For nearly two decades, one business model has shaped the software industry more than almost any other: Software-as-a-Service, or SaaS.
The idea was simple and powerful. A company built software, hosted it in the cloud, and charged customers a monthly or annual fee to use it. Businesses no longer had to install large software packages on their own servers, and employees could access the tools they needed from almost anywhere.
Boston became very good at building this type of company.
HubSpot built one of the world’s best-known customer platforms. Klaviyo created a major marketing and customer-data business. Toast built a restaurant operating platform around payments, ordering, payroll, marketing, and other daily tasks. Rapid7 became a major cybersecurity company, while PTC became deeply embedded in industrial software.
These companies were built around the same basic idea: software helps people do their jobs better.
Artificial intelligence is starting to change that relationship.
The most important question is no longer simply, “What software should an employee use?”
Increasingly, companies are asking a different question:
“What part of the employee’s work can the software complete by itself?”
That may sound like a small change, but it has enormous consequences.
It can change how software is designed, how companies charge for it, how customers measure value, how teams are structured, and how much time users actually spend inside an application.
Boston is becoming one of the most interesting places to watch this transition.
New companies such as 7AI, Day AI, Maven AGI, OpenHands, EnFi, and Code Metal are being built around AI systems that investigate, organize, analyze, code, or complete work. At the same time, established Boston software companies such as HubSpot, Klaviyo, and Toast are rebuilding parts of their products around agents, automation, and AI-driven workflows.
The result is not the death of SaaS.
Something more important is happening.
SaaS itself is beginning to move toward AI-native software.
That shift could reshape Boston’s software industry over the next decade.
The Short Version: Software Is Moving From Helping People Work to Doing More of the Work
Traditional SaaS gives employees tools.
AI-native software goes further by taking responsibility for parts of the workflow itself.
A traditional CRM helps salespeople store contacts, update deals, record conversations, and review reports. An AI-native CRM can begin collecting information automatically, understanding customer conversations, identifying what changed, and deciding what should happen next.
A traditional security platform produces alerts for analysts to investigate. An AI security agent can examine the alert, collect supporting evidence, compare it with other activity, summarize what happened, and recommend or carry out the next action.
A normal customer-service system helps agents manage tickets. An AI-native system can attempt to solve the customer’s problem before a human needs to become involved.
Developer software once focused mainly on helping programmers write code faster. Coding agents are beginning to take complete software tasks, modify files, run tests, fix problems, and return completed work.
Commercial lending software has traditionally helped teams organize applications and records. AI systems can now analyze parts of the deal, identify risks, summarize evidence, and prepare material for human review.
The common pattern is easy to see.
Software is moving through several stages.
| Software era | Main promise |
| On-premise software | Digitize the process |
| SaaS | Make the process easier to manage |
| SaaS with AI assistants | Help the employee complete the process |
| AI-native software | Complete larger parts of the process |
| Agentic software | Take responsibility for a defined outcome |
The difference between those final stages matters because it changes the economic unit of software.
If an application completes work instead of simply giving employees access to a tool, charging by employee seat may no longer make sense.
If one employee can supervise ten AI agents, a vendor cannot assume that customer growth will always mean more human licenses.
If the customer cares mainly about completed investigations, resolved tickets, processed documents, or finished coding tasks, the software company may eventually need to price around those outcomes.
That is why AI-native software is much more than SaaS with a chatbot added to the side of the screen.
It changes what customers are buying.
What Does “AI-Native” Actually Mean?
The term “AI-native” is already becoming overused.
Almost every software business now has some form of AI feature. Some products summarize information. Others generate text, answer questions, recommend actions, or create reports.
Those features can be useful, but they do not automatically make a product AI-native.
For this article, we use a stricter definition.
AI-Native Software Is Built Around Something That AI Makes Newly Possible
An AI-native product would either stop working or lose most of its value if the AI capability were removed.
That is different from traditional SaaS with several AI features added later.
Imagine a project-management platform.
The product may already allow employees to create tasks, assign owners, set deadlines, upload files, and track progress. If the company adds a button that summarizes project notes, the software becomes more useful, but the underlying product has not changed much.
Now imagine another system that watches project activity, identifies dependencies, notices missed deadlines, updates schedules, assigns routine tasks, collects missing information, detects delivery risk, and asks managers for help only when a decision requires human judgment.
That is not merely a better project-management interface.
The software has started participating in the operation of the project.
That is much closer to AI-native design.
A Simple Test: Remove the AI
One of the easiest ways to evaluate a product is to ask:
If the AI disappeared tomorrow, would the product still make sense?
For many traditional SaaS products, the answer is yes.
The core application would continue to work. Some useful features might disappear, but customers could still use the software.
For a true AI-native product, removing the intelligence may destroy the central value of the product.
An autonomous coding agent without AI would no longer be an autonomous coding agent.
A cybersecurity investigation agent without AI would not be able to perform investigations.
An AI-native CRM that automatically understands business activity would lose the feature that makes it different from traditional CRM systems.
This distinction is much more useful than simply asking whether a company uses artificial intelligence.
