New York’s next big AI story is not another chatbot.
It is software that actually does the work.
Across Manhattan and the wider New York tech market, a new group of AI companies is building software that researches companies, reviews legal rules, completes accounting work, processes patient referrals, talks to customers, evaluates financial decisions, analyzes investment opportunities, and even learns how to move through the physical world.
These products are usually called AI agents. The name can make them sound more futuristic than they really are. At the most useful level, an AI agent is simply software that can understand a goal, decide what needs to happen next, use tools or data, complete several steps, and return a result without asking a person to control every move.
The important word is not AI.
It is work.
New York is becoming especially interesting because many of its strongest agent startups are not trying to automate generic office tasks. They are going directly into expensive, difficult, high-value work in industries New York already knows extremely well: finance, law, accounting, healthcare, real estate, insurance, sales, compliance, and professional services.
NYC Tech Journal analyzed 13 New York companies that are either building AI agents directly or supplying critical infrastructure that agents need. Based on publicly reported funding figures, this group has raised roughly $2.52 billion.
Our most important finding is even more interesting.
About 67% of that capital has gone into companies working in regulated or high-stakes professional workflows.
That tells us something important about New York’s place in the AI economy.
Silicon Valley may still dominate the race to build foundation models. New York has a strong chance of becoming one of the places where those models are turned into actual workers.
The Short Version: New York Is Building AI That Does the Job
For the first phase of generative AI, most business software followed a simple pattern.
A person asked a question. AI generated an answer.
That was useful, but it rarely changed the structure of a company.
Agents are different because they can sit inside a workflow. They can watch for something to happen, gather information, make a judgment, perform an action, record what happened, and continue to the next step.
That changes the business case.
A bank does not really need another chatbot that can explain its lending policy. It may want an agent that reviews an application, checks the relevant information, applies the bank’s rules, identifies exceptions, documents its reasoning, and routes only uncertain cases to a person.
An accounting firm does not simply need AI that explains a tax form. It wants software that can work through the tax process.
A sales team does not only need AI-generated emails. It wants software that finds an account, researches it, identifies a buying signal, selects the right person, updates the CRM, and starts the right workflow.
That is the transition happening inside some of New York’s most valuable AI startups.
Clay is pushing agents deeper into go-to-market work. EliseAI is automating housing and healthcare operations. Rogo is building an AI system for financial professionals. Norm AI is bringing agents into law and compliance. Basis is building accounting agents. Taktile is applying agents to financial decisions. Regal is deploying voice agents at huge call volumes.
Around those companies, another layer is appearing. AIR is building security controls for agents. Jedify is building the context layer agents need to understand a company. General Intuition is working on agents that can reason about space and eventually operate in the physical world.
The market is moving from AI that knows to AI that acts.
For New York businesses, that may be one of the most important technology shifts of this decade.
Original Research: How NYC Tech Journal Built the New York AI Agent Dataset
There is a problem with ranking AI agent startups.
Almost every AI company now uses words such as “agent,” “agentic,” “autonomous,” or “AI employee.” If we simply searched for companies using those words, the result would be almost meaningless.
So we used a stricter method.
Our Inclusion Rules
For this analysis, a company had to meet two basic tests.
First, it needed a meaningful New York presence and to be commonly identified as New York-based or have its core operating presence in the city.
Second, its product needed to do more than generate text. It had to perform, coordinate, control, or enable multi-step work.
That includes companies whose agents directly complete tasks, companies building autonomous decision systems, and a small number of infrastructure companies solving problems that become important specifically because agents can take actions.
We excluded generic AI software simply adding a chat interface.
We also excluded major agent companies based elsewhere, even if they have New York employees.
What We Measured
We collected publicly available figures for disclosed funding, last publicly reported valuation where available, production usage, customer scale, and evidence that the technology was performing real work.
Because private-company reporting is inconsistent, the analysis should not be treated like audited financial statements. Funding totals can differ slightly between databases because some count secondary transactions, extensions, debt, or undisclosed rounds differently.
For that reason, we use the figures mainly to understand scale, concentration, and direction, not to pretend that every figure is precise to the final dollar.
Our dataset is current through September 8, 2026.
NYC Tech Journal’s 2026 AI Agent Startup Dataset
| Company | Main Agent Market | Approx. Disclosed Funding Used in Our Analysis | Last Public Valuation Signal Used | What the Agent Does |
| General Intuition | Physical AI | $454M | $2.3B closed round | Learns to navigate and act in digital and physical environments |
| EliseAI | Housing & healthcare operations | $391.9M | $2.2B closed round | Automates operational work across housing and healthcare |
| Rogo | Finance | $310.5M | $2B | Research and workflow agent for financial professionals |
| Norm AI | Law & compliance | $260M+ | $1.2B | Applies legal and regulatory rules to enterprise work |
| Taktile | Financial decisions | $213M | Not disclosed | Automates decisions in banking and insurance |
| Clay | Sales & go-to-market | $204M | $5B secondary valuation signal | Researches accounts, decides actions, and runs GTM workflows |
| Hebbia | Professional research | $161.1M | About $700M | Analyzes large collections of documents for professional work |
| Tennr | Healthcare operations | $160M+ | $605M | Automates patient referral and administrative workflows |
| Basis | Accounting | $137.7M | $1.15B | Completes long-running accounting, tax, and audit work |
| Regal | Customer experience | $83M | $350M older disclosed valuation | Voice agents for sales, support, and operations |
| Patlytics | Patent work | $65M | Not disclosed | Automates workflows across the patent lifecycle |
| AIR | Agent security | $50M | Not disclosed | Controls which tools and add-ons AI agents can safely use |
| Jedify | Agent context | $33M+ | Not disclosed | Gives enterprise agents structured business context |
Source note: Funding and valuation figures are assembled from company announcements and recent reporting. The dataset deliberately uses closed valuation events where possible rather than treating ongoing fundraising discussions as completed transactions.
