Clinical trials have a speed problem.
A promising drug can move through years of laboratory work only to hit a completely different set of problems when human testing begins. Researchers have to design the right study, choose sites, find patients who meet detailed rules, review thousands of medical records, collect clean data, measure whether the treatment worked, and keep dozens or sometimes hundreds of research sites moving in the same direction.
Every delay matters. A weak site can sit open without enrolling patients. A patient who looks eligible in a database may fail screening after a coordinator spends hours reviewing the chart. A poorly chosen trial endpoint can make a useful treatment look less effective than it really is. A protocol that seemed reasonable on paper may turn out to exclude most of the patients doctors actually see.
Boston’s AI industry is increasingly attacking these problems.
The important change is that this is no longer simply about using artificial intelligence to discover a molecule. A growing group of Boston, Cambridge, and Waltham companies is applying AI after a drug enters development. Their tools are being used to design trials, predict enrollment, choose research sites, search electronic health records, identify eligible patients, standardize medical images, discover biomarkers, monitor studies, and analyze real-world patient data.
That could turn clinical development into one of the most important commercial markets for Boston’s broader AI-and-biotech ecosystem.
BostonBizScene analyzed nine Boston-area companies with active AI products that directly touch clinical trials. Our analysis found that eight of the nine companies—89%—have products connected to patient identification or screening.
But only one of the nine has a clearly marketed AI product that extends across patient outreach and consent support.
That gap tells us something important.
Boston has become very good at using AI to answer the question, “Which patient might qualify?”
The next opportunity may be building the infrastructure that turns that answer into an enrolled participant quickly, safely, and reliably.
The Short Version: AI Is Moving Into the Messiest Parts of Clinical Trials
AI in clinical research is often described in very broad terms. That makes the market sound less interesting than it actually is.
The biggest opportunity is not one giant AI system that magically runs an entire clinical trial. The more realistic opportunity is dozens of focused systems removing specific delays from the development process.
A trial might begin with a platform such as PhaseV helping a sponsor test alternative study designs. TriNetX or ConcertAI might use real-world patient data to estimate whether the planned population can actually be found. Another system could identify sites with enough suitable patients.
Once the study starts, AI can search medical records to identify potential participants. Iterative Health can apply AI to gastrointestinal trials where endoscopy is important. PathAI can make pathology measurements more consistent. BostonGene can help identify biological signals linked to treatment response. Other platforms can monitor trial performance or turn real-world clinical information into evidence.
These are separate problems, but together they form something larger.
Boston is beginning to build an AI layer around clinical development itself.
Our research suggests that this emerging layer is currently strongest in four areas:
| Boston AI clinical-trial category | Companies in our sample with capability | Share of 9-company sample |
|---|---|---|
| Patient identification and screening | 8 | 88.9% |
| Site selection and feasibility | 4 | 44.4% |
| Biomarkers and trial endpoints | 4 | 44.4% |
| Protocol and trial design | 3 | 33.3% |
| Trial operations and monitoring | 3 | 33.3% |
| Real-world evidence | 3 | 33.3% |
| Patient engagement and consent | 1 | 11.1% |
Source: BostonBizScene analysis of publicly available company product pages, company announcements, scientific publications, and clinical-trial materials available through October 4, 2026. Categories are binary capability classifications and should not be interpreted as market share.
The imbalance is striking.
Patient identification / screening ██████████████████ 88.9%
Site selection / feasibility █████████ 44.4%
Biomarkers / endpoints █████████ 44.4%
Protocol / trial design ███████ 33.3%
Trial operations / monitoring ███████ 33.3%
Real-world evidence ███████ 33.3%
Patient engagement / consent ██ 11.1%Boston’s AI clinical-trial market is therefore not evenly distributed.
It is heavily concentrated around finding and qualifying the right patient.
That makes sense because patient selection sits at the intersection of several things Boston already has in unusual concentration: hospitals, electronic health records, biomedical research, specialist physicians, biotech companies, medical data businesses, and machine-learning talent.
Why Clinical Trials Are Such a Good Target for AI
Drug development contains enormous scientific uncertainty. AI cannot remove that uncertainty. A drug that does not work in humans cannot be rescued by a better dashboard.
But clinical trials also contain large amounts of operational friction.
That is different.
Operational friction includes reading charts manually, comparing dozens of inclusion and exclusion rules against medical records, estimating how many eligible patients a hospital actually has, deciding which sites to activate, checking images, reviewing pathology samples, preparing reports, and tracking whether enrollment is falling behind.
These are areas where software can potentially make a meaningful difference.
Patient eligibility is still surprisingly manual
One of the clearest Boston examples comes from Mass General Brigham.
