failures |
| Cost per approved medicine | Whether the entire development process becomes more productive | Requires years of outcomes and a full accounting of failures |
For investors and founders, the third measure is ultimately the most important. However, it is also the hardest to measure because many AI-designed medicines have not yet completed the long approval process.
Research Finding: Clinical Work Represents a Large Share of Direct Costs
The 2024 JAMA Network Open cost study found that clinical work represented about 68% of mean direct development costs per candidate. When failure costs were included, the share was approximately 60.5%, and after including the cost of capital, it was approximately 53%.
PMC
These numbers reveal something important about the challenge AI faces.
The largest opportunity does not necessarily sit in the work computers can automate most easily. Many high-cost activities involve patients, safety, physical manufacturing, and regulatory review.
That does not make AI less valuable. It means companies need to be precise about which costs their systems can actually reduce.
Research Finding: Phase 2 Remains a Major Barrier
The Biotechnology Innovation Organization, together with Informa Pharma Intelligence and QLS Advisors, studied clinical development success using programs tracked between 2011 and 2020.
The analysis covered 9,704 development programs and found an overall 7.9% likelihood that a drug entering Phase 1 would ultimately receive approval. It also found that only 28.9% of Phase 2 candidates advanced to the next stage.
BIO
+1
This gives AI developers a useful target.
Improving the design of molecules is important, but improving the ability to predict what will happen in patients could create far greater financial value.
The most valuable AI system may be the one that helps scientists decide not to proceed with an attractive but weak program.
Original Finding #1: Cutting Discovery Costs in Half Does Not Cut Total Development Costs in Half
The first part of BostonBizScene’s original analysis examines a simple but often overlooked mathematical problem.
If AI makes early drug research dramatically cheaper, how much money does that actually save across the entire development process?
To answer this, we created a cost model using the average direct-cost distribution reported in the 2024 JAMA Network Open study. We divided spending into two broad categories: clinical development and everything else.
The original research reported a clinical share of 68%. For our analysis, the remaining 32% represents nonclinical development and other included costs.
This is a simplified allocation. It does not mean that discovery alone accounts for 32% of total spending, because that category also includes other activities outside clinical development.
Our Methodology
We started with an illustrative development budget of $100 million. We allocated $68 million to clinical spending and $32 million to the remaining activities.
We then reduced the nonclinical portion by different percentages while keeping clinical spending unchanged.
This approach isolates one question: what happens if AI lowers costs before clinical development, but does not change the cost of running human trials?
| AI efficiency scenario | Nonclinical spending | Clinical spending | Total spending | Overall saving |
|---|---|---|---|---|
| No improvement | $32M | $68M | $100M | 0% |
| 10% nonclinical reduction | $28.8M | $68M | $96.8M | 3.2% |
| 25% nonclinical reduction | $24M | $68M | $92M | 8% |
| 50% nonclinical reduction | $16M | $68M | $84M | 16% |
| 75% nonclinical reduction | $8M | $68M | $76M | 24% |
BostonBizScene scenario analysis, October 2026. Dollar amounts are illustrative, not actual company spending. The cost split is drawn from the 2024 JAMA study; the saving assumptions are hypothetical.
What the Numbers Tell Us
If a company cuts nonclinical spending by 50%, total development spending falls by 16% in this model.
That is a meaningful improvement. A 16% reduction in a large development budget could save substantial amounts of money and extend the period a startup can operate before raising additional funding.
However, the remaining 84% of the original budget must still be financed.
And this is an optimistic way to interpret improvements in discovery efficiency. If the AI system only reduces a small portion of discovery spending rather than the entire nonclinical category, the effect on total costs would be smaller.
Why This Matters for Boston Startups
Many AI-biotech pitches emphasize how much faster computers can produce candidate molecules.
Faster design is useful, but it is not enough to establish overall cost savings. Investors should ask what share of the actual development budget becomes smaller and whether expenses are being reduced, delayed, or moved into another category.
A startup may save money on chemical testing while spending more on computing infrastructure. It might reduce early laboratory costs while advancing more candidates into expensive studies.
The question is not whether an activity becomes faster. It is whether the business can produce better development outcomes using less capital.
Original Finding #2: Combining Small Improvements Across Research and Clinical Development Has a Larger Effect
The second part of our model examines a different strategy.
Instead of concentrating all efficiency gains in early research, what happens when a company improves both nonclinical work and clinical execution?
We kept the same starting allocation of $32 million in nonclinical spending and $68 million in clinical spending. We then applied different reductions to each category.
How Combined Savings Change the Economics
Illustrative direct-cost model. Clinical and nonclinical reductions are assumed independently. Does not include changes in success rates, the cost of capital, or AI implementation expenses.
The model produces an important result.
A company that reduces nonclinical spending by 20% and clinical spending by 10% achieves a 13.2% reduction in total costs. By comparison, reducing nonclinical spending by 20% alone produces a 6.4% overall reduction.
The additional improvement comes from changing a larger part of the cost structure.
Why Clinical AI Could Matter More Than Discovery AI
Clinical trials contain many tasks where better information could reduce wasted time and money.
Companies need to identify suitable patients, select clinical locations, predict enrollment needs, monitor study progress, and detect data problems. Some of these activities could become more efficient with better forecasting and software.
AI might help researchers recognize that certain patients are unlikely to respond to a treatment. It might also support more accurate planning, helping study teams avoid opening unnecessary locations or making late changes to a trial.
These improvements would need to be demonstrated with actual outcomes.
An AI system that suggests a better trial design has not saved money until the design reduces real costs, avoids delays, improves useful evidence, or prevents an unnecessary study.
The Value of Avoiding a Bad Trial
Consider a company deciding whether to move a candidate into Phase 2.
It has promising laboratory results, but important questions remain about whether the treatment will work in the intended patients. The company could begin the clinical study immediately, or it could spend additional money gathering evidence first.
If the extra research identifies a major weakness and the company cancels a study that would have failed, that decision may preserve substantial capital.