Clay, for example, reached a $5 billion valuation through a January 2026 employee tender, while reporting in late August said it was raising fresh capital at a $7 billion pre-money valuation. We therefore use $5 billion in the closed-valuation analysis and treat $7 billion as a current market signal rather than a completed financing event.
Chart 1: Where the Agent Money Is Going in New York
The first thing that jumps out is the size of the capital pool.
Our 13-company cohort has attracted approximately $2.52 billion in disclosed funding.
| Company | Funding Used | Relative Scale |
| General Intuition | $454M | ████████████████████ |
| EliseAI | $391.9M | █████████████████ |
| Rogo | $310.5M | ██████████████ |
| Norm AI | $260M+ | ███████████ |
| Taktile | $213M | █████████ |
| Clay | $204M | █████████ |
| Hebbia | $161.1M | ███████ |
| Tennr | $160M+ | ███████ |
| Basis | $137.7M | ██████ |
| Regal | $83M | ████ |
| Patlytics | $65M | ███ |
| AIR | $50M | ██ |
| Jedify | $33M+ | █ |
General Intuition’s $320 million 2026 financing brought its total disclosed capital to $454 million and valued the company at $2.3 billion. The startup is unusual within this group because it is trying to train generalized agents that can reason about movement and the physical world, rather than focusing on one office workflow.
EliseAI’s reported equity funding is roughly $392 million after its $250 million Series E. The company has also become one of the strongest operating businesses in the group: it said in June 2026 that it had passed $200 million in annual recurring revenue.
Rogo raised $160 million in April 2026, bringing total funding above $300 million. Its platform is now used across more than 250 investment banks and investment firms, according to the company.
Norm AI raised $120 million in July 2026 at a $1.2 billion valuation and said total funding had passed $260 million.
Those are no longer experimental startup numbers.
This is serious capital being placed behind the idea that AI can participate directly in professional work.
Original Finding #1: New York’s Agent Market Is Surprisingly Concentrated
The largest five companies in our dataset account for approximately 64.6% of all disclosed funding in the cohort.
The largest eight account for roughly 85.4%.
That means the New York agent market already has a visible upper tier.
Funding Concentration in Our Dataset
| Group | Share of Total Funding |
| Top 5 companies | 64.6% |
| Top 8 companies | 85.4% |
| Remaining 5 companies | 14.6% |
This matters because agent software is still talked about as if the entire category were in its experimental seed-stage period.
It is not.
Some New York agent companies are already operating with hundreds of millions of dollars in capital, nine-figure revenue in certain cases, large enterprise customer bases, and billion-dollar valuations.
That creates a different competitive environment.
A new agent startup is no longer competing only with an incumbent SaaS company that has been slow to add AI. It may also be competing with a heavily funded AI-native company that has spent years collecting workflow data, building integrations, improving evaluation systems, and learning where automation breaks.
In other words, the moat is moving beyond the model.
It is moving into the workflow.
Original Finding #2: About Two-Thirds of the Capital Is Going Into High-Stakes Work
We grouped the companies by the type of work they are trying to change.
The result is one of the clearest pieces of evidence for why New York’s AI ecosystem is different.
Chart 2: Share of Funding by Work Category
| Work Category | Funding in Our Dataset | Share |
| Finance | $684.6M | 27.1% |
| Housing & healthcare operations | $551.9M | 21.9% |
| Physical AI | $454M | 18.0% |
| Legal, compliance & IP | $325M | 12.9% |
| GTM & customer experience | $287M | 11.4% |
| Accounting | $137.7M | 5.5% |
| Agent security & context infrastructure | $83M | 3.3% |
If we combine finance, accounting, legal/compliance, patent work, healthcare, and EliseAI’s regulated operating markets, approximately 67.3% of the funding in our cohort is connected to high-stakes or heavily regulated workflows.
That is a remarkable concentration.
It also makes sense.
New York has enormous pools of workers who spend their days reading, checking, comparing, documenting, reviewing, approving, calling, researching, and making decisions.
These tasks may look different across industries, but the software problem underneath them is often similar.
The worker needs to collect information from several sources. The worker must understand rules. The worker needs to make a judgment. The worker must record why the decision was made. Then something needs to happen.