Researchers developed RECTIFIER, a system that uses a large language model with retrieval to review unstructured clinical information and help determine whether patients meet clinical-trial eligibility rules.
The research team then did something especially valuable: it tested AI-assisted screening prospectively against manual screening in a randomized clinical trial.
A total of 4,476 patients were randomized between AI-assisted and manual screening. The study found that eligibility decisions happened significantly faster with AI assistance. The AI-assisted group produced 35 enrollments compared with 19 in the manual group.
That is far more useful evidence than a generic claim that AI “improves efficiency.”
BostonBizScene calculation: what the Mass General Brigham trial actually means
Using the published study numbers, we calculated the relative operational difference between the two groups.
| Measure | AI-assisted | Manual | BostonBizScene derived difference |
|---|---|---|---|
| Patients randomized to workflow | 2,242 | 2,234 | Nearly identical groups |
| Eligibility determinations | 458 | 284 | 61% more in AI group |
| Eligibility rate | 20.4% | 12.7% | +7.7 percentage points |
| Patients enrolled | 35 | 19 | 84% more enrollments |
| Approx. enrollment rate | 1.56% | 0.85% | ~84% relative lift |
| Time for >99% of completed screening | <15 days | <50 days | At least ~70% shorter |
Method: BostonBizScene calculations using published patient counts and timing reported in the randomized JAMA study. Relative differences are simple arithmetic calculations and are not substitutes for the study’s formal statistical analysis. The study itself reported an enrollment subdistribution hazard ratio of 1.79.
The timing difference may be the most important number.
Time to >99% completed eligibility screening
AI-assisted ██████ <15 days
Manual ████████████████████ <50 days
Approximate reduction in time: ≥70%That does not prove every AI screening tool will produce the same result. The study was conducted in one health system and in a particular heart-failure research setting.
But it proves something more fundamental.
Automating parts of trial screening is no longer only a theoretical idea.
It can be tested like any other intervention, and at least one Boston team has produced randomized evidence showing a meaningful operational improvement.
BostonBizScene Original Research: Mapping Boston’s Clinical-Trial AI Stack
To understand what is actually happening in Boston, we built a simple market map.
Rather than counting every company that mentions AI or clinical research somewhere on its website, we used a narrower definition.
A company had to meet three conditions.
First, it needed a meaningful operating presence in Boston, Cambridge, or the surrounding Boston life-sciences cluster.
Second, it needed an AI-enabled product that directly affects the design, execution, measurement, recruitment, or evidence layer of human clinical research.
Third, the capability had to be described in publicly available product material, research, or company announcements that could be checked.
That produced nine companies for this analysis:
ConcertAI, TriNetX, PhaseV, PathAI, Iterative Health, BostonGene, OM1, AIwithCare, and Bond Health.
This is not intended to be a complete list of every Boston company touching clinical trials. It is a focused sample designed to understand where AI capability is accumulating.
The seven-stage scoring model
We coded each company across seven functions:
- Protocol and trial design
- Site selection and feasibility
- Patient identification and eligibility screening
- Patient engagement and consent
- Biomarkers and trial endpoints
- Trial operations and monitoring
- Real-world evidence
A company received a “1” when we found a clearly marketed or documented capability and a “0” when we did not.
The score therefore measures breadth of public capability, not product quality, revenue, market share, regulatory acceptance, or clinical effectiveness.
Boston clinical-trial AI capability map
| Company | Trial design | Sites | Patient screening | Engagement / consent | Biomarkers / endpoints | Trial ops | RWE | Breadth score |
|---|---|---|---|---|---|---|---|---|
| ConcertAI | ✓ | ✓ | ✓ | — | — | ✓ | ✓ | 5/7 |
| PhaseV | ✓ | ✓ | ✓ | — | ✓ | ✓ | — | 5/7 |
| TriNetX | ✓ | ✓ | ✓ | — | — | — | ✓ | 4/7 |
| Iterative Health | — | ✓ | ✓ | — | ✓ | ✓ | — | 4/7 |
| BostonGene | — | — | ✓ | — | ✓ | — | — | 2/7 |
| OM1 | — | — | ✓ | — | — | — | ✓ | 2/7 |
| Bond Health | — | — | ✓ | ✓ | — | — | — | 2/7 |
| PathAI | — | — | — | — | ✓ | — | — | 1/7 |
| AIwithCare | — | — | ✓ | — | — | — | — | 1/7 |
Source and methodology: BostonBizScene manual coding of public company materials and scientific evidence available through October 4, 2026. A check means the company publicly describes a meaningful product capability in that category. Absence of a check does not prove the company has no capability; it means we did not count one under our public-evidence standard.
