Where can a new company still break in AI agents?

Why does the AI agent market feel so crowded now?

The AI agent market feels crowded today because the frontier labs have absorbed a surprising amount of what used to qualify as an agent startup.

OpenAI can now connect agents to company apps, files and tools, let them work for long periods and schedule recurring tasks. Google is building similar capabilities into Gemini and its enterprise agent stack. Microsoft has pushed memory, skills, organizational data and multi-step execution into Copilot Studio. Anthropic keeps extending Claude from conversation into coding, browsers and enterprise workflows. As labs bundle more agent features into their own products, we follow which technologies are gaining traction in our agentic AI market report.

A year or two ago, connecting an LLM to tools, adding memory and letting it complete several steps could justify a company. These days, much of that stack comes from the model provider.

Hebbia shows how quickly the bar is moving. The company recently overhauled Matrix after competitors reproduced parts of its original document-analysis experience. Matrix 2.0 now goes further into executing workflows, connecting databases and producing finished memos or presentations.

For a new entrant, the bar is clear: own something a frontier lab cannot reproduce with another model update or connector.

Are companies actually handing real work to AI agents yet?

Companies are already handing meaningful work to AI agents, although adoption is still much deeper inside large enterprises than across the average company.

McKinsey's latest global AI survey found that 40% of respondents at companies with more than $1 billion in annual revenue said they were scaling agents in at least one part of the business, up from 27% a year earlier. Smaller companies were much flatter at 22%.

OpenAI's latest enterprise usage data points in the same direction. Among its enterprise customers, Codex generated 64% of combined Codex and ChatGPT output tokens as of the latest measured quarter. Since February, weekly active enterprise Codex users had multiplied 108 times in legal, 41 times in sales, 41 times in recruiting and 26 times in marketing. Engineering, where agentic adoption started much earlier, grew fivefold over the same period.

That does not mean 64% of enterprise work has suddenly become autonomous. Agentic jobs naturally consume more tokens because the system works through many steps. Still, the change in usage is too large to dismiss.

The market has moved from “will businesses use agents?” to “which jobs will businesses actually trust agents to own?”

What is still too hard for generic AI agents to handle reliably?

The biggest remaining opening in AI agents is the gap between a model that can perform individual actions and a system that can finish a long business process correctly.

Companies already have models that can reason, browse, write code, read documents and call software tools. The remaining headaches are much less glamorous: figuring out which data is authoritative, keeping track of changing state, understanding company-specific rules, respecting permissions, recovering from exceptions and knowing when a human needs to intervene.

McKinsey's research still finds data limitations among the biggest obstacles to scaling agentic AI. Most businesses have important information scattered between email, documents, SaaS applications, databases and knowledge that exists only in employees' heads.

OSWorld 2.0 shows how large the execution gap remains. The benchmark contains 108 workflows that resemble actual computer work and take humans a median of about 1.6 hours to complete. The best tested configuration finished only 20.6% of the workflows perfectly, despite earning 54.8% partial credit.

The failure modes are exactly the ones that matter in business. Agents lose track of earlier constraints, miss information that appears later, guess when they should ask the user and fail to verify their work.

A claims process makes the opportunity concrete. Reading an insurance document is becoming easy for a frontier model. Completing the claim still means identifying the correct policy, checking coverage, collecting missing evidence, reconciling contradictory information, updating the insurer's systems and escalating unusual cases.

For a startup, the valuable work sits in making one of those long workflows reliable enough to trust.

Can a new general-purpose AI agent or agent builder still win?

A new general-purpose AI agent or generic agent builder can still win, but both are now among the toughest places for a startup to enter.

OpenAI, Anthropic, Google and Microsoft already have the distribution, models, compute budgets and integrations required to build increasingly general agents. OpenAI says Codex has spread far beyond software development, while ChatGPT Work is explicitly designed to take longer projects and execute them across company tools. Gemini and Claude are moving in the same direction.

The infrastructure underneath these products is also becoming standardized. Google now gives developers managed agents with background execution, remote MCP connections, functions and scheduled triggers. Microsoft Copilot Studio includes orchestration, memory, skills and access to organizational data. OpenAI's enterprise agents can work with apps, files and custom MCP servers. With every big platform now selling its own agent tools, we compare who is ahead and who is slipping in our agentic AI report.