But if the company repeatedly delays promising treatments for tests that provide little useful information, it may spend more money and lose valuable time.
The business problem is deciding when another experiment is worth its cost.
This is one of the strongest possible uses for AI: estimating the value of gathering additional information before making an expensive commitment.
Original Finding #3: Boston’s AI-Biotech Companies Are Demonstrating Different Kinds of Efficiency
Not all AI-biotech companies are trying to save money in the same way.
Some are redesigning how molecules are created. Others are building systems that perform laboratory experiments at greater scale. Some focus on identifying the right biological targets, while others seek to improve clinical development.
Boston’s ecosystem is useful because it contains companies working on several parts of this problem.
Our comparison below uses publicly described platforms and reported milestones. It is a qualitative framework, not a financial ranking. Most private platforms do not publicly disclose the complete costs needed to make a fair comparison.
| Company | Approach | Possible source of savings | What still needs proof |
|---|---|---|---|
| Generate:Biomedicines | AI-based protein medicine design | More focused design and testing | Lower lifetime cost per successful medicine |
| Relay Therapeutics | Computational modeling of protein movement | Better candidate selection | Savings attributable to the platform |
| Manifold Bio | Large-scale testing of designed proteins in biological systems | More information per experiment | Lower cost per validated drug program |
| Lila Sciences | AI combined with automated scientific laboratories | Faster, repeatable experiments | Lower fully loaded cost per useful discovery |
| Recursion (external benchmark) | AI, data generation, drug design, and clinical tools | Fewer compounds and better decisions | Reliable gains in approval rates and end-to-end cost |
Research comparison based on company disclosures through October 8, 2026. Recursion is headquartered in Salt Lake City and is included as an external benchmark, not a Boston company. Potential savings are hypotheses unless independently measured.
Generate:Biomedicines: Better Protein Design Does Not Eliminate Clinical Costs
Generate:Biomedicines is one of the clearest examples of the movement from AI-enabled research into AI-native drug development.
The company uses generative models to design proteins that could become medicines. Its approach is based on the idea that scientists can learn enough about protein structure and behavior to create useful molecules more systematically.
Rather than depending only on finding suitable molecules in nature or modifying previously discovered proteins, AI can help explore designs based on desired properties.
This could reduce some of the trial-and-error work required to identify promising candidates.
But the company’s progress also illustrates what happens when AI-designed drugs advance into expensive human studies.
The Clinical Spending Reality
Generate reported $64.3 million in research and development expenses for the second quarter of 2026, compared with $59.7 million in the same quarter of 2025.
The company attributed the increase mainly to ongoing investment in its Phase 3 asthma trials for GB-0895, partly offset by lower external discovery and other program costs. It reported approximately $457.4 million in cash, cash equivalents, and marketable securities at June 30, 2026.
SEC
This is an important real-world example.
An AI-driven company can reduce certain discovery expenses while its total research and development spending rises because successful programs are moving into costly clinical stages.
That is not necessarily a bad outcome. Advancing a promising drug into Phase 3 is a meaningful step toward commercial value.
However, it shows why total R&D spending cannot be expected to decline continuously as an AI-biotech company matures.
What Founders Should Learn
A platform should be evaluated according to where it sits in the drug development process.
During discovery, useful measures might include the cost per tested protein, the share of designed proteins that meet laboratory goals, and the time required to produce a candidate.
During clinical development, the measures need to change. Patient enrollment, study duration, spending per patient, safety outcomes, and the quality of evidence become more important.
Companies should not keep selling investors on early discovery speed after clinical spending becomes their main financial challenge.
Relay Therapeutics: A More Revealing Look at How Research Budgets Change
Relay Therapeutics is based in Cambridge and uses computational approaches to understand how proteins move and behave.
Its scientific approach is different from simply generating large numbers of new molecules. By studying how a protein changes shape, researchers hope to identify treatments that interact with important parts of its behavior.
Better information about those interactions could improve candidate selection and reduce some unnecessary experiments.
Relay also provides something particularly valuable for financial analysis: detailed research and development spending disclosures.
BostonBizScene Original Analysis: Relay’s Changing Cost Structure
We examined the research and development expense categories in Relay’s 2025 annual filing and compared them with 2024.
| R&D expense category | 2024 | 2025 | Change |
|---|---|---|---|
| External clinical program costs | $92.1M | $104.3M | +13.2% |
| External platform and preclinical costs | $76.4M | $38.5M | −49.6% |
| Employee-related expenses | $123.6M | $95.6M | −22.7% |
| Other R&D expenses | $27.0M | $23.0M | −14.8% |
| Total R&D expenses | $319.1M | $261.4M | −18.1% |
Source: Relay Therapeutics 2025 Form 10-K, fiscal years ended December 31. Amounts rounded to $0.1 million; changes calculated from the original reported figures.
SEC
Relay’s R&D Spending Shift, 2024–2025
■ 2024
■ 2025
Company-reported expenses. Differences reflect many operating and portfolio decisions and cannot be assigned to AI.
These numbers deserve careful interpretation.
Relay’s external platform and preclinical spending fell by nearly half, while external clinical program spending increased. Its total R&D expenses declined by approximately $57.7 million.
However, the company’s management explained that its overall expense reduction primarily reflected organizational changes, spending decisions, and changes in development responsibilities for specific programs.
Relay Therapeutics
We cannot claim that AI caused these savings.
What the data does demonstrate is how a computationally focused biotech company’s expense structure can change as research priorities and clinical investments shift.
Original Finding: Clinical Spending Became a Larger Part of Relay’s Research Budget
Using the reported figures, we calculated that external clinical program costs accounted for approximately 28.9% of total R&D expenses in 2024 and 39.9% in 2025.
2024 clinical share of R&D
28.9%
2025 clinical share of R&D
39.9%
External clinical program expense divided by total R&D expense for each year. BostonBizScene calculations from Relay’s SEC filing.
That is an increase of approximately 11 percentage points in one year.