Across 63 possible company-category combinations, we found 26 positive capability matches.
That produces a Boston Clinical Trial AI Coverage Index of 41.3%.
The percentage itself is not meant to be treated like an established industry benchmark. Its value is comparative. It shows that Boston companies are generally specialists rather than seven-stage, end-to-end platforms.
That may actually be healthy.
Clinical trials are too complex for one generic AI product to solve everything well.
Original Finding #1: Patient Screening Is Boston’s Clear AI Beachhead
Eight of the nine companies in our sample touch patient identification, qualification, or selection.
No other category comes close.
This means patient selection appears eight times more frequently in our dataset than AI-supported patient engagement and consent.
That pattern makes economic sense.
Finding eligible patients is a difficult information problem. Eligibility rules are written in protocols. Patient information is spread across laboratory values, medications, physician notes, pathology reports, imaging results, past diagnoses, procedures, and dates.
Traditional database searches work well when the answer sits inside a neat structured field.
Clinical reality is rarely that neat.
A patient may qualify because of something described in a physician’s note six months ago. Another may appear qualified until a radiology report reveals an exclusion condition. One trial might require prior failure of a specific class of therapies. Another might depend on a disease-severity score recorded inside a specialist report.
Large language models and clinical machine-learning systems are particularly interesting here because they can work across both structured and unstructured information.
AIwithCare shows how far this can go
AIwithCare is commercializing the RECTIFIER approach for trial screening.
The company’s public materials point to the Mass General Brigham research showing faster eligibility determination and higher enrollment with AI assistance. The earlier retrospective research found that RECTIFIER achieved strong sensitivity and specificity when judging trial eligibility from clinical records, while keeping final human review as an important safety step.
The lesson for biotech companies is not “replace coordinators.”
It is almost the opposite.
Use AI to remove the work that makes coordinators slow.
If software can place the most promising charts at the top of the queue and show the evidence supporting each match, humans can spend more of their time on judgment, patient communication, unusual cases, and actual trial execution.
Bond Health Is Pushing Beyond Matching Into the Recruitment Funnel
Boston-based Bond Health is interesting because it attacks a different part of the problem.
Its platform combines EHR screening, patient outreach, pre-screening, scheduling, and consent support. The company describes a three-part workflow called Identify, Engage, and Consent. Its system uses clinical records to rank possible participants and provide evidence behind the match before moving selected patients into outreach workflows.
Bond says its engagement layer can use voice and text agents for patient outreach and scheduling, while its consent system helps patients understand approved consent material in simpler language while leaving formal consent responsibility with research staff.
This matters because finding a candidate is not the same thing as enrolling one.
A typical recruitment funnel looks more like this:
10,000 patient records
↓
Potential matches
↓
Coordinator review
↓
Patient contacted
↓
Patient responds
↓
Pre-screen passes
↓
Site visit
↓
Full screening
↓
Consent
↓
RandomizationAI can improve the top of this funnel and the trial can still fail further down.
A perfectly matched patient who never answers the phone does not help enrollment.
Neither does a patient who receives a confusing message, cannot understand the study requirements, or drops out before completing screening.
That is why Boston’s low level of capability in AI-supported engagement may represent an important market opening.
Original Finding #2: Boston Is Building “Qualification Intelligence”
The phrase we would use for the market emerging in Boston is qualification intelligence.
Traditional clinical-trial software often starts after a site has been selected and a study is underway.
Boston’s newer platforms are moving upstream.
They are trying to answer questions before expensive decisions are made:
Is this protocol realistic?
Are enough patients available?
Which eligibility rule is destroying the addressable population?
Which sites really have the patients?
Which patients are most likely to qualify?
Which biological subgroup may respond?
Which endpoint can measure response most reliably?
These decisions are connected.
A poorly designed inclusion rule changes the patient pool. That changes site feasibility. Site feasibility changes enrollment speed. Enrollment changes trial cost. Patient heterogeneity affects statistical power. Measurement noise affects the size of the treatment effect researchers can see.
The best AI clinical-trial businesses may therefore be the ones that connect these decisions instead of optimizing them in isolation.
PhaseV Is Trying to Move AI Into Trial Design Itself
Boston-based PhaseV sits near the beginning of that decision chain.
Its platform uses machine learning and large-scale simulation to test different clinical-trial designs. The company says its Trial Optimizer can simulate millions of adaptive, Bayesian, and fixed trial designs and evaluate ways to reduce sample size while maintaining statistical requirements.
This is a fundamentally different use of AI from patient matching.
