What is a vertical AI agent, and when is one worth paying for?

The short version

• A vertical AI agent is built for one industry or one job, while a horizontal agent is built to work across many.
• What makes an agent vertical is the data, tools, rules and tests around the model, not usually the model itself.
• In Vals AI's blind benchmark of 200 US legal research questions, legal AI scored 6 points higher on authoritativeness.
• AI spending and adoption surveys mostly cover AI as a whole, so they say little about vertical agents.
• A vertical AI agent is not worth paying for when the task is simple, your process is unusual or a general tool already does the job.

What is a vertical AI agent?

A vertical AI agent is an AI agent built for one industry, or for one defined job inside that industry. An agent is software that works towards a goal by taking steps and using tools, rather than only answering a question. Vertical means the scope is narrow on purpose.

Examples include Freed's AI front desk receptionist for clinics and legal AI tools such as Alexi, Counsel Stack and Midpage. The narrowness is the product. A clinic front desk agent is expected to understand bookings, cancellations and when to hand over to a person. A general assistant has to be taught all of that.

The term is also a marketing label, and nobody polices it. A vendor can put an industry prompt on a general chatbot and call the result a vertical agent. So the useful question is not whether a product uses the word. It is what has been built around the model.

How is a vertical AI agent different from a horizontal agent?

A vertical AI agent goes deep on one domain, while a horizontal agent goes wide across many. A horizontal agent offers a general ability, such as drafting or answering from documents, and you supply the context. A vertical agent arrives with much of that context built in.

The trade is depth for flexibility. Vertical AI agents need less setup when your work matches their assumptions, and they fit badly when it does not. Horizontal agents need more configuration and testing, but they can move between teams and tasks.

What does each type of agent give you?

What makes an AI agent genuinely vertical?

An AI agent is genuinely vertical when its domain knowledge sits in its data, tools, rules and tests, not only in its prompt. A prompt saying 'you are a legal assistant' changes the tone of the answer. It does not give the model access to case law, and it does not stop the model inventing a citation.

The model itself is often the least specialised part. Many vertical products run on the same general models that anyone can license. The value is in the layers around the model, and those layers take real work to build and keep current.

The test set is usually the hardest layer to copy. It takes time from people who know the work, and it grows every time the agent fails in a new way. If a vendor cannot describe how they test against real domain cases, treat the word vertical with suspicion.

What should be built around the model?

• •Domain data: the sources the job depends on, kept current and retrieved under the right permissions.
• •Tools: working connections to the industry's systems of record, including write access where the job needs it.
• •Rules: the constraints of the trade, enforced in code rather than requested in a prompt.
• •Handover: defined points where the agent stops and a person decides.
• •Tests: real cases from the domain, scored by people who know the work.

Do vertical AI agents perform better than general AI tools?

Sometimes, on the tasks they were built for, but the public evidence is thin. One useful test comes from legal research. Vals AI put 200 US legal research questions from six law firms to Alexi, Counsel Stack, Midpage and ChatGPT Team. Lawyers and law librarians then scored the responses blind.

Legal AI scored 6 points higher on authoritativeness, with ChatGPT Team as the general tool in the test. That is where you would expect a specialist to gain, because an authoritative answer depends on which sources a tool can reach.

The result needs limits around it. It is one measure, from one benchmark, in one field, with responses collected in July 2025. It tested research answers, not an agent carrying out a task across several systems. It supports the idea that domain sources help. It does not prove that vertical agents beat general tools everywhere.

Most buyers will need to run a smaller version of the same test on their own work.

What do AI adoption and spending figures say about vertical AI agents?

Very little, because the figures we looked at measure AI as a whole or describe enterprises. None of them isolates vertical agents at small firms. They are useful only if you hold them to what they measured.

The US Census Bureau's Business Trends and Outlook Survey reports that firms with at least 20 employees are the biggest AI users. Its question covers AI use in any business function in the past two weeks, so it is a measure of AI use in general, not of agents. The article text gives no percentages for firms with 5 to 49 employees.

SBE Council surveyed 517 small business employers with 2 to 99 employees in February 2026, with a stated credibility interval of 4.4 points. That survey is closer to a small firm's position, but its spending figure covers all AI spend, not vertical agents.

Menlo Ventures' 2026 report sets its own bottom-up estimate at annualised run rate alongside a Morning Consult survey of nearly 400 US enterprise decision-makers. The estimate is modelled, it comes from an investor in AI vendors, and the web page gives no breakdown for small firms.

How much does a vertical AI agent cost?

There is no reliable market price for a vertical AI agent, and we do not know of a published figure that a small firm could budget from. Prices vary by vendor, by industry and by usage. Treat any single number as one vendor's offer, not a benchmark.

Freed's AI front desk receptionist for clinics shows how to read a public price. Freed publishes one on its product page, but it is one vendor's price, not a market average. When read in November 2026 the product was labelled beta, the plan button led to sales and no overage rate was listed.

The overage rate matters because agent usage is not flat. A seat costs the same in a quiet month and a busy one. Calls, conversations and completed tasks do not. If a price does not say what happens above the included volume, you cannot cost it for your busiest month.

The licence is also only part of the bill. Add the time to connect the agent to your systems, someone to review its work, someone to handle the cases it hands back, and retesting when the vendor changes the model. None of those appear on the pricing page.

When is a vertical AI agent not worth it?

A vertical AI agent is not worth it when the task is simple, your process is unusual, or a general tool already does the job. A specialist product earns its price by carrying domain work you would otherwise have to do. If there is little domain work in the task, you are paying for the label.

Drafting, summarising and answering questions from your own documents often need no vertical product. A general model, your own documents and a sensible review step can be enough, and that setup is easier to change later.

An unusual process is the opposite failure. A vertical agent encodes somebody else's idea of how the work runs. If your intake, pricing or approval chain differs from that idea, you will either fight the product or reshape your process around it. Neither is what you were paying for.

Volume matters too. A task that happens a few times a week may not repay the setup, testing and supervision. An agent also will not fix a process that is already broken. It will run the broken process faster, with fewer people watching.

Should you buy or build a vertical AI agent?

Buy when a product fits your workflow and the task is common in your industry. Build when the task depends on your own data, systems or rules and no product handles them properly. Try buying first, because a failed trial costs less than a failed build.

Surveys will not settle this for you. Menlo Ventures published a build or buy figure from a survey of 495 US enterprise AI decision-makers, fielded in November 2025, and the web page for its 2026 edition gives no updated figure. Menlo invests in AI vendors, and its respondents are enterprises. A small firm should read that as background at most.

There is also a middle path, and for most firms it is what building means. You license a general model and build only the domain layer: the data connections, rules, handover points and tests. Training a model from scratch is rarely part of it. The work you own is the part that makes the agent fit your business, and it can move to a different model later if it is built with that in mind.

How do you evaluate a vertical AI agent before you buy it?

Test it on your own real cases, not on the vendor's demo. A demo shows the paths the vendor chose. Your risk sits in the cases nobody chose: the odd request, the missing record, the customer who changes their mind halfway through.

What should the test include?

1. Collect real cases from your own work, including the awkward ones.
2. Have someone who does the job score the results without knowing which tool produced them.
3. Run a general tool on the same cases, so you can see what the specialist actually adds.
4. Check what the agent does when it is unsure, and whether it hands over or guesses.
5. Find out which actions it can take without approval, and whether you can limit them.
6. Ask where your data goes, who can see it and whether it is used to train models.
7. Get the overage rate and the notice period for model changes in writing.
8. Run it alongside your current process before you let it act alone.