That is almost the perfect environment for agentic software.
The Biggest AI Agent Startups in New York
1. Clay — Turning Go-to-Market Work Into an Agent System
Clay may be the clearest example of how quickly a software company can move from “AI feature” to “agent platform.”
The company historically became known for giving sales and growth teams a powerful way to combine data providers, enrich leads, research companies, and build customized outbound workflows.
That was already useful.
Agents make the model more powerful because the software can now make judgments inside those workflows.
Claygent Is Moving Beyond Simple Sales Automation
Clay says its Claygent research agent had passed 5 billion runs by August 2026. It can research the open web, analyze business data, qualify leads, interpret unstructured information, and return structured results.
The more important development is Clay’s move toward persistent Account Agents.
An Account Agent can start with everything a company knows about a prospect, reason about a new signal, decide which permitted action should happen, remember the conclusion, and then trigger the next step. Clay describes the process as a loop of knowing, deciding, remembering, and acting.
That is much closer to a digital worker than a writing assistant.
Why Clay Matters to the Future of Sales Jobs
Traditional sales software organizes work.
Agentic sales software can increasingly perform parts of it.
Imagine a sales territory with 20,000 potential accounts.
A human team cannot deeply monitor every company every day. An agent can.
It can watch funding news, hiring, product launches, management changes, earnings calls, web activity, CRM history, prior calls, and product usage. It can then tell the sales team which accounts matter today.
Eventually, the difference between “sales software” and “sales labor” starts becoming harder to define.
Clay’s valuation shows investors understand the opportunity. Its January 2026 tender valued the company at $5 billion, while an August 31 report said a new financing was being pursued at a $7 billion pre-money valuation.
For New York businesses, the lesson is not simply to buy an AI sales tool.
The lesson is to redesign revenue operations around continuous machine research and human judgment.
2. General Intuition — A New York Bet on Agents for the Physical World
General Intuition is the strange company on this list.
That is exactly why it deserves attention.
Most enterprise agents live inside software. They read documents, call APIs, update databases, send messages, or make recommendations.
General Intuition is trying to build agents that understand movement through space and time.
The New York-based company spun out of Medal and is using enormous amounts of gameplay data to train models that can learn how actions change an environment.
In June 2026, General Intuition raised $320 million at a $2.3 billion valuation, bringing disclosed funding to $454 million. TechCrunch later reported that investors were discussing another financing at a much higher $6 billion pre-money valuation, though that later valuation had not been closed when this article was prepared.
Why a Gaming-Trained Agent Could Matter to Work
At first, this may sound disconnected from “the future of work.”
It is not.
A large share of the economy does not happen inside a spreadsheet.
Warehouses, hospitals, factories, construction sites, stores, hotels, restaurants, logistics facilities, and cities all involve physical movement.
Software agents can transform knowledge work. Physical agents could eventually transform operational work.
General Intuition is therefore a useful reminder that the AI agent market may not stop at white-collar automation.
If agents can reliably understand physical environments, the long-term market becomes much larger.
3. EliseAI — Agents Move Into Housing and Healthcare Operations
EliseAI has quietly become one of New York’s most important examples of vertical AI.
Instead of selling a general-purpose AI system, the company started with the hard operational problems inside housing.
That decision now looks extremely important.
By August 2025, EliseAI said it supported more than 600 owners and operators and 75% of the NMHC Top 50 operators. The company said its technology powered around 10% of the U.S. apartment market.
In June 2026, EliseAI announced that annual recurring revenue had crossed $200 million.
Then, in September 2026, it introduced Apollo, which it describes as an agentic AI teammate capable of performing tasks across its multifamily platform.
Why EliseAI Is Important
The company’s biggest advantage may not be the model.
It may be the operating context.
Housing operations involve leasing, tenant questions, appointments, maintenance, payments, renewals, records, compliance, and many exceptions.
Healthcare has similar complexity.
An AI agent that understands those processes can potentially do far more valuable work than a general assistant that simply answers questions.
That distinction will become increasingly important.
The future enterprise agent may not look like one universal digital employee.
It may look like hundreds of highly specialized systems that understand the rules, data, language, exceptions, and economics of a specific industry.
EliseAI is one of the strongest New York examples of that model.
4. Rogo — Building an AI Agent for Wall Street
If there is one market where New York should have an unfair advantage in AI agents, it is finance.
Rogo is trying to turn that advantage into software.
The company started with a simple but valuable problem: highly paid financial professionals spend extraordinary amounts of time gathering information, reading documents, rebuilding analysis, preparing materials, and moving information between systems.
That is exactly the kind of work modern AI can attack.
Rogo raised $75 million in January 2026 and then another $160 million only three months later. Its April financing valued the company at roughly $2 billion, according to secondary-market financing data, and brought total funding above $300 million.
The company says its platform serves more than 250 global investment banks and investment firms and is scaling an AI agent called Felix.
The Real Opportunity Is Not Faster Research
The obvious benefit of a financial AI agent is speed.
That may not be the biggest benefit.
The larger opportunity is workflow continuity.