Instead of asking, “Who belongs in this trial?” PhaseV asks something closer to, “What should this trial look like in the first place?”
The company raised a $50 million Series A in 2025, bringing reported total funding to $65 million. At the time, PhaseV said it was working with more than 30 global customers. Its current site says the platform now supports more than 40 biopharma companies and CROs.
Simulating the trial before spending money on the trial
One PhaseV analysis illustrates why simulation is becoming important.
The company tested an adaptive Phase 2 dose-finding design across different possible treatment effects. It ultimately ran 60 million simulated trials across the broader analysis.
Its reported result was that the adaptive design achieved 71.1% assurance of success compared with 63.4% for the fixed design while using about 5% fewer patients on average.
Those are company-generated results rather than an independent head-to-head validation of the commercial platform, so they should be treated differently from the randomized Mass General Brigham screening study.
But the underlying idea is powerful.
A clinical-trial team cannot test 10 alternative protocols on real patients before choosing one.
Software can test enormous numbers of simulated versions.
That turns trial design into a search problem.
PhaseV is also moving downstream
PhaseV has expanded beyond statistical trial design.
Its ClinOps platform uses patient-level data for site selection and operational decision-making, while its 2026 Enrollment Lab is designed to model enrollment potential and test how changing protocol criteria affects the available patient population before the protocol is locked.
This is important because protocol design and recruitment should not be separate conversations.
A protocol can be scientifically beautiful and operationally impossible.
The better system asks both questions before enrollment begins.
TriNetX Shows Why Real-World Data Is Becoming Trial Infrastructure
TriNetX is one of the most important companies in Boston’s clinical-data ecosystem.
The company is headquartered in Cambridge and says its federated network surpassed 300 million patients across more than 240 healthcare organizations and more than 13,000 sites in 2026.
The size matters, but the bigger story is what can be done with the network.
TriNetX combines real-world clinical information with analytics and AI to help researchers evaluate protocols, understand patient populations, identify research sites, and generate evidence.
The company said in January 2026 that its platform had produced measurable reductions in protocol amendments and improvements in site identification efficiency. It also added conversational AI and API capabilities to make complex research analysis easier to perform.
Site selection is becoming a data problem
Historically, companies could select trial sites partly based on previous relationships and historical performance.
That does not always tell sponsors whether a site has enough suitable patients for a new protocol right now.
Imagine two hospitals.
Hospital A is famous, has a strong investigator, and has run ten trials in the therapeutic area.
Hospital B has less history but has three times as many currently eligible patients.
Which site should receive the study?
The answer should increasingly be driven by current patient data rather than reputation alone.
That is one of the areas where systems such as TriNetX are changing clinical operations.
Original Finding #3: Boston’s Clinical-Trial AI Market Is Concentrated in the Middle of the Funnel
When we grouped our 26 capability tags, another pattern appeared.
The largest concentration sits around feasibility, site selection, patient qualification, and measurement.
Those are the parts of a clinical trial where clinical information has to be converted into operational decisions.
Consider the flow:
Drug candidate
↓
Trial design
↓
Feasibility
↓
Site selection
↓
Patient identification
↓
Patient qualification
↓
Enrollment
↓
Treatment
↓
Endpoint measurement
↓
Analysis
↓
Regulatory evidenceBoston is particularly strong from feasibility through measurement.
This matches the strengths of the region.
Boston has hospitals producing real-world clinical data. It has companies specializing in pathology and biological measurement. It has biotech sponsors running complex studies. It has academic groups developing machine-learning methods. It has specialists who understand difficult diseases.
Those assets become more valuable when AI is applied to the connections between them.
ConcertAI Is Trying to Build a Broader Clinical-Trial AI Layer
Cambridge-based ConcertAI is one of the broadest platforms in our sample.
The company launched Accelerated Clinical Trials, or ACT, in February 2026. ConcertAI says the platform uses agentic AI, real-world data, and proprietary information across protocol design, site selection, and study monitoring. The company says the system is designed to shorten overall trial timelines by 10 to 20 months, although that is a company target and should not be read as an independently verified average result.
ConcertAI also expanded its CancerLinQ products with AI-driven trial matching in 2026.
The company says the capability can identify trial-eligible patients 3.3 times faster and works against a clinical-data base containing more than 13 million patient records. Again, those are company-reported product results rather than independent randomized evidence.
Even with that distinction, the strategy is worth watching.
ConcertAI is trying to connect the front and back ends of clinical development.
Instead of AI being a single point tool used by one department, it becomes an operating layer shared across protocol teams, clinical operations, sites, and evidence generation.
Iterative Health Shows Why Vertical AI Can Beat Generic AI
One of Boston’s most interesting clinical-trial companies is Iterative Health.
