AI agents are meant to complete tasks across several steps and tools, not only answer questions. To learn that, they need examples of real work: the sequence of actions, the tools used, the exceptions and the decisions made. Public web text describes work in general terms. Company workflow records show how it was actually done.
From answering to doing
The first widely used language models were built to produce text in response to a prompt. They were trained mainly on text from the public web and on books, which is a good source for learning language and general knowledge.
An AI agent is asked to do something different. It receives a goal, such as resolving a billing dispute or setting up a new customer account, and has to carry it out. That means choosing steps, calling software tools, reading the results, and adjusting when something unexpected happens.
Knowing about a task and being able to perform it are different skills. A person can read a description of how insurance claims are processed and still be unable to process one. The gap is filled by experience of real cases. For a model, experience means data.
This shift explains much of the current interest in company data. A model that only writes text can learn a great deal from published writing. A model that has to operate software and finish a job needs to see jobs being finished.
What an agent trajectory is
The term used for a record of a completed task is an agent trajectory. It is the ordered sequence of steps taken to get from the starting state to the result, across whatever tools and systems were involved.
A human trajectory for a support case might run as follows. The agent reads the ticket. They look up the account. They check the order history. They find a duplicate charge. They consult the refund policy, issue the refund, write to the customer and close the ticket with a category code. Each step has an input, an action and an outcome.
Company systems capture much of this without anyone intending it. Ticket histories record status changes and replies. Audit logs record which screens were opened and which fields were changed. Version control records each change to code and the review that followed. Linked together, these form trajectories of real work.
Multi-step tasks
Business tasks are rarely one step. Onboarding a client, closing the books for a month or handling a return each involve many actions, and later ones depend on earlier ones.
Errors compound over a long sequence. An agent that is right at each step most of the time can still fail a long task often, because one wrong turn early on sends everything after it off course. To be reliable, a model has to learn what a complete, successful sequence looks like, including the checks that experienced staff make along the way.
Public text seldom contains whole sequences. A help article explains a procedure in the abstract. It does not show two hundred instances of the procedure being carried out, each slightly different. Workflow records do.
Tool use
Real work happens inside software: ticketing systems, accounting packages, customer databases, code repositories, spreadsheets. An agent has to learn which tool to use for which purpose, in what order, and how to read what comes back.
Much of this knowledge is never written down. An experienced claims handler knows to check the policy system before the payments system, and knows what a particular status code means in practice. That knowledge shows up in the pattern of their actions.
Records that link activity across systems are especially informative for this reason. A ticket that can be tied to the account lookup, the knowledge base article consulted and the change made in the billing system shows the whole path. Each system's log alone shows only a fragment.
Timing is part of the pattern as well. Logs show how long a step took, where work waited for an approval, and which steps were repeated after a failure. That detail tells a model what normal progress looks like.
Exceptions
The standard case is the easy part of any job. The value of an experienced employee lies largely in handling what the procedure did not anticipate: the customer whose account is in an unusual state, the invoice that matches no purchase order, the request that falls between two policies.
Exceptions are rare individually and common collectively. Any single unusual case may appear once in thousands of records, but a meaningful share of all cases is unusual in some way. They are also where automated systems fail most visibly.
Only real operational history contains them in realistic proportions. Synthetic examples written to imitate work tend to cover the cases their authors thought of. Real records include the ones nobody thought of, along with what was done about them.
Judgment
Many decisions at work are not determined by a rule. Whether to escalate, whether to grant an exception, whether a piece of work is good enough to ship, how to word a difficult message: these are matters of judgment, shaped by context and experience.
Records capture judgment in several forms. Code review comments show what a senior engineer considered acceptable and why. Approval histories show what was signed off and what was sent back. Quality scores show how a supervisor rated a call. Internal discussion shows people reasoning through a hard case before acting.
This is some of the most useful material for both training and evaluation. It pairs a piece of work with an expert's assessment of it, which is hard to obtain any other way.
Why public web text is not enough
The public web is large, but it is a particular kind of record. It holds what people chose to publish. Operational work is mostly not published, for good reasons. The table compares the two as sources for what an agent has to learn.
There is also a simpler point. AI labs already have the public web. More of it adds little. Records from inside businesses are scarce by comparison, and they can only be obtained from the companies that hold them, with their agreement.
| What an agent has to learn | Public web text | Company workflow records |
|---|---|---|
| Full sequences of steps | Described in outline | Recorded case by case |
| Which tool to use, and when | Product documentation | Observed in real use across systems |
| Exceptions and edge cases | Sparse and anecdotal | Present in realistic proportions |
| Judgment and quality standards | Opinion and general advice | Reviews, approvals and scores tied to specific work |
| Outcomes | Seldom stated | Recorded: resolved, reopened, approved, rejected |
| Held-out test material | Likely already seen in training | Not public, so usable for evaluation |
What this means for a company's records
The records that matter most are those that show work in progress with its outcome attached, linked across systems and covering a long enough period to include the unusual cases.
This demand is real, but it should not be overstated. Not every record is useful, and data that cannot be cleared for licensing has no value to a careful buyer however rich it is. Records also contain personal and client information that has to be removed first. Outrich Data buys this kind of data directly from the companies that created it, after a rights review and de-identification.
Evaluation is a separate use from training and matters as much. To know whether an agent can do a job, a lab needs test tasks with known correct outcomes that the model has never seen. Real cases from a company's history, held back from training, serve that purpose. Their value depends on the records never having been public.
- Ticket and case histories with every status change, reply and resolution.
- Source code with commit history, review comments and linked issues.
- Workflow and audit logs that show actions taken inside business systems.
- Internal procedures together with the cases where they were applied.
- Approval chains, quality reviews and scored work.
- Internal discussion in which staff reason through a problem.