Original Research: How We Analyzed Boston’s AI-Native Shift
To understand what is actually happening in Boston, BostonBizScene created a small original dataset for this article.
The goal was not to count every software company in Massachusetts. Doing that accurately would require proprietary databases, private company information, and a much larger research project.
Instead, we started with a defined sample and applied the same classification rules to every company.
Dataset One: 25 Greater Boston AI Startups
We began with BostonBizScene’s earlier research into 25 Greater Boston AI startups active in 2026.
The sample included companies from enterprise software, cybersecurity, healthcare, biology, robotics, infrastructure, developer tools, scientific AI, and other categories.
For the current analysis, we asked a different question:
What is the company’s main product actually designed to do?
We divided the companies into two broad groups.
The first group contains companies whose primary product is digital software.
The second group contains companies where AI is closely tied to science, medicine, machines, factories, biology, satellites, robotics, or other physical systems.
That distinction gives us a better view of Boston’s AI economy than simply placing every company into one large software category.
Original Finding #1: Only 36% of the Sample Looks Like Primarily Digital Software
Our analysis produced the following split.
Chart 1: Product Architecture Across 25 Boston AI Companies
| Product architecture | Companies | Share |
| Primarily digital software | 9 | 36% |
| AI tied to science, health, or physical systems | 16 | 64% |
| Total | 25 | 100% |
The digital-software group includes Code Metal, 7AI, Maven AGI, Day AI, OpenHands, EnFi, Gradient AI, Rindler, and Suno.
The science, healthcare, and physical-systems group includes Lila Sciences, Physical Superintelligence, Osmo, Axiomatic AI, Walden Robotics, Tomorrow.io, Blue Water Autonomy, Reframe Systems, Dipole Labs, Beacon Biosignals, Clairity, Somite.ai, AI Proteins, Videa, Asimov, and Dyno Therapeutics.
This does not mean one model is better than the other.
Instead, the result shows something important about Boston.
Boston’s AI economy extends far beyond traditional SaaS.
Nearly two-thirds of the companies in this sample are building AI around scientific, medical, industrial, biological, or physical systems.
That fits Boston’s broader economic structure.
The region has major universities, research hospitals, biotechnology companies, robotics teams, engineering talent, defense contractors, financial institutions, software businesses, and research laboratories.
Boston therefore enters the AI-native era from a different starting point than a technology market built mainly around consumer apps or web software.
It can build intelligent software, but it can also connect that intelligence to difficult work in the physical world.
Original Finding #2: Task Execution Dominates Boston’s Digital AI Startup Sample
We then looked only at the nine companies in the primarily digital-software group.
Each company was classified according to the main value its product appears to produce.
Chart 2: What Boston’s Digital AI Startups Primarily Produce
| Primary output | Companies | Share of digital sample |
| Task execution or workflow completion | 6 | 66.7% |
| Prediction or scoring | 1 | 11.1% |
| Generative output | 1 | 11.1% |
| Agent infrastructure | 1 | 11.1% |
| Total | 9 | 100% |
The six companies placed in the task-execution group are Code Metal, 7AI, Maven AGI, Day AI, OpenHands, and EnFi.
The finding matters because these companies are not simply trying to make information easier to access.
They are moving deeper into the workflow.
The product may investigate a security event, manage customer information, complete a coding task, analyze a lending opportunity, or handle part of a customer-service process.
That does not mean the human disappears.
It means the software starts doing more before the human needs to become involved.
The Shift Is Bigger Than the Word “Agent”
The technology industry currently uses the word “agent” constantly.
That label may eventually change.
The more important transition will remain.
Software is moving from:
information → recommendation → execution.
A system that only provides information leaves almost all responsibility with the worker.
A system that recommends an action begins sharing some of the thinking.
A system that safely carries out the action starts taking responsibility for part of the work.
That final step is what software companies need to watch.
Traditional SaaS Was Designed Around Human Attention
Traditional SaaS quietly depends on a simple assumption.
A person is operating the product.
The employee logs in.
The employee opens the relevant page.
The employee enters information.
The employee reads the report.
The employee decides what should happen.
The employee clicks the button.
Software makes the process faster, but the responsibility for moving it forward stays largely with the person.
That design naturally supports seat-based pricing.
If twenty employees use the system, the customer buys twenty licenses.
As the company hires more people, the software company can sell more seats.
Artificial intelligence creates a problem for that model.
What Happens When One Worker Supervises Many Agents?
Imagine a security team with ten analysts.
A traditional security platform might sell ten user licenses.
Now imagine that an AI system handles much of the routine investigation work. Instead of ten analysts doing every investigation manually, four analysts supervise a larger number of automated workflows.
The customer becomes more productive.
However, the number of human software seats falls.
If the vendor charges only by seat, the software company could help the customer reduce headcount while also reducing its own revenue.
That creates a pricing problem.
A more natural pricing system may be based on the actual work completed.
The vendor could charge for investigations, cases, transactions, workflows, credits, resolved problems, actions, or another measurable unit.
That is one reason AI-native software could change SaaS economics.
Boston’s Established SaaS Companies Are Changing Too
The rise of AI-native startups does not mean established SaaS businesses are standing still.