The Cambridge company has focused heavily on gastrointestinal research rather than trying to solve every therapeutic area at once.
That specialization matters.
Clinical trials do not all fail for the same reasons.
An oncology trial can depend heavily on genomic biomarkers. A rare-disease trial may struggle mainly with identifying a tiny population. A gastrointestinal trial may depend on endoscopic imaging and disease-severity measurements.
Iterative Health built around that last problem.
Using AI to find IBD trial patients
Inflammatory bowel disease trials often use colonoscopy findings when deciding whether a patient qualifies and when measuring whether treatment worked.
Those assessments can involve human judgment.
Iterative Health uses AI to analyze endoscopy information and support more standardized disease scoring. The company has also built AI-powered recruitment services designed to help research teams identify likely candidates for IBD trials.
This is a good example of why vertical AI can create value.
A generic language model does not automatically understand the operational reality of an IBD trial.
A system built around the actual clinical workflow can.
The network may matter as much as the algorithm
Iterative Health has also built a research-site network.
Its 2026 research reported enrollment rates more than three times industry benchmarks across the trials evaluated. The company says the performance comes from combining operational support, site infrastructure, trial-specific training, and AI-enabled pre-screening rather than from AI alone.
That last point deserves attention.
The winning clinical-trial AI companies may not look like pure software businesses.
AI alone cannot make a research coordinator call a patient. It cannot create a functioning site. It cannot negotiate a contract, train staff, schedule procedures, or fix every operational bottleneck.
The best systems therefore may combine software with real-world research infrastructure.
PathAI Is Attacking Measurement Noise Inside Trials
Boston-based PathAI attacks a different problem.
Clinical trials often depend on pathology.
Researchers need to determine whether tissue shows a certain disease feature or whether that feature changed after treatment. When measurements depend heavily on manual interpretation, variability can weaken trial data.
PathAI uses machine learning to make those measurements more consistent.
In February 2026, the company expanded its AISight Clinical Trials platform with AI-powered tools for inflammatory bowel disease histopathology. PathAI said the tools were designed to support standardized, reproducible trial endpoints inside a GCP/GCLP-compliant workflow.
Why better endpoints can make trials faster
This is an underappreciated part of clinical-trial AI.
Suppose Drug A creates a real biological improvement.
If the measurement used to detect that improvement is noisy, the trial may need more patients to separate the drug signal from measurement variation.
More consistent measurement can potentially improve the quality of the signal.
That does not automatically mean every trial becomes smaller. Statistical requirements still depend on the disease, treatment effect, endpoint, design, and regulator.
But measurement precision matters.
This is why AI pathology belongs in the clinical-trial conversation even though it looks very different from patient recruitment software.
BostonGene Is Moving AI Into Biological Patient Selection
BostonGene, based in Waltham, represents another major piece of the emerging stack.
Its platform combines genomic, transcriptomic, immune, imaging, and clinical information to model cancer biology.
In drug development, the goal is not simply to describe the tumor.
It is to identify which biological signals may explain response, resistance, or toxicity and use those signals to make better clinical-development decisions.
The company announced several collaborations in 2026 that demonstrate this model.
A collaboration with Daiichi Sankyo is applying AI-driven multimodal analysis to an antibody-drug conjugate program, including patient-selection strategy and clinical-development insights. Another partnership with Ottimo Pharma is using BostonGene’s technology around the development of an immuno-oncology therapy.
BostonGene also announced work with Kyoto University connected to a multicenter Phase II study involving advanced gastrointestinal cancers.
Patient selection is moving from demographics to biology
The long-term change here could be significant.
Traditional trial eligibility might say:
Patient has disease X.
Patient is within age range Y.
Patient has received treatment Z.
Modern precision trials increasingly ask another question:
Does this patient’s biology make the mechanism of this drug particularly relevant?
That is a much harder problem.
But if AI can help find biological patterns that predict response, trials could gradually become better at concentrating enrollment around patients who are most informative.
The goal is not to manipulate a trial until it succeeds.
The goal is to understand the disease well enough to test the right treatment in the right biological population.
OM1 Shows How Clinical Data Can Find Patients Who Are Almost Invisible
Boston-based OM1 approaches clinical research through longitudinal real-world data and AI-driven phenotyping.
This is especially useful when simple diagnosis codes do not adequately identify the patient population researchers need.
OM1 has described a pilot in which its technology screened the full EHR population of a health system to identify patients at high risk for a rare lipid-storage disease for possible study enrollment.
This matters because rare-disease recruitment has a basic mathematical problem.