This does not establish that Relay’s platform lowered discovery costs. The comparison is affected by restructuring, portfolio changes, and other business decisions.
It does illustrate a broader issue that AI-biotech investors should expect.
As more candidates progress toward patients, clinical activities can take up a growing share of spending. A company may become more efficient at discovering drugs without becoming a low-cost business overall.
Manifold Bio: The Value of Getting More Information From Each Experiment
Manifold Bio is approaching the research problem from another direction.
One challenge with AI-generated molecules is that computers can propose more designs than a traditional laboratory can realistically evaluate. If scientists can generate a million ideas but physically test only a few hundred, laboratory testing becomes the bottleneck.
Manifold Bio is working on technologies that let researchers test large groups of protein designs and collect biological measurements at scale.
In March 2026, the company reported that it had worked with NVIDIA to test one million protein binder designs against 127 targets in a single large-scale experiment, measuring more than 100 million protein interactions. The company reported identifying specific binders for 68% of the targets tested.
Manifold Bio News
This is an impressive example of research scale, but the financial interpretation needs care.
Testing many designs in one experiment does not automatically prove that each successful drug candidate becomes cheaper. The experiment may require specialized equipment, complex preparation, expensive materials, and significant data analysis.
The Metric That Matters
A useful measure is the cost per validated, decision-ready candidate.
A decision-ready candidate is one supported by enough evidence for a company to make its next development choice. It is different from a molecule that simply shows activity in a single test.
Suppose one laboratory workflow costs $100,000 and produces ten candidates that pass a defined validation standard. Its cost per validated candidate is $10,000.
Now imagine another workflow costs $150,000 but produces thirty candidates meeting the same standard. Its cost per validated candidate is $5,000.
The second workflow costs more in total, but it is more efficient per useful result.
These amounts are hypothetical. They illustrate the financial measure that should be used when comparing different laboratory systems.
For Manifold and similar Boston companies, the long-term opportunity is not simply producing more experimental data. It is improving how much reliable decision-making information researchers obtain for each dollar spent.
Lila Sciences: Laboratory Automation May Change the Cost of Learning
Lila Sciences represents another direction in Boston’s AI ecosystem.
The company combines advanced AI models with automated scientific laboratories. Its systems are intended to help researchers form ideas, design experiments, run tests, and use the results to guide further experiments.
Flagship Pioneering introduced Lila publicly in March 2025, announcing $200 million in committed seed funding. Later that year, Lila announced the closing of a $350 million Series A, bringing total capital raised to $550 million.
Flagship Pioneering
+1
This level of investment makes an important point.
Building an AI-driven laboratory system can be extremely capital-intensive. Automation may reduce the cost of repeated experimental work after the infrastructure is established, but the initial investment can be substantial.
Automation Makes Economic Sense When Equipment Is Used Well
A robot does not create savings simply by replacing a manual action.
The full financial equation includes the equipment purchase, installation, maintenance, software, operating staff, consumables, and the cost of keeping the system running.
If a company buys an expensive automated platform but uses it only occasionally, the cost per experiment may remain high.
If the same platform runs useful experiments consistently, with fewer errors and less waiting time, the economics could improve considerably.
This creates a business model based on utilization, or how much productive work expensive infrastructure performs.
Boston’s opportunity may be to build shared or highly utilized research systems that serve multiple scientific programs, rather than requiring every startup to own its own complete automated laboratory.
That could make advanced experimental capabilities available to smaller companies that cannot justify the upfront investment.
Original Finding #4: Fewer Synthesized Molecules Can Be a Powerful Efficiency Signal, but Not Proof of Lower Drug Costs
One of the most concrete public claims about AI-driven discovery efficiency comes from Recursion Pharmaceuticals, which is headquartered outside Massachusetts.
We included Recursion in our research because its 2025 annual filing provides a specific comparison between the number of molecules its system synthesizes and an industry benchmark.
Recursion reported that its drug discovery approach uses approximately 330 synthesized molecules per advanced candidate, compared with an industry figure of about 2,500 molecules.
SEC
Using those disclosed figures, we calculated the difference.
Molecules Synthesized per Advanced Candidate
86.8%
Fewer molecules synthesized
7.6×
Ratio of benchmark to reported count
BostonBizScene calculations from the comparison disclosed in Recursion’s 2025 Form 10-K. The industry benchmark is company-cited, not an independently matched control group.
What This Comparison Really Means
The arithmetic is straightforward.
Reducing the number of molecules from 2,500 to 330 represents 2,170 fewer molecules, or approximately an 86.8% reduction.
That is a meaningful indicator of improved research efficiency if the candidates being compared meet similar quality standards.
But several important questions remain unanswered.
We do not know whether each molecule had the same production cost. We do not know whether the programs had comparable difficulty. We cannot determine from these figures alone how many physical experiments were performed, how much computing was required, or how often the resulting candidates succeeded in humans.
The comparison also comes from the company’s own disclosure.
It would be incorrect to interpret an 86.8% reduction in molecule count as an 86.8% reduction in drug development costs.
A More Useful Cost Translation
Consider a hypothetical research program that spends $2 million on making and testing compounds.
If compound-related costs fell by 86.8%, and every other expense remained unchanged, that portion of spending would fall to approximately $264,000.
The saving would be about $1.74 million.
However, suppose the full drug development budget were $100 million. That same $1.74 million reduction would represent only about 1.74% of the original total.
The benefit could still be very important for an early-stage company. Saving nearly $2 million may allow a startup to complete additional experiments without raising money.
But the overall impact depends on where the cost reduction occurs and what other expenses are required to achieve it.
Why Boston Should Pay Attention
Boston founders building molecule design platforms should measure both scientific and financial productivity.
Scientific productivity asks whether the system identifies promising molecules.
Financial productivity asks how much the company spends to reach a decision, whether that decision is reliable, and whether the resulting candidates perform better in later development.
A strong platform should eventually improve both.
Original Finding #5: Improving the Chance of Success Could Matter More Than Cutting the Cost of Each Experiment
Our next analysis examines something potentially more valuable than laboratory efficiency.