A traditional analyst might search a database, read a filing, create a spreadsheet, compare competitors, build a chart, prepare a presentation, check the numbers, update the presentation after a new filing, and then answer questions from senior bankers.
Every transition creates work.
An agent that lives across the entire process can potentially remove those transitions.
That is why agentic finance may be much more disruptive than simply putting a chatbot next to a research database.
5. Norm AI — Agents That Can Understand Rules
Most businesses want AI to become more autonomous.
Regulated businesses immediately encounter another question.
What is the AI allowed to do?
Norm AI attacks that problem from the legal side.
The company says it embeds law into AI agents so the systems can operate inside legal and compliance requirements. Its affiliated Norm Law practice uses agents under attorney supervision and charges based on outcomes rather than the traditional hourly model.
That combination is important.
Norm is not merely using AI to make lawyers type faster. It is experimenting with what the business model of legal work looks like when software performs a meaningful portion of the work.
Norm’s Scale Is Growing Quickly
Norm raised $120 million in July 2026 at a $1.2 billion valuation, taking total funding above $260 million.
The company said that during 2025 its customer base represented more than $30 trillion in assets under management and that its systems performed more than 25 million AI-powered legal and compliance determinations.
Those numbers matter because legal AI faces a much harder standard than ordinary productivity software.
Being useful is not enough.
The system must also be explainable, controlled, reviewable, secure, and consistently tied to the correct rules.
That makes legal and compliance agents a natural New York market.
6. Basis — AI Agents That Actually Do Accounting Work
Accounting is one of the most interesting agent markets because the work sits directly between structure and judgment.
There are rules.
There are documents.
There are numbers.
There are repeatable processes.
But there are also exceptions, messy inputs, reconciliations, missing information, and cases that require reasoning.
Basis was built specifically for this environment.
The New York company says its agents can operate for hours and complete end-to-end accounting work. The company has demonstrated an agent completing a partnership tax workbook and says its systems are used for journal entries, reconciliations, accounting memos, tax, and audit workflows.
Basis raised $100 million in February 2026 at a $1.15 billion valuation, bringing total funding to about $138 million. Public reporting says its technology is used by roughly 30% of the top 25 accounting firms.
Accounting Shows What Long-Running Agents Could Become
Many current AI systems are designed for interactions that last seconds.
That is not how most jobs work.
Real professional tasks can run for hours or days.
An accounting agent may need to review many files, compare values, investigate mismatches, document its work, and return to an earlier step when something changes.
That is a much harder technical problem.
It is also a much more valuable one.
If long-running agents become dependable, a large part of the value created by enterprise AI could come from reducing the number of hours needed to complete recurring professional workflows.
Basis is one of New York’s clearest tests of that idea.
7. Taktile — Moving Agents Into High-Stakes Financial Decisions
Most companies began their AI journey with low-risk tasks.
Write an email.
Summarize a call.
Search a document.
Taktile is going in the opposite direction.
The company is building systems for decisions that directly affect financial outcomes.
These include lending, insurance claims, customer onboarding, and financial-crime detection.
Taktile raised $110 million in June 2026 in a Series C led by Growth Equity at Goldman Sachs Alternatives. CB Insights lists total funding at approximately $213 million, although the company did not disclose a new valuation.
The company says financial institutions are moving toward organizations increasingly powered by autonomous agents.
This Is Where Agent Governance Becomes Critical
A writing agent can produce a bad sentence.
A lending agent can make a bad credit decision.
The risk is completely different.
That is why Taktile is strategically important.
The financial sector will be one of the clearest tests of whether autonomous systems can be surrounded by enough controls, monitoring, evidence, and human review to operate safely in high-stakes environments.
For banks, the goal should not be “maximum autonomy.”
The goal should be the maximum safe autonomy that can be measured and controlled.
8. Hebbia — The AI Analyst Model
Hebbia sits between research software and an AI analyst.
Its Matrix product was built to analyze large collections of documents and organize answers in structured formats.
That sounds simple until you consider the work it replaces.
An investment analyst might need to compare dozens of filings, legal documents, transcripts, agreements, and spreadsheets to answer one question.
Hebbia can work across that material at once.
The company raised $130 million in 2024 at roughly a $700 million valuation. At the time, TechCrunch reported that Hebbia had around $13 million in annual recurring revenue and was profitable. The company said it was already used by 30% of asset managers.
Sacra estimates Hebbia reached approximately $48 million in ARR by August 2026, although that figure is an estimate rather than a company-reported result.
Why Research Agents Matter
Research is often treated as a small task.
Inside finance, legal work, consulting, life sciences, and government, it can be a major labor cost.
The problem is not finding one answer.
The problem is creating a reliable chain of evidence across hundreds or thousands of documents.
The winner in this category may therefore not be the system that produces the nicest paragraph.
It may be the one professionals trust enough to use when the answer has real financial consequences.
9. Tennr — Agents for the Invisible Work Behind Healthcare
Some of the most valuable AI opportunities are buried inside workflows that consumers almost never see.
Patient referrals are a good example.
The process involves medical records, insurance rules, documents, communications, eligibility, scheduling, and administrative follow-up.