Some of Boston’s largest software companies are already changing how their products work.
That is important because incumbents have something startups often lack: huge amounts of customer data, existing workflows, integrations, distribution, and trust.
The future may therefore be shaped by both new AI-native companies and existing SaaS companies that successfully rebuild themselves.
HubSpot Is Turning CRM Data Into an Agent Platform
HubSpot grew by helping companies manage marketing, sales, service, and customer relationships.
Its original model was familiar SaaS.
Employees logged into the software, worked with customer data, managed campaigns, updated deals, and used reports to make decisions.
That model still exists.
However, HubSpot is increasingly adding agents that can work across its customer-data system.
The important strategic point is not simply that HubSpot now offers AI.
The real advantage comes from the data beneath the AI.
CRM Data Gives Agents Context
A general AI model may understand sales.
It does not automatically understand your sales process.
A CRM contains that missing context.
It knows which company the salesperson is talking to.
It knows what products the customer has purchased.
It knows what happened during previous conversations.
It knows which stage the deal has reached.
It may know what the customer asked for, who owns the account, and what happened the last time someone followed up.
When agents can use that information, the CRM database becomes more than a record-keeping system.
It becomes operational context.
That gives established SaaS companies an important advantage.
Existing SaaS Companies Already Own Valuable Context
HubSpot owns customer and sales context.
Toast owns restaurant and transaction context.
Klaviyo owns marketing and customer-behavior context.
Other software companies may own financial, employee, logistics, security, or project data.
As AI systems become better at taking action, that existing information can become much more valuable.
The result is not simply startups replacing SaaS companies.
It is SaaS companies turning their systems of record into systems that can understand and act.
Klaviyo Shows What Happens When SaaS Becomes More Autonomous
Klaviyo provides another strong Boston example.
The company grew around marketing automation and customer data.
Traditional marketing software helps teams build campaigns, segment customers, prepare messages, review performance, and manage workflows.
AI changes how much of that work the software can handle.
Klaviyo has introduced agentic systems designed to create campaigns, support customers, and work across customer information.
The company has also moved toward allowing customers to interact with its capabilities through other AI interfaces.
That may sound like a technical change, but it could reshape the relationship between users and software.
SaaS Used to Own the Screen
Software companies historically wanted users inside their applications.
They spent years improving navigation, dashboards, onboarding, mobile experiences, and product engagement.
A user who logged in frequently was often considered a healthy user.
AI creates a different possibility.
A customer may receive more value while spending less time inside the application.
Imagine a marketing manager asking an AI system:
“Find customers whose repeat-purchase rate is falling, identify the most likely cause, create three recovery campaigns, estimate the expected value of each campaign, and prepare the safest option for approval.”
The manager may still depend heavily on the marketing platform.
However, the user may not need to click through six screens to complete the task.
The software remains important even when the interface becomes less visible.
Applications Can Become Capabilities
Traditional SaaS often follows this path:
Human → application → data → action
AI-native software can create a different path:
Human → agent → multiple systems → action
The application does not disappear.
Instead, its capabilities become available through an intelligent layer that can work across several products.
This is why APIs, tool connections, agent protocols, and structured access to software functions are becoming more important.
The software company may no longer control the first screen the user sees.
It may still control an essential capability underneath that screen.
Toast Shows Why Vertical SaaS Could Be Extremely Hard to Replace
Toast offers another useful lesson.
Restaurant software is not simply a dashboard.
It touches payments, ordering, menus, staff, customers, payroll, marketing, and other daily operations.
That creates deep operational context.
A new AI startup can build a restaurant assistant quickly.
Replacing an operating system deeply connected to payments and daily workflows is much harder.
Vertical SaaS Has a Hidden Defense
Vertical SaaS companies often have years of specialized data and integrations.
They understand how a particular industry works.
They know the exceptions.
They know the messy details.
They may also control transactions, permissions, or systems that cannot easily be replaced.
Artificial intelligence can make these businesses stronger because agents need exactly that kind of context.
A restaurant agent that understands orders, customers, staffing, menus, payments, and marketing can become much more useful than a general chatbot.
The same idea applies to healthcare, insurance, finance, logistics, manufacturing, construction, and other specialized industries.
This is why some of the strongest AI-native companies of the next decade may grow out of existing vertical SaaS businesses.
SaaS Is Not Dying, but Weak SaaS Is Becoming Easier to Attack
Claims that AI will “kill SaaS” are too broad to be useful.
Some software companies have strong defenses.
Others are much more exposed.
Administrative SaaS Is Particularly Vulnerable
Consider software whose main job is helping employees perform repetitive information work.
The user may need to:
enter data,
copy information,
prepare standard reports,
classify documents,
search internal knowledge,
update records,
schedule tasks,
or move information between different applications.
These workflows are vulnerable because AI can increasingly perform much of the activity directly.
If an employee spends three hours every week operating software simply to keep a process moving, an AI-native competitor can ask a powerful question:
Why does the employee need to operate this workflow at all?
That is the challenge traditional SaaS companies need to answer.
Strong SaaS Companies Usually Own Something Difficult
The safest software companies often control assets that are hard to copy.