What happens to drug development economics when more candidates succeed?
To study this, we built a simple portfolio model showing how the cost per successful medicine changes as the probability of success increases.
The model is deliberately separate from our earlier expense calculations. Rather than estimating the cost of an individual development activity, it examines how many development programs must be financed, on average, to produce one success.
The Starting Assumptions
We assume that each candidate entering clinical development has the same expected total development spending of $100 million, measured across the paths programs may take. This is a modeling assumption, not an estimate of actual spending on a Phase 1 candidate.
We also assume that candidates have independent outcomes and that each candidate’s probability of success can be represented by one fixed number.
The starting approval probability is 7.9%, drawn from BIO’s historical study of programs entering Phase 1.
For the sensitivity analysis, we then test higher hypothetical success probabilities of 10%, 12%, and 15%.
These higher probabilities are not reported AI clinical approval rates. They are scenarios designed to estimate what improved candidate quality might be worth.
Our Portfolio Model
The model uses the following equation:
\[ \text{Expected cost per approval} = \frac{\text{Expected spending per candidate}} {\text{Probability of approval}} \]
This simplified equation assumes that expected spending per candidate remains unchanged. It excludes the time value of money and does not model stage-by-stage spending.
| Hypothetical approval probability | Expected candidates per approval | Expected spending per approval | Reduction from baseline |
|---|---|---|---|
| 7.9% historical reference | 12.66 | $1.266B | — |
| 10% | 10.00 | $1.000B | 21.0% |
| 12% | 8.33 | $833M | 34.2% |
| 15% | 6.67 | $667M | 47.3% |
BostonBizScene portfolio sensitivity model, October 2026. The historical 7.9% figure is from BIO’s 2011–2020 analysis; all higher success probabilities and the $100M expected per-candidate spending assumption are hypothetical. These are not estimates of real industry costs per approval.
How Higher Approval Rates Change Expected Portfolio Costs
Illustrative expected cost per approval. Approval probabilities after the 7.9% reference are hypothetical, not observed AI outcomes.
Why the Improvement Looks So Large
When the probability of success rises, a company needs fewer attempts, on average, to produce one successful treatment.
At a 7.9% success probability, the expected number of candidates needed per approval is approximately 12.7.
At 12%, that figure falls to about 8.3.
This is why better candidate selection could change the economics of the pharmaceutical industry more than faster molecule generation alone.
However, the model does not tell us whether AI can achieve these higher success rates. It only measures their possible financial value.
The Main Limitation: Success Rates Are Not Enough by Themselves
Our model assumes that each candidate has the same expected spending, regardless of its probability of success.
Real drug development is more complicated.
Most failed candidates do not consume the same total funding as candidates that survive every clinical stage. A drug that fails in Phase 1 usually costs less than one that fails near the end of Phase 3.
This means improved success rates can change the number of programs reaching expensive later stages. They may also change total portfolio spending and the time needed to generate useful evidence.
A complete model must account for when each candidate fails, not just whether it eventually succeeds.
The 2026 Nature Reviews Drug Discovery discussion of AI economics also emphasizes the importance of later clinical outcomes, particularly Phase 2, when evaluating the potential effect of AI on development costs.
Nature Reviews Drug Discovery
For Boston investors, the practical lesson is clear.
Do not assign a higher approval probability to an AI-designed drug simply because it was designed by AI. Require evidence that the company’s candidates have better outcomes than comparable programs.
Original Finding #6: The Most Valuable Prediction May Be Knowing When to Stop
Drug discovery companies often celebrate the creation of new candidates. But one of the biggest potential financial benefits of AI is helping researchers reject weak candidates early.
This requires a different way of thinking about productivity.
A company that produces fifty promising molecules has not necessarily created more value than one that produces five. The first company may simply have more candidates that still need expensive testing.
The more important question is how accurately each company can identify candidates worth further investment.
A Simple Example of Early Failure Detection
Imagine that a biotech company has ten programs approaching a major development decision.
Each program would require another $10 million in spending to reach its next milestone. The company is considering committing $100 million across all ten.
Now suppose additional testing identifies three programs that are very unlikely to succeed.
If the company stops those programs, it could avoid $30 million in planned spending.
However, that is only a gross estimate.
We must subtract the cost of collecting the information, and we must consider whether the screening system mistakenly rejects a candidate that would have succeeded.
An overly aggressive screening system might save money in the short term while reducing the company’s chance of producing a valuable medicine.
The Financial Value of Better Decisions
A reliable AI decision system needs to be measured using more than prediction accuracy.
It should be evaluated according to how much unnecessary spending it avoids, how often it correctly identifies genuine problems, and how often it rejects valuable opportunities.
For example, imagine two systems evaluating one hundred candidates.
System A rejects thirty candidates and correctly identifies twenty-five as unsuitable. System B rejects forty candidates but incorrectly eliminates several candidates that would have succeeded.
System B may appear more aggressive and less expensive because it allows fewer projects to continue. But the company may lose far more economic value through incorrect decisions.
The best system is not necessarily the one that stops the most projects. It is the one that improves expected outcomes after accounting for both the savings and the cost of missed opportunities.
What Founders Should Measure
Boston startups should track whether AI recommendations improve decisions compared with a clearly defined alternative.
For every decision, researchers should record the information available at the time, the model’s recommendation, the scientific team’s recommendation, the action taken, and the eventual outcome.
Over time, this creates a record of whether AI changes decisions in useful ways.
Without such a record, companies risk confusing impressive predictions with demonstrated business value.
Why AI’s Early Clinical Success Data Is Encouraging but Incomplete
The pharmaceutical industry is beginning to accumulate real clinical evidence from AI-assisted drug programs.
That is a major change from the period when most claims focused on computer predictions or laboratory results.
However, the available clinical evidence still has limitations that matter for business decisions.