If a disease affects only a tiny share of the population, a research team cannot afford to manually inspect huge numbers of records.
The value of AI rises as the search problem becomes harder.
The same logic applies beyond rare diseases.
AI phenotyping can potentially identify patients whose real clinical pattern matches a research population even when the necessary information is distributed across years of records.
Original Finding #4: AI Could Change the Economics of Trial Sites
Clinical trials are often discussed from the sponsor’s perspective.
There is another side.
Sites have limited staff.
A coordinator spending half a day reviewing patients who clearly do not qualify has lost time that could have been spent on participants who do.
This makes false-positive candidate lists expensive.
The important metric is therefore not simply:
How many patients did the AI find?
A better measurement system includes:
| Metric | Why it matters |
|---|---|
| Qualified patients per 100 records reviewed | Measures screening precision |
| Coordinator minutes per qualified patient | Measures staff burden |
| Contact rate | Shows whether matches can actually be reached |
| Pre-screen pass rate | Tests candidate quality |
| Full-screen failure rate | Shows how much bad qualification remains |
| Days from identification to first contact | Measures execution speed |
| Consent completion | Measures funnel progression |
| Randomizations per active site-month | Measures actual enrollment |
| Cost per randomized patient | Connects AI to economics |
This is where clinical-trial AI buyers should become much more demanding.
A vendor saying it processed one million records sounds impressive.
The question is whether it produced more randomized patients with less staff time and acceptable accuracy.
Why Site Selection May Be the Next Big AI Battleground
Sponsors regularly open more sites because they are worried enrollment will be slow.
More sites can help.
They can also add cost, contracting work, monitoring burden, and management complexity.
If AI improves site forecasting, sponsors could potentially concentrate resources on sites that have both the right patients and the operational ability to enroll them.
TriNetX, PhaseV, ConcertAI, and Iterative Health all touch this problem from different directions in our sample.
That makes site intelligence the second-most crowded part of Boston’s clinical-trial AI stack.
The old question was “Has this site performed before?”
The new question is increasingly:
“Will this site perform for this exact protocol?”
Those are not the same thing.
Historical enrollment in lung cancer does not prove a center has enough patients with a particular mutation, treatment history, disease stage, laboratory range, and combination of exclusions required by a new protocol.
AI can help calculate that difference before a site is activated.
That could prevent months of low-productivity trial activity.
The Biggest Mistake: Treating AI as a Layer Added After the Protocol Is Finished
Clinical teams should not wait until recruitment fails before using data.
By then, some of the most important decisions have already been made.
A much stronger approach begins during protocol development.
Step 1: Run the proposed eligibility criteria against real patients
Before locking a protocol, sponsors should estimate how many patients in real healthcare datasets meet the proposed rules.
If one exclusion criterion removes 35% of otherwise suitable patients, the clinical team should know that before the final protocol is signed.
That does not mean relaxing necessary safety rules.
It means understanding their operational cost.
Step 2: Model where the patients actually are
The next step is geographic.
Sponsors should identify the health systems and research sites with the strongest concentration of suitable patients.
That allows the site strategy to start with patient availability rather than reputation.
Step 3: Measure competing trials
A site may have 300 apparently eligible patients.
That number is less attractive when six competing trials are recruiting the same population.
Enrollment forecasting therefore needs to include competition.
PhaseV’s Enrollment Lab is one example of the industry moving toward this kind of analysis before protocol lock.
Step 4: Automate pre-screening, not final judgment
AI should remove repetitive chart work while keeping clear human responsibility.
The Mass General Brigham RECTIFIER research is important partly because the model is not simply presented as an autonomous enrollment machine. The researchers explicitly discuss the importance of mitigation measures such as clinician review before patient engagement.
That is likely to remain the more credible operating model.
AI narrows.
Humans decide.
Original Finding #5: The Market Is Moving From Point Tools Toward Connected Decision Systems
The most important strategic trend may be convergence.
Three years ago, it was easier to describe many clinical-trial AI products as separate categories.
One system selected sites.
Another matched patients.
Another analyzed pathology.
Another produced real-world evidence.
Those lines are beginning to blur.
ConcertAI is expanding across design, site selection, patient matching, monitoring, and evidence.
PhaseV began with trial optimization and now operates in site selection and enrollment forecasting.
Iterative Health combines AI with a research-site network and trial operations.
Bond Health is connecting screening, outreach, scheduling, and consent support.
TriNetX connects real-world data with feasibility, site identification, research analytics, and evidence generation.
This is not accidental.
Each clinical-trial decision changes the next one.
Connected systems can learn from the entire chain.