They may own a system of record.
They may control payments.
They may have deep industry integrations.
They may understand regulated workflows.
They may have proprietary data.
They may hold transaction history.
They may be tied to physical hardware.
They may have difficult enterprise permissions.
They may have an ecosystem of partners and customers.
Artificial intelligence can increase the value of those assets rather than destroy them.
Boston Funding Shows That Investors Are Already Backing AI-Driven Software
The AI-native transition is not only visible in product design.
Capital is also moving toward the sector.
BostonBizScene’s 2026 AI funding tracker identified at least $2.373 billion across 36 high-confidence Greater Boston AI financing events through September 22.
Twelve enterprise and developer-AI financings represented roughly $553 million.
Four security and observability financings represented another $355 million.
Together, these two software-heavy categories accounted for about $908 million.
Chart 3: Software-Heavy Categories in Boston’s 2026 AI Funding Dataset
| Category | Financings | Capital |
| Enterprise and developer AI | 12 | $553M |
| Security and observability | 4 | $355M |
| Combined | 16 | $908M |
| Total tracked Boston AI funding | 36 | ≥$2.373B |
| Combined share of tracked capital | — | ~38.3% |
The figure should not be interpreted as the share of all Boston technology investment going into AI-native SaaS.
The funding categories contain several different business models, and the dataset includes only financing events that met our confidence threshold.
Even with those limits, the pattern is clear.
Software remains a major part of Boston’s AI investment market.
The more interesting question is what type of software investors are now backing.
Many companies are being built around automation, agents, developer workflows, cybersecurity operations, and other forms of work that move beyond traditional software interfaces.
Original Finding #3: AI Labels Now Outnumber SaaS Labels in YC’s Boston Directory
Y Combinator’s Boston directory provides another useful signal.
As of September 2026, the directory lists more Boston companies under artificial intelligence than under SaaS.
Chart 4: YC Boston Directory Labels
| YC Boston category | Companies listed |
| Artificial intelligence | 22 |
| SaaS | 9 |
That means AI-labelled companies outnumber SaaS-labelled companies by roughly 2.4 to 1.
This is not a perfect market census.
Some companies appear under more than one category. A company can be both AI and SaaS, and the YC directory covers only YC-backed startups.
However, the labels still tell us something about how founders increasingly describe their companies.
Ten years ago, “SaaS” itself was often a major part of the startup’s identity.
Today, founders are more likely to describe the intelligence, automation, or work the product performs.
That change in language reflects a deeper change in product strategy.
The SaaS Dashboard Is Losing Its Monopoly
One of the most important changes happening in software has little to do with raw model intelligence.
It is the interface.
Traditional SaaS organizes the world around pages.
Contacts.
Customers.
Reports.
Projects.
Tickets.
Campaigns.
Settings.
Users learn where each function lives and navigate between screens to complete a task.
AI starts from a different place.
It starts with intent.
A user can say:
“What happened?”
“Find the problem.”
“Prepare the campaign.”
“Investigate this alert.”
“Update the customer.”
“Build the feature.”
“Tell me what needs attention.”
The software then decides which information and tools are required.
Intent Can Replace Navigation
If the system understands the user’s goal, the user does not always need to know which page contains the required function.
That changes which parts of software become most valuable.
Navigation becomes less important.
Data quality becomes more important.
Permissions become more important.
APIs become more important.
Historical context becomes more important.
Audit trails become more important.
The visual interface remains useful, but it may no longer be the only place where the work happens.
AI-Native Software Changes the Meaning of Good User Experience
Traditional user-experience design often asks how to make an action easier.
Can a twelve-click process become a five-click process?
Can a form become shorter?
Can navigation become clearer?
AI-native design asks a more radical question.
Can we remove the action from the human workflow entirely?
Suppose an employee submits an expense.
Traditional software might make reimbursement faster by improving the form and reducing the number of screens.
An AI-native system could automatically recognize the receipt, match it to the transaction, check the expense policy, prepare the reimbursement, and contact the employee only if something looks unusual.
The process has not simply become easier.
Most of the process has disappeared from the user’s view.
Exceptions Could Become the New Dashboard
This creates a new interface model.
Instead of showing users every workflow, software may increasingly show only the situations that require judgment.
What failed?
What needs approval?
What looks unusual?
What exceeds the policy?
Where is the AI uncertain?
Which decision has financial or legal risk?
Routine work can remain mostly invisible.
Humans can focus on exceptions.
For many enterprise applications, that could become the dominant design pattern of the AI-native era.
Time Spent in the Product May Become a Misleading Metric
Software companies have traditionally treated engagement as a positive signal.
If users spend more time inside the product, the assumption is often that the product has become more valuable.
That logic becomes weaker when software starts doing the work.
Suppose a manager currently spends five hours every week inside a payroll system.
After introducing AI, the same manager spends only fifteen minutes reviewing unusual cases.
Time in the application falls dramatically.
Yet the value of the software has increased.
That means software companies need a different set of metrics.