Phase 1 Success Is Not the Same as Drug Approval
The 2024 Drug Discovery Today analysis reported Phase 1 success rates of approximately 80% to 90% among the AI-discovered molecules included in its assessment.
The finding suggests that AI methods may be effective at identifying or designing molecules with properties suitable for early human testing.
But Phase 1 and Phase 2 ask different questions.
Early human trials often focus on safety, how the drug behaves in the body, and how much can be administered. Later trials need to establish whether the treatment actually produces a useful medical benefit.
A molecule can perform well in early safety testing and still fail to treat the disease.
Small Samples Can Create Misleading Comparisons
When only a limited number of AI-discovered drugs have progressed through a particular clinical phase, the reported success rate can change dramatically as additional results become available.
Imagine ten programs entering a study phase. If eight succeed, the success rate is 80%.
If another ten programs enter and only four succeed, the combined success rate falls to 60%.
This does not mean the original analysis was necessarily wrong. It means small samples provide uncertain estimates of how future programs will perform.
Researchers also need to account for differences in the diseases being studied, the types of medicines, the companies involved, and how success is defined.
Comparing an AI-heavy portfolio focused on certain diseases with a historical industry portfolio containing many other diseases can produce misleading conclusions.
Why Independent Evaluation Matters
Company announcements are useful sources for understanding what a platform is doing. They are not substitutes for independent comparisons.
A company has strong reasons to highlight successful programs and promising milestones. Its disclosures may not provide enough detail about programs that were stopped or experiments that produced weak results.
A rigorous comparison should track all eligible programs, including failures and programs still awaiting outcomes.
It should also distinguish between drugs created primarily through AI-based design and conventional drugs that used AI for a supporting activity.
Until larger and more comparable clinical datasets exist, claims about dramatic improvements in approval rates should be treated as possibilities rather than proven industry-wide results.
Original Finding #7: AI Can Lower the Cost per Experiment While Increasing Total Research Spending
This is one of the most important business lessons emerging from AI-driven science.
A company can make each experiment cheaper and still spend more money on experiments overall.
That may sound contradictory, but it is a normal outcome when technology makes an activity easier to perform.
When Lower Costs Encourage More Research
Imagine a biotech company that runs 1,000 experiments per year at an average cost of $10,000 each.
Its annual experimental spending is $10 million.
Now suppose AI and automation reduce the average cost per experiment to $6,000.
If the company continues running 1,000 experiments, spending falls to $6 million.
But if it increases activity to 2,000 experiments, total spending rises to $12 million.
Cheaper Experiments Do Not Always Mean Lower Spending
BostonBizScene illustrative experiment-volume model. Original cost per test: $10,000. AI-assisted cost per test: $6,000. No company-specific costs implied.
In this example, unit costs decline by 40%, but total spending increases by 20% when the number of experiments doubles.
That is not automatically a problem.
The company may be collecting twice as much useful scientific information, creating more candidates, or answering questions that were previously too expensive to study.
But investors must distinguish between cost efficiency and overall spending.
The Better Metric Is Useful Information per Dollar
Experiment volume is not a measure of scientific success.
Some experiments provide clear answers that help a research program move forward. Others produce results that are difficult to interpret or too uncertain to guide decisions.
If automation increases the number of experiments without improving data quality, the company may create more information than its scientists can use.
A more meaningful measure is the cost of obtaining a reliable answer to an important research question.
This is particularly relevant to Boston’s autonomous laboratory companies, which are developing systems capable of running larger numbers of experiments.
The long-term advantage may come from learning more quickly, rather than from simply lowering the laboratory budget.
Original Finding #8: Boston’s Biggest Advantage May Be Connecting AI Predictions With Physical Experiments
Boston’s AI-biotech ecosystem has an advantage that is difficult to reproduce through software alone.
The region brings together computer scientists, biologists, medical researchers, hospitals, biotech founders, pharmaceutical companies, and specialized research service providers.
These groups do not always work together smoothly. But their proximity creates opportunities to combine computational predictions with laboratory testing and clinical knowledge.
AI drug discovery depends heavily on that connection.
The Problem With AI Models That Only Learn From Old Data
Many AI systems are trained using information that scientists have already collected.
That information can contain missing measurements, inconsistent laboratory methods, and biases toward molecules that researchers chose to study in the past.
An AI model can learn patterns from these datasets, but it may perform less reliably when researchers ask it to design molecules that are substantially different from those in its training data.
The best way to improve the model is often to generate new experimental evidence.
This creates a cycle in which AI suggests an experiment, scientists test the idea, the system analyzes the results, and the next prediction improves.
Why Physical Testing Remains Essential
Biology is full of unexpected interactions.
A molecule that works in a computer simulation might fail because it breaks down too quickly, behaves differently inside cells, or interacts with something that the model did not account for.
A protein that binds strongly to its intended target may also bind to another protein in a harmful way.
A treatment that works in one laboratory system may not perform the same way in a human body.
Physical experiments provide information that models cannot reliably invent.
The key is to use computation to make those experiments more informative.
Boston’s Opportunity Is Reducing the Cost of the Next Useful Answer
The strongest AI-biotech systems may combine three capabilities.
They need models that suggest useful experiments, laboratory systems that execute those experiments reliably, and data infrastructure that allows new results to improve future predictions.
When these capabilities work together, companies can reduce the amount of repeated work needed to solve a research problem.
However, each part of the system adds costs.
Building the AI model requires specialized people and computing resources. Operating the laboratory requires equipment and supplies. Combining the two requires reliable data handling and scientific oversight.
The economic advantage only becomes clear when the improvement in results is large enough to justify the full cost of the system.
The Hidden Costs of AI Drug Discovery That Founders Cannot Ignore
AI is often described as a substitute for expensive laboratory work. In practice, the technology introduces new cost categories that need to be included in financial planning.
A company that fails to account for those costs may overestimate its actual savings.
Computing Can Become a Major Research Expense
Training and operating advanced biological AI models can require substantial computing resources.
The costs depend on model size, the amount of data being processed, the hardware used, and how frequently scientists need to run predictions.