MCP reinforces the trend. Anthropic created the protocol and later acquired Stainless, which helps turn APIs into SDKs, CLIs and MCP servers. Connecting an agent to a software application is becoming much easier.

There may still be room around open-source agents, sovereign deployments or genuinely new model architectures. For a normal application company, though, owning domain logic around a narrow job looks much safer than competing on generic orchestration.

AI agent approach	Room for a new startup	Why
General consumer agent	Low	ChatGPT, Gemini and Claude already have massive distribution
General enterprise agent	Low	Microsoft, OpenAI and Google can bundle agents into existing enterprise relationships
Open or sovereign agent stack	Medium	Some customers need deployment and model control the large platforms cannot fully provide
Vertical workflow agent	High	Domain rules, integrations and exceptions still have to be solved
New agent model architecture	Potentially high	A real technical breakthrough can still create a company, but the capital and research bar is much higher

Can a new AI coding agent still break through?

A new AI coding agent can still break through today, but generic coding is already one of the most competitive markets in AI.

Cognition is the clearest proof that independent coding-agent companies can become enormous even while competing with the frontier labs. When Cognition raised $1 billion earlier this year, the company said Devin had reached a $492 million annualized revenue run rate and enterprise usage had grown roughly 50% month over month for six consecutive months. The financing valued Cognition at $26 billion post-money.

The market has kept moving since then. Cognition has reportedly discussed another round at a valuation of at least $40 billion if its annualized revenue approaches $1 billion. Factory raised $150 million at a $1.5 billion valuation to focus specifically on enterprise engineering teams. Meanwhile, Cursor, Claude Code, Codex, Replit and several other products are fighting for the same developer budgets.

That competition pushes a new entrant toward harder engineering work. Devin itself is heavily used for tasks such as legacy modernization and migrations. Similar openings exist in regulated codebases, security remediation, testing against proprietary hardware, industrial software and old internal systems.

A broad “AI software engineer” is a much harder entry point now than a specialized engineering agent built around one ugly backlog.

Is AI customer service already too crowded?

Generic AI customer service is already crowded enough that a new startup needs a much narrower reason to exist.

Sierra recently raised $950 million at a valuation above $15 billion and says more than 40% of the Fortune 50 now use its platform. Its agents already handle billions of customer interactions across areas such as returns, mortgages, insurance claims and fundraising.

Decagon has also moved fast. The company raised $250 million at a $4.5 billion valuation and added more than 100 enterprise customers during the previous year. Recent third-party estimates put its annualized revenue around $100 million, although Decagon itself has not publicly confirmed that figure.

Gartner has estimated that roughly 17 million people work in contact centers worldwide, so the underlying market is huge.

The remaining openings are in customer interactions where the agent has to perform regulated or operational work: insurance claims, collections, healthcare coordination, financial servicing or industry-specific authentication.

Are AI sales agents already a bad place to start?

Generic AI sales agents look like one of the weakest places for a new AI agent company to start now.

Prospect research, personalized emails, meeting preparation and CRM updates were obvious early agent use cases. They were also relatively easy to copy. The result is a market packed with AI SDR startups, AI-native CRMs and incumbent products from Salesforce, HubSpot, ZoomInfo and others. When products are this easy to copy, telling what works from what is hype matters, and we do that in our agentic AI reality check.

The frontier models make the problem worse for new entrants. Researching a company and writing a decent outbound email is exactly the kind of task that improves quickly as models improve.

More interesting businesses sit closer to the transaction: renewal management, pricing, quote generation, account expansion, industry-specific sales operations or proprietary buyer-intent data.

An agent that owns a revenue outcome is much more interesting than one that produces another sales artifact.

Is healthcare administration one of the best openings for AI agents?

Healthcare administration is currently one of the strongest places for a new AI agent company because the work is expensive, repetitive, fragmented and still difficult enough that generic agents cannot simply take it over.

Assort Health started around patient scheduling and has expanded into intake, referrals, document processing, medication refills, eligibility, lab requests and payments. The company raised $120 million at a $1.2 billion valuation, then recently pushed further into referral automation for specialty practices and health systems.