A Better KPI Dashboard for AI-Native Software
| Traditional SaaS metric | AI-native equivalent |
| Seats | Workloads served |
| Daily active users | Tasks or outcomes completed |
| Time in application | Human time removed |
| Feature adoption | Workflow penetration |
| Clicks | Successful actions |
| Messages generated | Problems resolved |
| Reports viewed | Decisions improved |
| Tickets handled | Tickets resolved |
| Software utilization | Automation rate |
| Expansion seats | Expansion workflows |
| Cost per seat | Cost per outcome |
| User productivity | Economic value created |
Traditional SaaS metrics do not suddenly become useless.
However, they no longer tell the whole story.
If the software is doing more work, companies must measure the work.
Pricing Could Become One of AI Software’s Hardest Problems
Traditional SaaS pricing became relatively predictable.
Companies charged by seat, plan, storage, or usage.
AI-native software creates new complexity because AI actions have real variable costs.
Inference costs money.
Search costs money.
Code execution costs money.
External APIs cost money.
Long-running agents can consume much more infrastructure than a short AI answer.
That changes software economics.
Seat Pricing May No Longer Match Customer Value
Imagine an AI product that removes half of an employee’s repetitive workload.
Charging $30 per employee could badly underprice the value.
At the same time, if the software runs expensive models constantly, that same $30 price might not produce attractive margins.
AI companies therefore need to align three things:
customer value,
product usage,
and infrastructure cost.
That is difficult.
Why Credit Systems Are Becoming More Common
Credits provide one possible solution.
The customer may still pay a predictable subscription fee.
However, AI-heavy work consumes additional credits based on the amount or type of work performed.
This gives vendors more flexibility than pure seat pricing while still providing customers with some predictability.
Other companies will use different models.
Some will price by completed task.
Some will price by transaction.
Some will charge by resolved case.
Others may price according to the economic value the system creates.
There is unlikely to be one universal model.
AI-Native Companies Need to Measure Gross Margin Differently
Classic SaaS companies became famous for very high gross margins.
AI-native software can be more complicated.
A traditional page load costs almost nothing.
An agent that spends several minutes researching, reasoning, searching, calling tools, and checking its work can cost far more.
That means AI software teams should stop focusing only on raw model cost.
They should measure the cost of producing a successful business outcome.
Measure Cost Per Successful Outcome
Consider two hypothetical models.
Model A costs $0.08 per attempt but succeeds only 60% of the time.
Model B costs $0.15 per attempt but succeeds 97% of the time.
If the company looks only at inference price, Model A appears cheaper.
If failed attempts require retries, human review, or customer corrections, Model B may actually produce better economics.
The useful metric is not simply:
cost per model call.
It is:
cost per successful task.
That is a much better way to manage AI economics.
Boston Has an Advantage Because Its Industries Contain Expensive Work
AI does not need to automate cheap tasks to create a valuable company.
In many cases, the best opportunities may involve expensive work performed by highly skilled professionals.
Boston has a large amount of that work.
Commercial lending.
Cybersecurity.
Healthcare.
Biotechnology.
Engineering.
Insurance.
Scientific research.
Drug development.
Enterprise software development.
Industrial operations.
These workflows often cost far more than simple administrative work.
That creates more room for AI vendors to create value.
Expensive Problems Can Support More Expensive AI
Imagine a company spends $100,000 each year on a repetitive analytical process.
If an AI system can reduce that cost by $40,000 while maintaining quality and control, the vendor does not need the underlying inference to cost almost nothing.
The product can spend meaningful amounts on compute and still create attractive economics.
That matters in Boston because many of the region’s strongest industries already contain difficult, expensive workflows.
Boston’s Strongest Moat May Be Domain Knowledge
Most Boston software startups will not build a foundation model that is better than every major AI lab.
Fortunately, they do not need to.
A startup can create a strong advantage by understanding a specific workflow better than anyone else.
Consider insurance.
Insurance contains specialized language, rules, historical information, risk models, and unusual workflows.
A generic model may understand insurance concepts.
A system connected to years of insurance data, validated decisions, domain rules, workflow history, and real outcomes can become much more useful.
The same idea applies to healthcare.
Finance.
Cybersecurity.
Manufacturing.
Biotechnology.
Scientific research.
Enterprise software.
The model itself may be widely available.
The workflow knowledge is not.
The Data Flywheel Is Changing
Traditional SaaS companies usually collected data because employees entered it.
AI-native software can also create valuable data by performing work.
A security agent investigates an incident.
That creates a record of what it checked and what it found.
A lending agent analyzes a deal.
That creates evidence around the decision.
A coding agent modifies software.
That produces tests, failures, patches, and outcomes.
A customer-service agent tries to solve a problem.
That creates data about which actions worked and which did not.
This makes completed work part of the company’s learning system.
Completed Work Can Become a Feedback Loop
The pattern looks like this:
AI acts → outcome occurs → outcome is measured → system improves → future action becomes better
That feedback loop could become one of the strongest competitive advantages in AI software.
Access to a public model is not difficult to copy.
Years of workflow history and measured outcomes are much harder to reproduce.
Autonomy Creates a Permission Problem
Giving software better intelligence is one challenge.
Giving it permission is another.