A small model that ranks known molecules may be relatively inexpensive to operate. A large system that performs repeated protein design calculations or complex simulations may require far more resources.
Companies need to measure computing costs per useful result, not simply the total amount of computing they purchase.
If a more expensive model produces only a small improvement in prediction quality, a simpler approach may be better for the business.
Scientific Data Is Expensive to Create
High-quality AI models often depend on experimental data that cannot be downloaded from a public database.
Companies may need to perform thousands of measurements, standardize laboratory procedures, and repeat tests to confirm reliability.
This work can be expensive, but it may also create a defensible business advantage.
A company that owns reliable experimental data for a difficult biological problem may be able to make predictions that competitors cannot easily reproduce.
The financial question is whether the value of those predictions justifies the cost of collecting and maintaining the data.
AI Specialists Do Not Replace All Other Scientists
An AI-native biotechnology company still needs people who understand biology, chemistry, medicine, and experimental design.
It may also need specialists in software, machine learning, data engineering, laboratory automation, manufacturing, and regulatory work.
These teams can be more productive when they collaborate well, but they can also be expensive to assemble.
The savings from automating one type of research work may be partly offset by additional technical staffing.
Founders should avoid assuming that adding AI automatically means operating with a much smaller team.
Validation Creates Its Own Expenses
Every AI prediction that influences an important drug development decision needs an appropriate level of testing.
Models can make mistakes. Their performance can change when new data becomes available or when researchers apply them to unfamiliar biological systems.
Companies must therefore invest in quality checks, repeatability, data controls, and documentation.
These activities are particularly important when AI is used to support decisions involving human safety.
The goal should not be to remove scientific oversight. It should be to make scientific oversight more effective.
What Boston’s Funding Environment Tells Us About the Economics
Massachusetts remains a major biotechnology financing market, which gives AI-driven companies access to investors and industry partners familiar with the long timelines of drug development.
MassBio’s 2025 Biopharma Funding & Pipeline Report recorded approximately $6.85 billion in venture capital investment across 197 rounds involving Massachusetts-headquartered companies.
The report also noted that Massachusetts companies increased their drug candidate pipelines by nearly 14% year over year.
MassBio
These numbers provide important context, but they do not prove that AI is making drug discovery cheaper.
More Funding Does Not Necessarily Mean Greater Research Efficiency
Venture investment can increase because companies are developing promising medicines, building expensive platforms, or pursuing more ambitious research programs.
A funding round tells us that investors are willing to commit capital. It does not establish how efficiently that capital will be used.
Similarly, an increase in the number of drug candidates does not necessarily mean the industry is becoming more productive.
Those candidates still need to survive clinical development.
Original Analysis: The Average Reported Funding per Round Was About $34.8 Million
Using MassBio’s annual totals, we divided $6.85 billion by 197 reported rounds.
The resulting arithmetic average is approximately $34.8 million per round.
This number is not the median funding round, and it does not describe a typical Massachusetts startup. A relatively small number of large transactions may account for a substantial portion of the total capital raised.
It also covers the broader Massachusetts biopharmaceutical ecosystem, not AI-biotech companies alone.
Nevertheless, the calculation helps explain why research efficiency matters so much.
For companies raising tens of millions of dollars at a time, even a modest improvement in capital use could meaningfully extend their operating life or fund additional development work.
Why Capital Efficiency Could Become More Important Than Platform Size
In periods when investors strongly favor AI, startups may be rewarded for building ambitious platforms, collecting large datasets, and attracting technical talent.
Over time, investors are likely to pay closer attention to whether those investments produce useful drug candidates and valuable partnerships.
A company with a smaller platform but strong clinical evidence may become more attractive than one with impressive computing infrastructure and little validated output.
Boston’s established pharmaceutical and venture capital communities are well positioned to evaluate that distinction.
They understand that a new research method needs to produce results that survive the transition from laboratory discovery to human medicine.
How Boston Biotech Startups Should Measure AI Cost Savings
Founders need a practical way to determine whether their AI investments are improving the business.
The best approach is to measure costs and outcomes before introducing AI, then compare those results with a carefully defined AI-assisted workflow.
The comparison should use similar research problems, similar quality standards, and consistent accounting methods.
Otherwise, a company may attribute normal changes in its research portfolio to its technology.
Step 1: Establish the Cost of the Existing Workflow
Before adopting an AI tool, a company should determine how much it currently spends to complete the relevant activity.
This might be the cost of selecting a biological target, identifying a promising molecule, validating a protein, or preparing a candidate for clinical development.
The baseline should include employee time, external research services, laboratory materials, equipment usage, data processing, and other direct costs.
Founders should also record how long the activity takes and how often it produces a result that meets the company’s quality standards.
Without a reliable starting point, it is impossible to measure savings credibly.
Step 2: Include the Full Cost of AI
An AI-assisted workflow needs its own cost calculation.
That should include computing, software, model development, data generation, integration, maintenance, staff, and experimental validation.
Where a platform supports several programs, companies need a consistent way to allocate shared expenses.
One-time development costs should be distinguished from recurring operating expenses.
This prevents a common mistake: comparing the direct laboratory cost of an AI-assisted experiment with the full cost of a traditional workflow while ignoring the new platform’s overhead.
Step 3: Measure Outcomes, Not Just Activity
The quality of the result matters as much as its cost.
For example, a system that identifies promising molecules quickly may still perform poorly if most fail later validation.
Companies should track how many candidates pass predefined scientific standards, how long it takes to make a decision, and how often early predictions match later results.
The ultimate goal is to produce more reliable development progress per dollar invested.
Step 4: Test Against a Meaningful Comparison Group
A strong evaluation needs a control.
The control could be a conventional research workflow, a simpler computational method, or a previously established process. Where practical, researchers should assign comparable tasks to different methods without choosing easier problems for the AI system.
Companies should define success before running the evaluation.
They should also record failed experiments and unexpected costs rather than reporting only favorable outcomes.