Prosper AI gives us another useful data point. After raising $30 million, the company said revenue had grown fivefold in six months, more than 40 healthcare organizations had joined the platform and its software was operating across more than 150,000 providers. Prosper combines scheduling, insurance verification and patient billing rather than selling each task separately.

Parallel is attacking hospitals from another direction. Its agents work on top of existing hospital systems and initially focus on medical coding. The company raised $20 million after deploying across several dozen hospitals.

The most attractive wedges are the ones that can expand from one administrative task into several adjacent ones after the startup earns system access and trust.

Healthcare AI agent wedge	What we see today	How open it looks
Basic scheduling	Already competitive	Medium to low
Patient access and referrals	Strong demand, several fast-growing startups	Medium
Insurance verification and billing	Large labor burden and messy payer interactions	High
Hospital coding and administrative workflows	Legacy systems remain extremely manual	High
Specialty-specific operations	Rules vary heavily by specialty and provider	Very high where the niche is large enough
Generic healthcare assistant	Frontier models can increasingly provide the basic experience	Low

Why is insurance back-office work still so attractive for AI agents?

Insurance back-office work remains unusually attractive for AI agents because insurers still pay large numbers of people to move information between documents, portals, emails and old core systems.

Pace is one of the clearest examples. The company raised $46 million from investors including Thrive and Sequoia and says its agents have already completed more than a quarter of a million insurance workflows. Palomar reportedly uses Pace to resolve 90% of some policy-servicing work without increasing customer-service headcount at the same rate. Doing policy work without adding staff is a business model in itself; we compare which models are winning in our agentic AI market report.

The economics become especially obvious during demand spikes. When a major storm created thousands of new claims for one claims-servicing customer, Pace says its agents processed the work without the usual backlog that would have required a temporary operations team.

A startup can begin with document collection or policy servicing and then expand into policy checks, correspondence, fraud review, settlement preparation or renewals. That gives insurance agents a natural path from a narrow wedge into a much larger operating role.

Could finance and procurement become the next big AI agent vertical?

Finance and procurement are becoming some of the most interesting AI agent markets right now because the work combines structured rules with enormous amounts of repetitive human execution.

Rillet is a good recent example. The AI-native accounting company raised $100 million at a $1 billion valuation after saying its annualized revenue rate doubled in a single quarter. It now serves more than 600 customers, and roughly half of those customers reportedly switched from incumbent systems including Intuit, NetSuite, Oracle, SAP, Workday and Microsoft products.

Procurement is showing a similar pattern. Lio raised $30 million to build agents that work through supplier evaluation, compliance checks, negotiations and purchasing. The company says one global manufacturer automated 75% of previously outsourced procurement operations within six months. Freehand recently raised $75 million to build AI agents that manage supply-chain spending for large customers including Meta, Unilever, Johnson & Johnson and Pfizer.

Ramp is also moving aggressively into agentic finance, including accounts payable, procurement, close, accounts receivable and expense management. That raises the competitive bar, but it also validates the size of the opportunity.

The most attractive finance workflows are probably those where an agent can own a recurring process with a clear definition of correct completion. Reconciliation, close preparation, invoice exceptions, procurement operations and collections fit much better than vague “AI for the CFO” products.

Finance or procurement workflow	Why agents fit	Main risk
Accounting operations	Repetitive, rules-heavy and easy to verify	Rillet and incumbent finance platforms are moving quickly
Procurement execution	Work crosses contracts, ERP systems, suppliers and email	Large procurement suites can add agents
Supply-chain spend management	Large labor budgets and measurable savings	Enterprise integrations can be slow
Accounts payable and receivable	High volume with clear outcomes	Crowded fintech market
Generic finance copilot	Easy for existing finance software to bundle	Low standalone defensibility

Is legal AI already closed to new agent startups?

Legal AI is crowded at the horizontal layer, but several parts of legal work are still open enough for a new company to grow quickly.