Tennr is automating that work.
The New York company raised $101 million in June 2025 at a $605 million valuation, taking total funding above $160 million.
Its current product combines agentic workflow automation, document understanding, coverage criteria, triage, and controlled automation to help healthcare providers move patients through the referral process.
YC currently lists the company at about 205 employees in New York City.
Healthcare Agents May Win in the Back Office First
When people imagine AI in healthcare, they often imagine diagnosis.
Administrative operations may be a much easier place to create immediate value.
The work is expensive.
The processes are fragmented.
A great deal of information is already digital or can be digitized.
And much of the work involves interpreting information and deciding what should happen next.
That makes healthcare administration one of the strongest markets for specialized agents.
10. Regal — Voice Agents Reach Real Production Scale
Voice agents are one of the easiest forms of agentic AI to understand because the work has historically been done by people in contact centers.
A customer calls.
Someone answers.
They ask questions, access information, solve a problem, record the result, and decide what should happen next.
Regal is attempting to automate that sequence.
The New York company had raised $83 million as of its 2024 funding round, when it said it had around 100 employees and hundreds of customers.
More interesting is what happened afterward.
Regal says its customers ran more than 350 million calls and 570 million agentic workflows during 2025, with those interactions tied to $8 billion in revenue.
Those are company-reported figures, but even with that caveat, the scale is important.
Voice Agents Show Why Measuring Outcomes Matters
The question for a call-center agent should not be whether the voice sounds human.
The business question is whether it solves the problem.
Did the customer receive an answer?
Did the lead qualify?
Was the appointment booked?
Did the sale happen?
Was the interaction compliant?
Did the customer need to call again?
Those are much better measures of an AI agent than how impressive the demo sounds.
The Emerging Layer: Companies Building the Agent Economy Around the Agents
The most interesting part of an emerging technology market is often not the first application.
It is the infrastructure that appears once the application becomes important.
New York is beginning to develop that layer too.
Patlytics — Agentic Work for Intellectual Property
Patlytics is building AI around one of the most specialized areas of professional work: patents.
The New York company raised a $40 million Series B in April 2026, bringing total funding to approximately $65 million.
It says its platform now serves more than 40% of Am Law 100 IP practices and supports work across invention harvesting, patent drafting, infringement analysis, invalidity work, due diligence, and portfolio management.
The significance goes beyond patents.
Patlytics is another example of the same New York strategy: take a complex area where professional labor is extremely expensive, learn the workflow deeply, and use AI to automate more of the process.
AIR — Building a Firewall for AI Agents
As agents become more powerful, they also become more dangerous.
A chatbot that produces a bad answer creates one kind of risk.
An agent that has access to files, browsers, tools, systems, plugins, and enterprise data creates another.
AIR emerged from stealth in September 2026 with $50 million in funding.
Its product discovers agents inside a company and checks the tools, skills, plugins, Model Context Protocol servers, and other components those agents use. It can block connections that do not meet security requirements.
That sounds like infrastructure rather than the future of work.
But the two are directly connected.
Companies cannot safely give agents more authority unless they also build stronger control systems around them.
The bigger the agent economy becomes, the more valuable that control layer could become.
Jedify — Giving Agents the Context They Need to Understand a Business
Agents face another major problem.
They may be intelligent, but they do not automatically understand a company.
They do not know what “revenue” means inside one particular business.
They do not know which database is authoritative.
They do not know that a certain customer has a special contract.
They do not automatically understand internal permissions, operating rules, team language, or the relationship between information stored in different systems.
Jedify is trying to solve that problem.
The New York company raised $24 million in June 2026, taking total funding above $33 million. It builds what it calls a context graph connecting databases, warehouses, SaaS systems, documents, Slack conversations, recordings, and other sources so agents can work with company-specific knowledge.
This could become a critical category.
The smarter AI models become, the less defensible raw model access may be.
Knowing the company could become more valuable than knowing the internet.
Original Finding #3: The Best-Funded NYC Agents Are Not Generic
Look again at the leading companies.
Clay understands go-to-market operations.
Rogo understands finance.
Norm understands law and compliance.
Basis understands accounting.
Tennr understands healthcare referrals.
Taktile understands financial decisions.
Patlytics understands patents.
EliseAI understands housing and healthcare operations.
That pattern is too consistent to ignore.
New York’s strongest agent startups are mostly not trying to build a digital employee that works everywhere.
They are building systems that work extremely well somewhere.
Chart 3: What the Leading Companies Actually Automate
| Company | Primary Work Unit |
| Clay | Account research and GTM action |
| Rogo | Financial research and analysis |
| Norm AI | Legal and compliance determinations |
| Basis | Accounting, tax, and audit work |
| Taktile | Financial decisions |
| Hebbia | Professional document research |
| Tennr | Patient referral processing |
| Regal | Customer calls and CX workflows |
| EliseAI | Property and healthcare operations |
| Patlytics | Patent workflows |
The implication is important for founders.
A generic agent company has to compete with the world’s largest AI labs.