What Boston Biotech Companies Should Actually Do With AI
The wrong strategy is to create a company-wide “AI initiative” and then search for something to automate.
Start with the bottleneck.
If enrollment is slow, measure the funnel before buying software
Teams should know how many patients move through every recruitment step.
If 10,000 charts produce 1,000 potential candidates, 300 coordinator-approved candidates, 100 contacts, 40 screens, and 10 randomizations, the company should identify where the biggest loss occurs.
An EHR screening product does not fix a contact problem.
An outreach agent does not fix an impossible protocol.
A better endpoint does not fix bad site selection.
The technology should match the bottleneck.
If screen failure is high, audit the eligibility logic
Sponsors should find out which inclusion and exclusion conditions are causing the most failures.
This can reveal criteria that look harmless individually but dramatically reduce the addressable population when combined.
That analysis belongs before the next protocol is finalized.
If sites are underperforming, stop looking only at historical enrollment
Compare historical performance with current patient availability.
Then monitor whether patients are actually entering the funnel.
A site with excellent historical performance but no eligible population is still a bad site.
If measurement is noisy, investigate AI endpoints
For trials involving pathology, imaging, endoscopy, or other assessments with meaningful reader variation, teams should test whether validated computational methods can improve consistency.
This is where companies such as PathAI and Iterative Health become especially relevant.
How Sponsors Should Evaluate Clinical-Trial AI Vendors
Buying clinical AI requires more skepticism than buying ordinary enterprise software.
A slick demonstration is not enough.
Ask what evidence supports the performance claim
There is a large difference between these statements:
“We believe our software improves enrollment.”
“Our internal retrospective analysis found higher enrollment.”
“A customer case study found higher enrollment.”
“A peer-reviewed prospective study found higher enrollment.”
“A randomized prospective study found higher enrollment.”
All five can be true.
They are not equally strong evidence.
Boston companies themselves demonstrate the full range, from product claims to randomized clinical evidence.
Sponsors should know which one they are looking at.
Ask whether performance was measured on your disease
Clinical language differs by specialty.
Patient records differ.
Eligibility criteria differ.
A system performing extremely well in cardiology does not automatically perform equally well in oncology or a rare neurological disease.
Buyers should ask for disease-specific validation whenever possible.
Ask what happens when the AI is uncertain
This is one of the most important questions.
A strong clinical AI system should know how to escalate uncertain cases instead of quietly converting uncertainty into a confident-looking answer.
Human review must be designed into the workflow.
Ask whether every recommendation is traceable
Clinical users need to understand why a patient was matched.
“AI score: 91%” is not enough.
A much more useful interface says:
The patient satisfies criterion A because of this laboratory result.
The patient appears to satisfy criterion B because of this physician note.
Criterion C is uncertain because the relevant imaging report is missing.
That type of evidence makes AI useful to a coordinator.
Regulation Is Moving in the Same Direction
AI in drug development is no longer outside the regulatory conversation.
The FDA issued draft guidance in January 2025 describing a risk-based framework for assessing the credibility of AI models used to produce information supporting regulatory decisions.
In January 2026, the FDA and European Medicines Agency also published ten guiding principles for good AI practice in drug development.
Those principles include having a clear context of use, strong data governance, risk-based performance assessment, lifecycle management, multidisciplinary expertise, and human-centered design.
The FDA has said its drug center dealt with more than 500 submissions containing AI components between 2016 and 2023.
That number matters.
AI in drug development has moved beyond experiments happening outside the regulatory process.
Regulators are already seeing it.
The regulatory question is becoming more specific
The debate is no longer simply:
“Can AI be used in drug development?”
The more important questions are:
What is the model being used for?
How much risk does a wrong output create?
What data trained it?
How was it validated?
Does performance change across populations?
Who reviews its output?
Can the decision be traced?
What happens when the model changes?
Those are much more mature questions.
Why Boston May Have a Structural Advantage
Boston does not have a monopoly on clinical-trial AI.
Major companies are being built in New York, San Francisco, Europe, Israel, and many other markets.
Boston’s advantage is structural rather than exclusive.
The region combines several pieces of infrastructure that are unusually useful for this category.
Hospitals produce the raw material
AI clinical-trial systems need clinical data.
Boston has major academic health systems, specialist centers, and research hospitals that generate large amounts of complex patient information.
The Mass General Brigham RECTIFIER work demonstrates what can happen when clinical researchers and AI developers work inside the same ecosystem.
Biotech creates immediate buyers
Boston also has a dense concentration of drug developers.
That matters because clinical AI is not useful in isolation.
It becomes valuable when sponsors have real protocols, real patient populations, real sites, real endpoints, and real development deadlines.