Harvey reached about $190 million in annual recurring revenue by the end of last year and later raised at an $11 billion valuation. Legora crossed $100 million in ARR and was valued at $5.6 billion before recently exploring financing at an even higher valuation. According to the Financial Times, Legora has maintained more than 50% annual recurring revenue growth for seven consecutive quarters.

Sandstone raised $30 million only months after its seed round by focusing on in-house legal teams rather than fighting Harvey and Legora directly for private-practice workflows. The company says its revenue increased more than 40 times over roughly four months as it brought legal intake, triage, drafting, review and institutional context into one system.

Norm has gone further. After raising $120 million at a $1.2 billion valuation, it is building an AI-native law firm where attorneys supervise AI agents and clients pay based on outcomes rather than billable hours.

Another general-purpose law-firm copilot looks difficult. In-house workflows and AI-native legal services still look much more open.

Can AI agent security and testing become major standalone markets?

AI agent security and testing can both become major markets because companies are deploying autonomous software faster than they are learning how to control and verify it.

On the security-operations side, Torq raised $140 million at a $1.2 billion valuation around AI-driven security operations. Security analysts already spend much of their day collecting context across tools, investigating alerts and running response procedures, which maps naturally onto agentic work.

A newer opportunity is controlling the agents themselves. NewCore emerged with $66 million to build identity infrastructure for a workforce that increasingly includes humans, machines and AI agents. Enterprises need to know which agent is acting, who owns it, what information it can access and which actions it is allowed to take. Rounds like this show money following the need, and we look at who is getting funded, and by whom, in our report on agentic AI.

Microsoft has already started attaching Entra identities to Copilot Studio agents, while finance and payments companies are adding identities, budgets and spending policies for autonomous software. That makes cross-platform control more interesting than a narrow identity feature tied to one ecosystem.

Testing is developing alongside security. Patronus AI raised $50 million after its revenue grew fifteenfold in a year. The company builds simulated environments where businesses can test how agents behave across difficult scenarios before giving them access to real systems.

OSWorld 2.0 explains why that category exists. Long-running agents still lose state, miss information and skip verification. Every major model update can also change agent behavior, forcing companies to test workflows again.

The strongest control layer would combine some mix of permissions, auditability, evaluation, regression testing and runtime intervention across several agent platforms.

Is agentic commerce still open to startups?

Agentic commerce is open above the transaction rails, while the core payment layer is already filling up with very large incumbents.

Visa can now let AI agents initiate purchases through its payment network, with controls for spending limits, approvals and merchant restrictions. Mastercard is building its own agentic payment architecture and working with companies such as Google and Microsoft on standards for machine-driven commerce.

Ramp is pushing the same idea into corporate finance. Its current agent infrastructure can assign an autonomous agent an identity, owner, budget and approved payment methods while recording each financial action.

Stripe executives are now openly talking about a future where the conventional checkout page becomes much less important as agents move from recommending products to completing purchases.

Trying to build another generic payment rail for agents therefore looks difficult. The better opportunities sit around the decision and control layers: B2B purchasing agents, merchant tools for agent traffic, fraud systems designed for autonomous buyers, product-data infrastructure and industry-specific procurement.

The payment itself may become standardized. Deciding what the agent is allowed to buy, from whom, under which commercial terms and with which evidence is a much richer startup problem.

Is replacing outsourced labor a better AI agent opportunity than selling a copilot?

Replacing outsourced or back-office labor is one of the strongest AI agent opportunities because the startup can sell against a much larger cost base than a normal software subscription.

Pace effectively competes with insurance operations teams and outsourced processors. Lio says one customer automated three quarters of previously outsourced procurement work. Norm is building legal services around AI agents and charging for outcomes rather than lawyer hours. Sierra increasingly prices around completed customer-service work.

A software tool that makes ten employees 10% more productive has to fight for part of the existing software budget. An agent that safely takes over work costing a company $1 million a year can capture a much larger amount of value.

The trade-off is that the startup now owns the result. If the agent misses an insurance claim, pays the wrong invoice or mishandles a legal process, “the model made a mistake” is not an acceptable answer.

That makes these companies harder to operate, but potentially much more valuable.

What kind of AI agent moat survives when models keep getting better?

The best AI agent moats today are the ones that become stronger when OpenAI, Anthropic and Google release better models.