A vertical agent startup can compete on something the model provider does not automatically have: workflow knowledge.
That includes industry data, integrations, evaluation systems, permission structures, customer relationships, compliance logic, domain experts, and thousands of examples of what successful work actually looks like.
Those things compound.
Original Finding #4: New York’s Agent Economy Is Really a Professional-Services Automation Economy
New York has spent decades building enormous industries around expensive human expertise.
Investment banking.
Asset management.
Law.
Insurance.
Accounting.
Consulting.
Advertising.
Real estate.
Healthcare administration.
The old software model sold these professionals tools.
The emerging agent model can sell completed work.
That is a much bigger shift than it first appears.
Software Seats Versus Software Work
Traditional enterprise software usually charges for access.
The company has 100 workers.
It buys 100 seats.
Agentic software creates a different possibility.
What if the customer pays for 50,000 completed analyses?
Or 500,000 customer conversations?
Or 20,000 processed referrals?
Or 10,000 completed tax workflows?
The economic unit can move from the number of humans using the software to the amount of work the software completes.
Norm’s AI-native law firm is especially interesting here because it prices based on outcomes rather than traditional attorney hours.
That model could spread.
When software becomes capable of performing work instead of only supporting workers, pricing can increasingly follow the work.
Original Finding #5: The Agent Winners May Own the Workflow, Not the Model
A year or two ago, an AI startup could create a large technical gap simply by getting early access to a better model.
That advantage is getting harder to defend.
Models are improving quickly.
More models are available.
Open-weight systems are becoming more capable.
The cost of intelligence is falling.
That means the valuable part of an enterprise AI company may move upward in the stack.
What Could Become the Real Moat?
| Layer | How Defensible Is It Likely to Be? |
| Access to a general-purpose model | Low and falling |
| Basic prompt interface | Very low |
| Generic chatbot | Very low |
| Workflow integrations | Moderate |
| Proprietary industry context | High |
| Evaluation data from real work | High |
| Distribution inside an industry | High |
| Regulatory and permission controls | High |
| Ability to complete a workflow reliably | Very high |
| Historical outcome data | Potentially very high |
Clay’s product direction illustrates this clearly.
Its agents do not operate alone. They sit on top of CRM data, warehouses, call transcripts, enrichment providers, signals, workflows, and action systems. Clay says it maintains integrations with more than 200 data providers and is increasingly allowing agents to decide what action should occur next.
The agent is valuable because it lives inside a system.
That may become the defining feature of enterprise AI.
What Does This Mean for Jobs in New York?
The easiest prediction is that agents will “replace jobs.”
Reality will be messier.
Most jobs are bundles of tasks.
Some tasks are easy to automate.
Others require accountability, trust, relationships, negotiation, judgment, creativity, persuasion, or physical presence.
The result will probably vary enormously by occupation.
Entry-Level Knowledge Work Will Change First
Junior professional roles contain many tasks that agents are becoming good at.
Research.
Data gathering.
First drafts.
Document review.
Basic modeling.
Call summaries.
Reconciliation.
Form processing.
Qualification.
Information transfer.
That does not mean every junior job disappears.
It does mean companies may need fewer human hours to produce the same volume of work.
The bigger question is what happens to training.
If analysts become good bankers by spending years building models, what happens when an agent builds the models?
If junior attorneys learn by reviewing documents, what happens when software reviews most of the documents?
If accountants learn through repetitive reconciliation work, what happens when agents handle the first pass?
Companies will need to redesign career ladders, not simply remove tasks.
Managers Could Become Managers of Humans and Agents
One likely change is surprisingly practical.
Managers may manage digital labor.
A sales operations leader could supervise dozens of agent workflows.
A finance professional could assign research tasks to several specialized agents.
An accounting manager could review exceptions produced by agents completing routine work.
A contact-center leader might spend less time scheduling employees and more time evaluating agent behavior.
This requires different skills.
Employees will need to define tasks clearly, choose what the agent can access, design controls, evaluate results, and know when the software should stop and ask for help.
That may become a major new form of management.
The Most Important Question Is Not “Can an Agent Do This?”
Businesses frequently start AI projects with the wrong question.
Can AI write this?
Can AI answer this?
Can an agent perform this task?
Those questions are too narrow.
The better question is:
Can we redesign this workflow so a combination of software and people produces a better business result?
That creates a different conversation.
A company should not automate a seven-step workflow just because AI can handle step three.
It should map the entire process.
Where does information enter?
Where does someone wait?
Where are mistakes common?
Where does expensive labor perform repetitive work?
Which decisions carry risk?
What evidence must be recorded?
Which actions can safely be automated?
Where is human approval valuable?
Once the workflow is visible, the best place for the agent usually becomes much clearer.
Where New York Businesses Should Deploy Agents First
The best first workflow is rarely the flashiest one.
It is usually the workflow that is expensive, frequent, measurable, and constrained enough to evaluate.