This makes the results more useful for internal decisions and more credible to investors.
Step 5: Connect the Results to Portfolio Decisions
An AI platform becomes financially valuable when its performance changes investment decisions.
A company might use the technology to reject weak candidates earlier, identify better candidates, reduce unnecessary testing, or improve the design of a clinical study.
Founders should document which decisions changed because of AI and estimate their financial consequences.
That allows management to distinguish between improvements in technical performance and actual reductions in business risk.
The AI Drug Discovery Cost Dashboard Boston Founders Should Build
A useful dashboard should show whether AI is improving research outcomes, financial efficiency, and clinical progress.
The following framework can be used by early-stage startups, established biotechnology companies, or investors reviewing AI-enabled drug development platforms.
| Metric | How to calculate it | Why it matters |
|---|---|---|
| Cost per validated candidate | Total relevant costs ÷ validated candidates | Shows research productivity |
| Time to development candidate | Months from project start to candidate selection | Measures speed |
| Experimental success rate | Experiments meeting preset standards ÷ total experiments | Measures research quality |
| Cost per useful experiment | Full workflow cost ÷ experiments producing decision-ready evidence | Reveals real experimental efficiency |
| Model prediction accuracy | Validated correct predictions ÷ evaluated predictions | Tests AI reliability |
| Cost of AI infrastructure | Allocated computing, model, data, and support spending | Captures hidden expenses |
| Clinical transition rate | Programs advancing ÷ eligible programs with resolved outcomes | Tracks development progress |
| Cost per clinical milestone | Total relevant program spending ÷ milestones achieved | Connects science with finances |
| Portfolio cash efficiency | Cash spent relative to defined development progress | Supports funding decisions |
| Cost per approved medicine | Full portfolio investment ÷ approvals | Measures long-term productivity |
The most useful dashboard will also show results by disease area, research method, and development stage.
A company working on cancer therapies should not automatically compare its clinical success rates with those of a company developing treatments for a completely different disease.
The cost and risk profiles may be fundamentally different.
Why Every Metric Needs a Clear Definition
The dashboard will only work if everyone agrees on what counts as success.
For example, a validated candidate might mean a molecule that passes one laboratory test at one company. Another business might use the same term only after completing a much broader set of experiments.
Those definitions produce very different cost figures.
Founders should write down their criteria, apply them consistently, and disclose changes when methods evolve.
This is particularly important when reporting research efficiency to investors or pharmaceutical partners.
What Investors Should Ask AI-Drug Discovery Companies Before Funding Them
Investors reviewing Boston AI-biotech startups should look beyond the number of models, the size of the dataset, or the speed at which candidate molecules can be generated.
Those capabilities can be valuable, but they need to connect to development outcomes.
Ask Which Part of the Cost Structure Improves
A company claiming that AI reduces drug discovery spending should be able to identify the specific expenses being reduced.
Does the technology lower spending on molecule synthesis? Does it reduce the number of animal studies needed to answer a question? Does it help identify weak candidates before clinical development?
Each claim requires different evidence.
Investors should be skeptical when a narrow technical improvement is presented as a broad reduction in the cost of developing medicines.
Ask Whether the Savings Survive Full-Cost Accounting
A startup might report that AI reduces laboratory work by 60%, but that figure may exclude model training, data preparation, external validation, and specialized staff.
The business case should account for those expenses.
A strong company will be able to show where the savings occur and which new costs are necessary to produce them.
Ask What Happens When the Company Scales
Some AI systems become more efficient as they support additional programs.
A model or laboratory platform developed for one research program might also support others, spreading its initial costs across more work.
But scaling can introduce new expenses and technical challenges.
A platform may need additional computing, equipment, quality controls, or staff as its workload grows.
Investors should examine whether the cost per useful result improves when activity expands.
Ask for Evidence Beyond the Company’s Best Examples
A startup’s most successful experiment may not represent its normal performance.
Investors should request results across a predefined set of programs, including failures.
If the company has completed twenty research projects, the full set may reveal much more about the platform than its two strongest outcomes.
The goal is to understand how consistently the technology works, not simply whether it has ever produced an impressive result.
What Pharmaceutical Companies Should Do Differently
Large pharmaceutical companies face a different challenge from startups.
They often have established research systems, existing clinical pipelines, valuable datasets, and complex operating processes.
For them, the question is not whether they should replace everything with AI.
It is where AI can improve decisions without disrupting scientific quality or introducing unnecessary risks.
Start With Expensive and Repeated Problems
The best initial applications are often tasks performed many times across several programs.
For example, better tools for analyzing experimental results, comparing molecular properties, or organizing clinical information may create savings that accumulate across a large portfolio.
Companies should choose problems where they can measure both cost and quality.
This makes it easier to determine whether an AI investment deserves further funding.
Build Feedback Systems Between Research and Clinical Teams
Discovery teams often learn valuable lessons when drug candidates fail in later development.
Those lessons should improve how future candidates are selected.
AI may help connect clinical observations with earlier experimental data, allowing companies to identify patterns that are difficult to recognize manually.
However, the underlying information needs to be accurate, accessible, and suitable for the proposed analysis.
The value comes from improving future decisions, not simply collecting more data.
Treat Failed Programs as Sources of Information
A failed clinical program can produce valuable evidence about biology, patient selection, drug behavior, or trial design.
Companies should capture that information in a structured way.
If AI systems can learn from both successful and unsuccessful programs, they may help prevent repeated mistakes.
This could become one of the most important long-term sources of improvement in drug development productivity.
What Boston Should Watch Over the Next Three Years
The next phase of AI-driven biotechnology will likely be judged less by the number of molecules generated and more by what happens to those molecules in the laboratory and the clinic.
For Boston, several developments deserve particular attention.
More AI-Designed Medicines Reaching Later Clinical Trials
As AI-assisted candidates advance through clinical development, researchers will have more information about safety, effectiveness, and failure rates.
These outcomes should make it possible to compare AI-heavy drug development programs with more conventional approaches.