A chatbot can write almost anything.
A business cannot safely allow an automated system to do everything.
Imagine an AI that can:
refund customers,
change contracts,
approve credit,
delete production data,
modify software,
contact clients,
move money,
change payroll,
or alter medical workflows.
Every one of those actions creates risk.
Businesses need to know what the agent is allowed to do and when.
They also need answers to basic questions.
Who authorized the action?
What information did the agent use?
Which systems did it access?
What limits were in place?
Why did it choose that action?
Can the decision be reversed?
Who reviews unusual cases?
These questions create a new software layer around AI.
The Next SaaS Category May Manage AI Workers
As companies operate more AI agents, they will need tools to manage those agents.
That could create major new software categories.
Companies may need systems for:
identity,
permissions,
budgets,
monitoring,
logging,
evaluation,
routing,
security,
escalation,
and model selection.
This creates an interesting irony.
AI-native systems may reduce the need for some traditional SaaS products while creating entirely new SaaS markets.
The difference is that the users of those systems may increasingly include software agents as well as humans.
Established SaaS Companies Still Have a Major Advantage: Distribution
Startups often focus heavily on product quality.
That matters.
However, enterprise distribution is extremely difficult.
An AI agent must still pass a security review.
It still needs data access.
It needs integrations.
It may need procurement approval.
It needs implementation support.
It has to earn trust.
Established SaaS companies already have customer relationships.
They already have integrations.
They often already sit inside important workflows.
That gives them a powerful channel for distributing AI.
Existing Customers Can Accelerate AI Adoption
When an established software company adds a useful agent to a trusted platform, the customer does not need to adopt an entirely new vendor.
The system may already contain the data and permissions required.
That can shorten the path from experiment to production.
This is one of the strongest advantages incumbent SaaS companies have over new AI startups.
Startups Have a Different Advantage: They Can Remove the Old Workflow
Incumbents also have weaknesses.
They have existing products.
Existing customers.
Existing pricing.
Existing architecture.
Existing revenue.
Existing sales compensation.
Existing expectations.
A startup has less to protect.
That allows it to redesign the workflow from the beginning.
Do Not Ask Only, “How Can We Add AI?”
A stronger question is:
“If this workflow were created today, knowing AI exists, what would we remove?”
Would users still fill in every field?
Would they still build the report manually?
Would they still route each ticket?
Would they still move information between six systems?
Would they still prepare the same weekly summary?
Would they still need five separate applications?
Those questions lead to more ambitious products.
A Simple AI-Native Test for Boston Founders
Boston founders can evaluate their products using five questions.
1. Does the Product Own Work or Mainly Display Information?
Displaying information can still be valuable.
However, information is becoming easier to generate.
Products that own a workflow can create a deeper relationship with the customer.
2. Can the Software Understand What Should Happen Next?
If every action requires a human command, the software remains primarily a tool.
If the system can recognize the next safe step, it begins participating in the work.
3. Can It Perform the Action?
Intelligence without tools creates advice.
Tools without intelligence create automation.
Combining both creates something much closer to an agent.
4. Can Success Be Measured?
A company should be able to measure what the software actually improves.
That could mean money saved, revenue generated, time removed, cases completed, errors prevented, or another clear business result.
5. Does Usage Improve the Product?
If every competitor can use the same models and the same public data, the product will be easier to copy.
A proprietary feedback loop creates a stronger defense.
What Boston SaaS Companies Should Do Over the Next 90 Days
Existing software companies do not need to rebuild everything immediately.
They should begin by identifying where customers still perform unnecessary manual work.
Map the Ten Most Common Workflows
Do not start with a feature list.
Start with jobs.
For every important workflow, document what triggers it, what information is required, what decisions happen, which actions follow, and where human judgment is actually necessary.
This makes the automation opportunity much easier to see.
Measure Human Touches
Count how many times a person becomes involved.
Suppose a customer-support process requires twelve human steps.
Perhaps only three of those steps involve meaningful judgment.
The other nine may eventually be automated.
That insight is more useful than simply adding another AI button to the interface.
Identify Your System-of-Record Advantage
Ask what information your company owns that a generic AI model does not.
Customer history?
Transactions?
Workflow data?
Permissions?
Product telemetry?
Industry-specific data?
That context can become the foundation of a much stronger agent.
Build Around One Outcome
A common mistake is adding AI to many parts of the product at once.
Summaries here.
Text generation there.
Chat somewhere else.
That creates features but not necessarily a new product model.
Instead, choose one meaningful outcome.
For example:
Resolve 40% of routine customer-service requests without human handling.
That forces the company to solve data, permissions, actions, accuracy, monitoring, and economics as one complete system.
The AI-Native KPI Dashboard Boston Founders Should Build
| KPI | Why it matters |
| Task success rate | Shows whether the system actually works |
| Human intervention rate | Measures how much work remains manual |
| Cost per completed task | Links AI spending to business economics |
| Time to outcome | Measures real productivity improvement |
| Error rate | Shows reliability |
| Escalation accuracy | Tests whether the AI knows when to stop |
| Gross margin per workflow | Reveals whether automation is economically healthy |
| Repeat workflow rate | Measures recurring customer value |
| Customer correction rate | Exposes hidden quality problems |
| Expansion workflows | Shows whether customers trust the system with more work |
| Economic value created | Connects software activity with ROI |
| Model dependency | Measures exposure to outside AI providers |
A software company that cannot measure completed work will have difficulty pricing completed work.
