Why do AI agents need business data?

Explanation#

Ticketing, CRM, code review and approval systems record sequences of actions and outcomes.

These records can provide training examples and realistic evaluation tasks for agents.

Rights and privacy#

• 1Confirm customer, vendor and employee agreements allow the use
• 2Remove or de-identify personal details before anything leaves the company
• 3Approve the final dataset and permitted uses in writing

How licensing works#

A business describes its data, buyer demand is assessed, scope and permitted use are proposed, the data is prepared and de-identified, an agreement is signed, and the approved dataset is delivered. See data licensing for detail.

What is an AI agent?#

An AI agent is a system in which a large language model decides its own next steps and which tools to use, rather than following a fixed script. Anthropic draws the line this way: workflows orchestrate models and tools through predefined code paths, while agents let the model dynamically direct its own process and tool usage.

In practice an agent runs in a loop. It reads the goal, calls a tool such as a search, an API or a code runner, looks at the result, and decides what to do next. Anthropic notes that agents need ground truth from the environment at each step, such as tool results or code execution, to judge whether they are making progress.

AI agent vs LLM: what is the difference?#

An LLM is the model that produces text from a prompt; an agent is a system built around an LLM that acts on tools, checks results and keeps going until the task is done or a stopping rule is hit. The same model can power a chatbot or an agent; what changes is the loop, the tools and the responsibility for an outcome.

Which business records map to which agent tasks?#

The examples below are illustrative, not a list of what any particular buyer wants; each record shows the request, the steps taken and the outcome.

• 1The record shows the starting request, not only the final answer.
• 2Intermediate steps and status changes are timestamped.
• 3The outcome is recorded, including failures and reopened cases.
• 4The rules the team followed (policies, approval limits) can be described.

How are AI agents tested on real work?#

Agent benchmarks are increasingly built from real work records, with success measured by whether the end state is correct. SWE-bench, for example, uses 2,294 problems drawn from real GitHub issues and the pull requests that resolved them across 12 popular Python repositories.

tau-bench tests agents in a simulated conversation with a user, giving the agent domain-specific API tools and policy guidelines, and scores it by comparing the final database state with an annotated goal state. Both designs depend on knowing what correct work looked like, which is exactly what business system histories record.

When a company licenses records for this use, it approves exactly what is released, once per deal, and sees what would leave the company before anything is shared.