A Practical Agent Opportunity Scorecard
| Question | Weak Agent Candidate | Strong Agent Candidate |
| How often does the work happen? | Rarely | Daily or thousands of times |
| Is the process repeatable? | Different every time | Similar structure each time |
| Is the required information digital? | Mostly offline | Mostly accessible through software |
| Can quality be checked? | Highly subjective | Clear evidence or outcome |
| Is labor expensive? | Low cost | High cost |
| Does delay hurt the business? | Little impact | Revenue, cost, or customer impact |
| Can permissions be limited? | Needs unrestricted access | Actions can be tightly controlled |
| Can humans review exceptions? | No clear owner | Clear escalation path |
A strong first agent workflow might be lead research, invoice reconciliation, intake processing, appointment scheduling, document review, compliance checks, customer qualification, account monitoring, or internal research.
A weak first deployment might be a highly unusual strategic decision that occurs twice a year and cannot be objectively evaluated.
A Better 90-Day AI Agent Plan for NYC Companies
Businesses should resist the urge to launch 20 agents at once.
That makes it difficult to learn what is actually working.
A better approach is to turn one workflow into a controlled experiment.
Days 1-30: Map the Work Before Buying the Tool
Start with one business process.
Measure how it works today.
Document the number of cases per month, human hours required, average completion time, error rate, escalation rate, cost per case, and outcome.
This creates the baseline.
Without a baseline, an AI pilot can easily look impressive without creating real value.
Identify the Decision Boundary
Separate the process into three groups.
The first group contains steps that can be automated with very low risk.
The second contains steps where the agent can act but should remain inside predefined limits.
The third contains decisions that still require a person.
This boundary matters more than the model.
Days 31-60: Run the Agent Beside the Existing Process
Do not immediately give the agent full control.
Run it in parallel.
Let the AI process real cases while the existing team continues working.
Compare the outputs.
Where does the agent fail?
Which inputs create confusion?
What information is missing?
Does it make the same mistake repeatedly?
How often does a person need to intervene?
The goal of this phase is not maximum automation.
The goal is understanding the failure modes.
Days 61-90: Automate the Safe Middle
Once the company knows where the agent is reliable, allow it to perform those actions automatically.
Keep the uncertain cases routed to people.
Then measure the full business impact.
The Metrics That Matter
| Metric | What It Tells You |
| Cost per completed workflow | Whether the agent creates economic value |
| Human minutes per case | How much labor is actually removed |
| Completion time | Whether the process gets faster |
| Exception rate | How often humans still intervene |
| Error rate | Whether automation creates hidden risk |
| Rework rate | Whether the first answer was really useful |
| Customer outcome | Whether people receiving the service benefit |
| Revenue impact | Whether the agent creates or protects revenue |
| Escalation accuracy | Whether the agent knows when to stop |
Do not optimize for the number of prompts.
Do not optimize for the number of agents.
Optimize for business outcomes.
What Investors Should Watch Next
The New York agent market has already produced billion-dollar companies.
The next stage will probably be harder.
During the early AI boom, fast growth and a compelling demo could attract substantial capital.
As the market matures, investors will need better ways to separate durable companies from model wrappers.
Watch the Amount of Work Completed
Revenue matters.
But with agent companies, another metric could become equally important: completed work volume.
How many calls did the system handle?
How many financial analyses did it complete?
How many decisions did it process?
How many tax workflows did it finish?
How many hours did it operate without intervention?
How many cases reached the correct outcome?
These numbers tell us whether an agent is being trusted with actual work.
Clay’s billions of Claygent runs and Regal’s hundreds of millions of calls are useful because they show usage at a scale that goes far beyond pilot programs.
Watch Human Intervention Rates
A supposedly autonomous agent that requires a person to repair half its work is not very autonomous.
This is one of the most important numbers in the category, yet companies rarely disclose it.
Over time, investors should pay attention to the percentage of workflows completed without human intervention, the percentage escalated correctly, and the percentage that later require human correction.
That will reveal whether agent economics are improving.
Watch Gross Margin Differently
Traditional SaaS investors became used to high gross margins because each additional software user was cheap to support.
AI changes this.
Long-running agents consume model inference, data, tools, browsing, storage, and often human review.
A company can have extraordinary revenue growth while also carrying large variable costs.
Investors therefore need to ask not only whether agent usage is growing.
They should ask whether the cost of completing each unit of work is falling.
Why New York Could Remain One of the Best Cities for Agent Startups
New York does not need to beat Silicon Valley at everything.
It needs to be unusually good at turning AI into businesses.
It already has the ingredients.
Customers Are Close
A finance AI founder can meet banks, hedge funds, private equity firms, asset managers, insurers, and fintech companies without leaving Manhattan.
A legal AI founder is surrounded by major law firms.
A real estate AI founder can work with some of the world’s largest property owners.
A healthcare operations startup can reach major health systems and specialty practices.
An advertising AI startup can work with agencies and brands.
That proximity matters when the product is deeply tied to workflow.
Agents are not simple software installations.
They often need to be trained around real processes.
The closer the startup is to the work, the faster it can learn.
New York Has Expensive Labor
This sounds like a weakness.
For automation startups, it can be an advantage.
The return on investment from automation rises when the human work being automated is expensive.
