Why vertical AI agents need vertical company history

A general-purpose agent can draft an email or summarize a ticket. An agent built for freight brokerage, insurance claims or construction project management has to do more: follow the specific steps, vocabulary and exceptions of one industry. The only place those patterns are recorded in full is in the operating history of companies that work in that industry.

Every industry has its own shape. The tools differ, the words differ, the people outside the company differ, and the rhythm of the year differs.

Take a hypothetical 60-person freight brokerage. Its days are built around loads, carriers, rate confirmations, detention charges and proof of delivery. Its team talks to shippers, carriers and dispatchers, often several times about the same shipment. Its busy seasons follow the retail calendar and produce volumes.

Compare that with a 60-person dental practice management company, or a 60-person commercial roofing contractor. The headcount is the same. Almost nothing else is. Their tickets, their emails, their spreadsheets and their arguments are about completely different things, in completely different language.

An agent that is meant to work in one of these businesses has to learn that shape. It can pick up some of it from public documentation, but public material mostly describes how things are supposed to work. Company records show how they actually work, including the exceptions that take up most of the day.

For a vertical agent, the useful material is the sector-specific work that runs across a company's systems. In the freight example, that might include:

• email threads with carriers negotiating rates and resolving disputes
• tickets or tasks tracking a problem shipment from first report to final credit
• CRM records showing how shipper accounts are won, priced and renewed
• chat threads where dispatchers and account managers coordinate in real time
• finance workflows for invoicing, carrier payments and month-end reconciliation
• internal docs and wikis describing procedures, and the edits made when those procedures changed

Each of these is useful on its own. What makes them valuable together is connection depth: the same shipment, customer or dispute traceable across several tools. That is what lets an agent learn not just what a rate confirmation looks like, but what happens when one is wrong. We cover this idea in what connected data looks like in practice.

Many industries run on cycles longer than a few months. Insurance renews annually. Construction projects take a year or more from bid to closeout. Agriculture, tax preparation, education and retail all have seasonal peaks that look nothing like the rest of the year.

A dataset with a few months of history may catch a company only in its quiet season, or only in the middle of a project. Several years show the whole cycle, more than once. They also show how the company changed: a new pricing model, a new regulation to comply with, a tool migration, a key customer lost and won back.

For an agent that is supposed to operate over long horizons, that repetition is the signal. It sees how the same kind of work was handled the first time and the third time, and what changed in between. We go further into this in long-horizon tasks.

There is a second reason vertical agents need careful sourcing. One company's history shows how that company works, including its habits and quirks. An agent trained heavily on one freight brokerage may learn that brokerage's internal shorthand and assume every brokerage uses it.

A better approach is several companies in the same sector, ideally with some variation: different sizes, different regions, different tool stacks, operating and winding down. That variation helps a model separate what is true of the industry from what is true of a single team.

Finding several companies in one niche sector is a sourcing problem, not a data problem. Most of them have never thought about licensing their data and are not listed anywhere.

A lab building a vertical agent starts with a sourcing brief: the sector, headcount range, required tools, minimum years of history, regions and whether to include companies that are winding down. Martal builds a target list from company and technographic data, and our onshore reps, who have experience prospecting across 24 industries, call the people in those companies who can sign.

Sector experience helps on the calls. A rep who understands how a brokerage or a contractor runs can explain why that company's records are useful in terms the founder recognizes, and can answer the questions that come up in that industry, such as how customer and partner names are handled. Every person, company, address, email, phone and account becomes a consistent token, and sellers can exclude specific clients, channels or folders before any data moves.

Data is then evaluated, anonymized in-house and delivered on a rolling basis, so the lab can check early companies against its needs and adjust the brief as it goes.

Vertical agents will be judged on whether they handle the real work of an industry, not a tidy description of it. That requires training and evaluation data drawn from companies that did the work, across enough years and enough companies to separate the industry's patterns from one team's habits.

If you are building for a specific sector, send us a sourcing brief and we will scope what a campaign in that vertical would look like.