What Boston Buyers Should Ask AI Vendors
Customers need to change how they evaluate software too.
Traditional software buying often focused on features.
AI software should be evaluated around performance and control.
Ask About Failure
Do not only watch a polished demo.
Ask how often the agent fails.
Ask what failure means.
Ask how the system detects the problem.
Ask what happens next.
Ask whether the action can be reversed.
Compare the AI With the Existing Human Process
If the system completes a task in twenty seconds, that sounds impressive.
However, the business needs a baseline.
If a human already completes the task in forty seconds, the improvement may not matter.
If the task normally takes two hours of specialist time, the economics look much stronger.
Understand Permissions
A system that can read data carries one level of risk.
A system that can change data carries another.
A system that can approve a payment, contact a customer, or modify production software carries far more risk.
Permissions should expand as trust grows.
Understand the Vendor’s Pricing Logic
AI pricing can change quickly as usage grows.
Businesses should understand what happens when usage increases.
Will the bill increase at the same rate?
Does more automation mean more cost?
Does the price depend on tokens, tasks, actions, or results?
Understanding the economics early can prevent expensive surprises later.
Massachusetts Is Building Public Infrastructure Around AI
The transition is also receiving support outside private companies.
Massachusetts has created public programs aimed at AI adoption, research, infrastructure, entrepreneurship, and workforce development.
That matters because the next generation of AI-native companies will require more than software engineering.
They need compute.
They need data.
They need research talent.
They need industry experts.
They need customers.
They need governance.
They need workers who understand how to redesign processes around intelligent systems.
Boston already has many of those pieces.
Original Finding #4: Boston’s Best AI Opportunity May Be Where SaaS Becomes Vertical
Horizontal SaaS became successful by solving broad business functions.
CRM.
Marketing.
HR.
Accounting.
Project management.
AI creates an especially strong opportunity in vertical markets because models can now understand information that older software struggled to process.
Medical notes.
Insurance documents.
Security alerts.
Engineering files.
Customer calls.
Credit applications.
Scientific papers.
Legacy code.
Images.
Audio.
The software can move from organizing the workflow to understanding the workflow.
Once it understands the workflow, it can begin participating in the workflow.
Vertical SaaS Can Become Vertical Intelligence
This transition is already visible across Boston.
Gradient AI focuses on insurance.
EnFi focuses on commercial lending.
Videa focuses on dentistry.
Code Metal focuses on difficult engineering software.
Other Boston companies are applying AI to healthcare, science, manufacturing, biotech, and industrial systems.
This model fits Boston extremely well because the region already has deep domain expertise.
The Real Battle Is Not AI-Native Versus SaaS
The title of this article creates a useful contrast, but the market is more complicated.
AI-native software and SaaS are not opposites.
A company can sell a subscription while being completely AI-native.
The real difference is between two product philosophies.
The Traditional SaaS Philosophy
Give customers software.
Make it easier to use.
Add more features.
Sell more seats.
Increase engagement.
The AI-Native Philosophy
Understand the customer’s goal.
Complete as much of the work as possible.
Ask for human judgment when needed.
Measure the outcome.
Charge in a way that matches the value created.
The strongest software companies may combine both models.
They can keep the recurring revenue, reliability, integrations, and enterprise controls of SaaS while adopting the execution model of AI-native products.
Software Teams Will Need to Change Too
This transition affects more than the product.
Product managers traditionally design workflows for humans.
Now they must design workflows shared between humans and agents.
Engineers are used to software that behaves predictably.
AI introduces uncertainty.
Designers traditionally focus on interfaces.
They now need to design approval, trust, review, and escalation.
Finance teams are used to predictable cloud costs.
AI workloads can vary significantly.
Sales teams traditionally sell features and seats.
They may increasingly sell economic outcomes.
Customer-success teams once focused on teaching customers how to use software.
They may increasingly help customers redesign workflows around automation.
This is not simply an AI feature project.
It is an organizational change.
Boston Founders Should Avoid Building Another Generic AI Wrapper
The easiest AI product to build is often the easiest one to copy.
A startup can connect to a model API, add a prompt, create a polished interface, and launch quickly.
That may work for a period of time.
However, the competitive barrier is usually weak.
Another startup can use the same model.
The model provider can build the same feature.
An incumbent can copy the workflow.
Some customers may be able to reproduce the result themselves.
Boston has better opportunities.
The region is surrounded by hard problems, specialized data, difficult industries, scientific work, regulations, and complex systems.
Those challenges are harder to solve.
That is exactly what can make the resulting companies more defensible.
Where Boston’s Most Durable AI Companies Could Emerge
Several categories appear especially suited to the region.
Cybersecurity
Security contains enormous amounts of repetitive analytical work.
However, mistakes matter.
That makes the market attractive for agents that combine automation with strong controls.