Saving an hour of a highly paid banker, attorney, accountant, salesperson, analyst, or medical administrator can support meaningful software spending.
That gives New York agent companies large economic targets.
New York Has Hard Problems
The best agent businesses may come from hard environments.
Finance has regulation.
Healthcare has fragmented data.
Law demands accuracy.
Insurance involves high-stakes decisions.
Real estate has operational complexity.
Accounting has detailed rules.
These problems create friction.
They also create defensibility.
A startup that learns to operate reliably inside a difficult industry may build a much stronger business than a company automating an easy, generic task.
The Biggest Risk: Confusing Autonomy With Value
The AI industry sometimes treats more autonomy as automatically better.
Businesses should not.
An agent that autonomously performs the wrong action is worse than a system that asks for approval.
The right amount of autonomy depends on the task.
A research agent may be allowed to search freely but not send an external email.
A customer-service agent may be allowed to issue refunds below a certain amount but escalate larger ones.
A financial agent may recommend a lending decision but require human approval for unusual cases.
A legal agent may complete routine work but route uncertain interpretations to an attorney.
The strongest agent systems will probably not remove all human control.
They will make the boundary between machine authority and human authority extremely clear.
Original Finding #6: Security and Context Could Become the Next Major NYC Agent Categories
AIR and Jedify are much smaller than Clay, EliseAI, Rogo, or Norm today.
But they may point toward the next phase of the market.
Once businesses deploy many agents, two problems become urgent.
The agents need to understand the company.
And the company needs to control the agents.
Jedify attacks the first problem.
AIR attacks the second.
The Emerging Enterprise Agent Stack
| Layer | Main Question | NYC Example |
| Model | Can the AI reason? | Often supplied by outside model labs |
| Context | Does it understand our business? | Jedify |
| Workflow | Can it complete useful work? | Clay, Rogo, Basis, Norm, Tennr |
| Decision | Can it choose the right action? | Taktile, Clay |
| Interface | Can it interact naturally? | Regal |
| Security | Can we control what it touches? | AIR |
| Governance | Can we review and explain its actions? | Built into several vertical platforms |
This stack could become extremely important.
The more autonomy companies give to AI, the less willing they will be to accept black-box behavior.
That creates opportunities for monitoring, permissions, evaluation, identity, audit trails, context management, security, and agent orchestration.
In other words, the agent boom could produce another software boom around the agents themselves.
What Business Leaders Should Learn From New York’s Biggest AI Agent Startups
There are several lessons hiding inside the companies in this article.
The first is that vertical knowledge matters.
The second is that integration matters.
The third is that completed work matters more than impressive conversation.
The fourth is that agents become more valuable when they operate inside an existing workflow rather than beside it.
And the fifth is that the best use cases tend to have a measurable economic result.
Clay can connect research to pipeline.
Rogo can connect financial research to analyst workflows.
Tennr can connect documents to patient movement.
Basis can connect AI reasoning to completed accounting tasks.
Regal can connect conversation to customer outcomes.
That is the pattern businesses should look for.
Do not ask where you can add AI.
Ask where information enters the company and then gets pushed through a long chain of human work before something useful happens.
That chain is where agents become interesting.
What NYC Business Leaders Should Do Now
The agent market is moving fast enough that ignoring it is no longer a sensible strategy.
But rushing into random pilots is not a strategy either.
The best approach is disciplined.
Choose a painful workflow.
Measure its current economics.
Give the agent access only to what it needs.
Create clear boundaries around the actions it can take.
Run it against real work.
Measure failures.
Escalate uncertainty.
Increase autonomy only when the evidence supports it.
Then move to the next workflow.
Businesses that follow this pattern will gradually build something much more powerful than a collection of AI tools.
They will build an operating system in which people and agents share the work.
The Bigger Story: New York Is Building the Machine Workforce One Industry at a Time
The most important thing about New York’s AI agent market is not the amount of money being raised.
It is what that money is trying to automate.
Our analysis of 13 New York agent companies found approximately $2.52 billion in disclosed funding.
About 67% of that capital is connected to high-stakes or heavily regulated professional workflows.
The five best-funded companies account for roughly 64.6% of the capital in the dataset.
Among companies with usable public valuation signals, several have already crossed the billion-dollar mark: Clay, General Intuition, EliseAI, Rogo, Norm AI, and Basis.
These are not all competing to build the same agent.
That is the important part.
They are dividing the economy into workflows.
Sales.
Finance.
Law.
Accounting.
Healthcare.
Housing.
Customer service.
Patents.
Physical operations.
Security.
Context.
That may ultimately be how the agent economy develops.
There may never be one universal AI worker that suddenly replaces the office.
Instead, software may slowly move deeper into thousands of individual workflows until an enormous amount of work that once required continuous human effort is being handled by machines.
New York is well positioned for that future because it already has the industries, customers, money, talent, and high-cost professional work needed to train the next generation of vertical agents.
The race is therefore not simply about who builds the smartest AI.
It is about who understands the work well enough to give that intelligence something useful to do.
And in New York, that race is already well underway.