Universities create technical talent
MIT, Harvard, and the broader research environment continue to produce machine-learning researchers, physicians, statisticians, computational biologists, and biomedical engineers.
Clinical-trial AI needs all of them.
A strong model built without clinical knowledge can fail.
A strong clinical workflow built without modern data engineering will not scale.
A statistically clever design that regulators cannot interpret has limited value.
The Boston market contains talent across those boundaries.
The Bigger Story: Clinical Trials Are Becoming Software-Defined Systems
The most important long-term change is not that trials will become “AI trials.”
It is that more parts of trial execution will become measurable, programmable, and continuously optimized.
Consider what happens when the different layers connect.
A sponsor creates a draft protocol.
Software simulates alternative designs.
Real-world data estimates the eligible population.
AI tests which eligibility criteria shrink that population.
Site-selection software identifies hospitals with the strongest patient pools.
Clinical NLP continuously searches EHRs for candidates.
AI-assisted outreach moves those candidates into screening.
Computer vision evaluates trial images.
Digital pathology standardizes tissue endpoints.
Biological AI identifies responder subgroups.
Operational agents monitor enrollment.
Real-world data follows patients after the controlled study.
Each system makes the next system smarter.
That is much bigger than automating administrative work.
It begins to turn clinical development into a learning loop.
What AI Still Cannot Fix
This shift needs some caution.
AI cannot make a bad drug effective.
It cannot make unsafe eligibility criteria safe.
It cannot replace proper statistical design.
It cannot make biased data representative.
It cannot automatically turn correlation into causation.
It cannot decide whether a sick patient should join a study without the appropriate clinical and ethical process.
And it cannot eliminate uncertainty from human biology.
Those limits are important because clinical trials exist precisely because researchers do not yet know whether an intervention is safe and effective.
AI can improve the system used to answer that question.
It cannot predetermine the answer.
What Boston Business Leaders Should Watch Through 2027
The next stage of the market will be less about whether companies say they use AI and more about measurable proof.
Watch randomized validation.
The Mass General Brigham screening study has raised the standard. Vendors making major performance claims will face growing pressure to show prospective evidence rather than retrospective demonstrations.
Watch protocol intelligence.
The earlier AI enters the trial-design process, the more cost it may prevent later.
Watch integration.
Point tools will increasingly be judged on whether they connect with EHRs, clinical-trial management systems, electronic data capture systems, laboratory infrastructure, imaging systems, and sponsor workflows.
Watch agentic AI carefully.
ConcertAI and newer Boston startups are beginning to apply AI agents to multi-step clinical workflows. These products could reduce administrative work significantly, but autonomous actions around patients and regulated evidence will require much stronger controls than ordinary office automation.
And watch the patient layer.
Our Boston dataset found only one of nine companies with a clearly marketed AI capability spanning patient engagement and consent.
If patient identification becomes highly automated while actual engagement remains mostly manual, the bottleneck may simply move one step downstream.
Final Takeaway: Boston Is Building an AI Operating Layer for Clinical Development
The most important thing happening in Boston clinical-trial AI is not a chatbot.
It is the gradual conversion of clinical development from a chain of disconnected manual decisions into a connected data system.
PhaseV is attacking trial design and simulation.
TriNetX is using large-scale clinical data for feasibility, site intelligence, and research.
ConcertAI is trying to connect AI across multiple stages of trial execution.
Iterative Health is combining specialized AI with real clinical-research infrastructure.
PathAI is making difficult medical measurements more consistent.
BostonGene is applying multimodal biology to patient selection.
OM1 is using AI and longitudinal data to identify hard-to-find populations.
AIwithCare is bringing AI-assisted chart screening into clinical research workflows.
Bond Health is pushing automation deeper into the patient-recruitment funnel.
The companies are different, but their direction is similar.
Clinical development is becoming a computing problem layered on top of a biology problem.
Our BostonBizScene analysis found that 88.9% of the Boston-area companies studied are already working on patient identification or screening, making it by far the most common clinical-trial AI capability in our sample.
That is probably only the first stage.
Once sponsors can identify eligible patients faster, they will demand faster site activation, faster patient engagement, better endpoint measurement, better study monitoring, more flexible trial designs, and stronger evidence.
The competitive advantage will increasingly come from connecting those pieces.
The companies that win may not be the ones with the biggest general-purpose AI model.
They may be the ones that understand exactly where a clinical trial loses weeks, patients, money, or information—and build systems that remove that loss without weakening scientific rigor.
That is why Boston is worth watching.
The region is not simply using AI to invent drugs.
It is beginning to redesign the machinery used to prove whether those drugs actually work.