Healthcare
Healthcare contains large volumes of unstructured information and expensive specialist work.
AI can create major value, but safety and evidence requirements make the category difficult.
That difficulty can become a barrier to entry.
Life Sciences
Boston already has researchers, labs, hospitals, investors, and biotechnology companies.
That makes the region well positioned to connect AI with biological data and experimental workflows.
Insurance and Finance
Both industries contain large amounts of repetitive document analysis combined with strict rules and expensive decisions.
That creates opportunities for AI systems that can assist with or automate parts of professional workflows.
Industrial Software
Boston’s engineering and robotics base creates opportunities where AI must interact with machines, factories, sensors, and physical processes.
Developer Infrastructure
As companies operate larger numbers of agents, they will need tools for testing, orchestration, observability, security, identity, and permission management.
These markets are harder than generic productivity software.
That difficulty is one reason they may produce stronger companies.
The Bigger Story: SaaS Is Becoming Infrastructure for Digital Workers
For years, SaaS companies built software primarily for humans.
That is beginning to change.
Customer data may increasingly be read by agents.
Marketing systems may increasingly be operated by agents.
Developer tools may increasingly coordinate coding agents.
Security software may increasingly manage investigation agents.
Commerce systems may increasingly receive requests that begin inside an AI assistant rather than a traditional website.
The role of SaaS therefore starts to change.
The software becomes less like a destination and more like infrastructure.
That could be one of the most important changes in enterprise technology over the next decade.
The Best SaaS Platform May Be the One Agents Can Use
For years, software companies asked:
How do we get more people into our product?
The next question may be:
How do we make our data and capabilities available wherever people and agents already work?
That changes software architecture.
It changes interface design.
It changes pricing.
It changes what companies measure.
And it changes the meaning of product engagement.
Boston’s largest SaaS companies are already beginning to move in this direction.
That should matter to every software leader in the region.
What Boston Software Leaders Should Do Now
Do not spend too much time asking whether AI will replace SaaS.
The question is too broad.
Instead, examine your product and identify every place where humans still perform repetitive work simply because older software required them to.
Look for activities such as reading, copying, searching, classifying, summarizing, routing, checking, entering, and updating.
Then ask which of those actions truly require human judgment.
Do not automate something simply because the technology can do it.
Automate where the economics are clear, the process is measurable, and the risk is manageable.
Keep humans involved when decisions are expensive, sensitive, or difficult to reverse.
Measure what happens.
Then expand gradually.
A 90-Day AI-Native Transition Plan
Days 1–30: Map the Work
Choose three important workflows.
Measure the current time, cost, error rate, and business value.
Break each workflow into actions and decisions.
Identify which steps require genuine human judgment.
Do not begin with model selection.
Begin with the process.
Days 31–60: Automate One Narrow Outcome
Choose one repeated, measurable task.
Give the system only the permissions it needs.
Create clear escalation rules.
Compare its performance with the existing process.
Do not judge the pilot by how impressive the conversation feels.
Judge it by completed work.
Days 61–90: Test the Economics
Calculate the AI infrastructure cost.
Add the cost of human review.
Include failure costs.
Include implementation work.
Then compare those costs with the time saved, revenue created, and errors prevented.
At the end of the pilot, answer one question:
Does the workflow become economically better as the AI does more of the work?
If the answer is yes, expand carefully.
If the answer is no, understand the problem before adding more AI.
Final Takeaway: Boston’s Software Industry Is Moving From Selling Tools to Selling Work
The most important change in software is not that every application now includes an AI button.
The deeper change is that software is beginning to take responsibility for larger parts of the job.
Traditional SaaS gave employees better tools.
AI-native systems increasingly help complete the work itself.
Our analysis of 25 Boston AI companies shows how broad that transition has become.
Only nine of the 25 companies in our sample, or 36%, fit primarily into the digital-software category. The remaining 64% connect AI with science, healthcare, biology, machines, or other physical systems.
Within the nine-company digital-software group, six companies, or roughly two-thirds, are primarily centered on task execution or workflow completion rather than simple prediction or information generation.
Boston’s funding market points in a similar direction.
Enterprise and developer AI, together with security and observability, accounted for approximately $908 million of the $2.373 billion in high-confidence Greater Boston AI financing tracked by BostonBizScene through September 22, 2026.
At the same time, Boston’s established software companies are changing.
HubSpot is building agents around CRM workflows.
Klaviyo is pushing toward more autonomous marketing and customer operations.
Toast is adding intelligence directly to restaurant workflows.
The divide is therefore not old companies against startups.
It is not even SaaS against AI.
The more important divide is between software that waits for a human to operate it and software that increasingly understands what needs to happen next.
That distinction will affect pricing.
It will affect user interfaces.
It will affect gross margins.
It will affect staffing.
It will affect product design.
It will affect how customers buy software.
It will affect which companies own the workflow.
The next great Boston software company may still charge a subscription.
It may still have a dashboard.
It may still describe itself as SaaS.
But the reason customers value the product could be completely different.
They may not love it because employees spend hours using it.
They may love it because the software gets more of the work done.