The comparisons will still need to account for disease types, treatment methods, patient populations, and the age of each program.
But the evidence should become more useful than early laboratory results alone.
Better Public Reporting of Research Productivity
Investors increasingly need to understand how AI platforms generate economic value.
Companies that report consistent measures of experimental cost, candidate selection, development time, and clinical outcomes will make it easier to evaluate their progress.
There is also an opportunity for independent researchers to build broader datasets using published studies, clinical trial registries, and regulatory records.
Better reporting could help separate genuine improvements from marketing claims.
Stronger Connections Between Automated Laboratories and Clinical Research
Boston has an opportunity to combine several parts of the drug development process that are often separated.
AI models can guide early research. Automated laboratories can test predictions. Clinical researchers can provide evidence about how treatments behave in patients.
Connecting those stages could help companies improve the quality of their decisions.
The biggest gains may come when information from one stage changes what happens in the next.
More Focus on Capital Efficiency
Drug development remains expensive even when early research becomes faster.
Companies advancing into later clinical trials will continue to need substantial funding, manufacturing capabilities, regulatory expertise, and clinical operations teams.
Investors will want to know whether AI helps companies make better decisions about which programs deserve that investment.
This creates an opportunity for Boston startups that can demonstrate reliable improvements without requiring enormous upfront platform spending.
Research Methodology: How BostonBizScene Conducted This Analysis
This article combines published research, company financial disclosures, and original scenario calculations prepared for BostonBizScene in October 2026.
The goal was to examine where AI-driven efficiency might affect the economics of drug discovery, rather than to estimate an unsupported industry-wide saving.
Research Sources
The primary sources include the 2024 JAMA Network Open analysis of drug development costs, the BIO clinical development success study, the 2024 Drug Discovery Today analysis of AI-discovered drug candidates, and a 2026 Nature Reviews Drug Discovery examination of AI’s potential impact.
Company-specific research relied on SEC filings, investor announcements, and official corporate disclosures. Massachusetts funding information came from MassBio.
Original Calculations
Our analysis produced five principal sets of calculations, along with supporting comparisons.
| Analysis | Primary inputs | Calculation method | Main limitation |
|---|---|---|---|
| Nonclinical cost reduction | 68% clinical cost share | Apply hypothetical reductions to remaining 32% | Nonclinical share includes more than discovery |
| Combined clinical and nonclinical savings | Same spending allocation | Apply separate hypothetical reductions to both groups | Does not model success rates |
| Relay expense analysis | 2024 and 2025 Form 10-K R&D data | Calculate year-over-year changes and spending shares | Cannot establish AI causation |
| Molecule efficiency comparison | Recursion’s 330 vs. 2,500 disclosure | Calculate count reduction and ratio | Not a controlled, cost-matched comparison |
| Approval probability scenarios | BIO 7.9% reference; hypothetical higher rates | Divide assumed expected candidate spending by approval probability | Does not model stage-specific failure costs |
We also calculated experiment-volume scenarios, Generate’s reported spending changes, and average Massachusetts venture funding per round.
What This Research Does Not Claim
The analysis does not establish a measured causal reduction in drug development costs from AI.
It does not show that AI-designed medicines have higher ultimate approval probabilities than comparable conventional medicines.
It also does not imply that a company’s declining R&D spending proves its AI platform is becoming more efficient.
Our cost models are deterministic sensitivity analyses, not forecasts or statistical estimates. Because we did not construct a matched dataset of individual AI-enabled and conventional drug programs, we cannot estimate a causal AI savings effect or a confidence interval for that effect.
The original contribution is the transparent application of public financial and clinical benchmarks to specific economic questions. The results identify where savings would matter most and what evidence future research needs to collect.
Reproducibility and Data Limitations
All financial comparisons use publicly reported figures and straightforward arithmetic. Percentages were calculated from unrounded inputs where available and then rounded for presentation.
The company examples were selected to illustrate different approaches and were not drawn from a complete, randomly sampled cohort of Boston AI-biotech companies.
The models should therefore be used to guide questions and decisions, not to predict the actual cost of developing a particular drug.
A more comprehensive follow-up study would require program-level data on research spending, AI implementation costs, experimental outcomes, clinical transitions, failures, and approvals. It would also need comparable non-AI programs and enough years of observation to account for candidates still in development.
Final Takeaway: AI Will Make Drug Discovery Cheaper Only If It Makes Drug Development Smarter
The most important lesson from Boston’s AI-biotech ecosystem is that faster research and cheaper medicines are not the same thing.
AI can help scientists explore more ideas, design molecules more efficiently, and reduce some of the repetitive work involved in drug discovery. Companies are already reporting promising technical improvements, including fewer synthesized molecules and the ability to evaluate large numbers of designs.
But the pharmaceutical industry does not spend money only on discovering molecules.
It spends money determining whether those molecules are safe, whether they help patients, whether they can be manufactured reliably, and whether the evidence supports regulatory approval.
Those later steps remain expensive and uncertain.
Our analysis shows why companies need to look beyond individual research savings. In a model where clinical development represents 68% of direct costs, cutting the remaining spending in half reduces total costs by only 16%. Improvements that affect clinical decisions, failure rates, and portfolio selection could be more financially important, although their actual impact still needs to be demonstrated.
Boston has an opportunity to lead because its biotech ecosystem extends far beyond AI software. The region contains drug developers, experimental scientists, clinical researchers, experienced investors, and institutions capable of producing the physical evidence AI systems need.
The winning companies may not be the ones that generate the most molecules or operate the largest models.
They may be the companies that consistently answer important scientific questions with less wasted work, identify weak candidates before major spending begins, and select treatments with a better chance of helping patients.
The real promise of AI is not simply making drug discovery faster. It is making expensive drug development decisions more reliable.
If Boston’s AI-biotech companies can prove that improvement across enough programs and clinical outcomes, cheaper drug development could become a measurable reality rather than an attractive prediction.
And that would be a far more important achievement than producing another impressive AI-designed molecule.