Workflow history is one. A company that has processed hundreds of thousands of insurance cases accumulates examples of exceptions, failure modes and correct resolutions. Those examples can be turned into evaluations and operating rules that a new entrant does not have.

Deep integrations help too. Once an agent works across an insurer's core platform, email, document systems and external data providers, replacing it involves much more than switching models. Integrations that are hard to swap out are one reason some companies keep winning, and we weigh the likely future winners in our report covering agentic AI.

Owning the system of record can be even stronger. Rillet is interesting because it wants to replace parts of the accounting stack itself, rather than sitting permanently as a thin AI layer above another vendor's software.

Distribution and trust become important in regulated markets. Hospitals, insurers and law firms are unlikely to give autonomous software broad permissions every few months just because another startup's benchmark score is slightly higher.

The weakest moat is still the prompt. McKinsey's latest survey found that 32% of respondents said their organizations had already decided against buying at least one software product or feature because agentic coding tools allowed them to build it internally. That should scare any founder whose product can be recreated by a competent employee with a frontier model and a few APIs.

We would ask one question before building: if the models become twice as good next year, does the startup become twice as useful or half as necessary? The best agent companies should benefit from the first outcome.

Where can a new company still break in AI agents?

A new company can definitely still break in AI agents, but the best openings have moved toward vertical workflows, operational control and completed outcomes.

Healthcare administration sits near the top for us. Assort, Prosper and Parallel show strong demand, while the underlying market still contains huge amounts of manual work spread across old systems. Insurance has the same characteristics, with Pace already demonstrating that agents can absorb real back-office volume rather than simply assist employees.

Finance, accounting and procurement have become much more interesting lately. Rillet's rapid growth, Lio's procurement automation and Freehand's push into supply-chain spending suggest that agentic finance is moving beyond demos into serious enterprise deployment. There is plenty of competition, but there is also an enormous amount of manual work to replace.

Cybersecurity is another strong bet because the agent boom itself creates demand for identity, permissions, monitoring and control. Testing and evaluation should grow for the same reason: companies will deploy more agents before those agents become perfectly reliable.

We would still consider specialized engineering agents, although generic coding already has extremely strong companies. We would be much more cautious around generic sales agents, horizontal customer-service agents, basic MCP connectors, generic agent builders and broad general-purpose assistants.

The pattern across the strongest openings is remarkably consistent. A valuable agent company takes responsibility for a real job that has a clear end state, works through the ugly systems surrounding that job and knows how to deal with the cases where the model gets confused.

Frontier labs can provide more of the intelligence every quarter. That leaves a new startup with a much clearer job: own the part between intelligence and a correctly completed piece of work.

Rank	Where we would look now	Why a new company can still break in	Main risk
1	Healthcare administrative workflows	Huge labor burden, fragmented systems, clear outcomes and many specialty-specific niches	Slow enterprise deployment and regulation
2	Insurance operations	Large back offices, repetitive workflows and strong expansion from one task into many	Incumbent insurance software vendors
3	Finance, accounting and procurement agents	Recent startup growth shows real demand and the workflows have measurable completion criteria	Ramp, Rillet and large ERP vendors are moving fast
4	Agent security, permissions and runtime control	More autonomous agents directly create more security demand	Microsoft and cybersecurity incumbents can bundle parts of the stack
5	Vertical agents replacing BPO and outsourced work	Much larger economic target than ordinary SaaS and strong outcome-based pricing	The startup has to accept operational responsibility
6	Agent testing, evaluation and control	Long-horizon reliability remains weak and every deployment needs regression testing	Basic observability will become a platform feature
7	Specialized engineering agents	Coding demand is proven and difficult enterprise backlogs remain	Cognition, Cursor, Anthropic, OpenAI and others are exceptionally strong
8	Agentic commerce applications above payment rails	Agents are beginning to transact, creating new purchasing and control workflows	Visa, Mastercard, Stripe and Ramp are taking the infrastructure layer
Avoid	Generic agents, generic builders, generic AI SDRs, basic connectors and thin copilots	Model and platform improvements can reproduce the core product too easily	Commoditization can happen in a single product release