Can a company license agent trajectories to AI labs?
Yes, when the steps were recorded inside systems the company owns and the people involved are removed. A trajectory is the full path of a task: each click, command, search and decision from the start to the finished result. Labs building AI agents need to see how real work moves through real software. Many companies already log those steps in audit trails, ticket histories and automation runs.
Where task paths are already recorded
| System audit logs | The record a business system keeps of who changed what and when: an order edited, a status moved, a field corrected. |
| Ticket and case histories | Every action on a case, in order: assigned, looked up, replied, escalated, closed. |
| Automation and script runs | Logs from scheduled jobs, robotic process tools or internal scripts, with inputs, steps and results. |
| Runs of AI assistants the company deployed | Prompts, tool calls and outputs from an internal assistant, along with whether a person accepted the result. |
| Screen and command recordings of tasks | Recordings or shell histories of staff doing a defined task, such as setting up an account or fixing a server. |
What makes a trajectory complete
A useful trajectory has a goal, the starting state, every step taken with its result, and the outcome. If any of these is missing, the path cannot be replayed or judged. The most valuable ones include mistakes: the wrong search, the backtrack, the fix.
The outcome needs a verdict. A ticket closed and not reopened, an order shipped without a return, a script run that passed its checks: each tells a lab whether the path worked.
Pulling trajectories out of logs
Logs from different tools have to be joined by a shared ID and put in time order. An order ID can link a CRM change, a warehouse scan and an email in one path. This works best while the systems are live and the IDs still match.
Pick a few task types and export them well rather than everything. Count how many complete runs each type has, the date range and the share that ended in success.
Cleaning and approval
Logs are full of user names, IP addresses, customer IDs and free-text notes. These are replaced with consistent stand-ins so one person or customer keeps the same stand-in across every tool in a path. Passwords, keys and tokens that show up in command histories are removed.
Runs of a third-party AI tool also follow that tool's terms, which may limit how its outputs are used. Counsel checks those terms before any are included.
What stays out of a trajectory set
| Secrets and credentials | Passwords, API keys and access tokens are removed from every log. |
| Partial paths | Runs with missing steps or no outcome are left out, since they cannot be judged. |
| Tool outputs whose terms forbid reuse | AI tool runs are included only when the provider terms allow it. |
| Customer systems the company only operated | Logs from a client's own systems belong to that client. |
More on this: how personal details are removed, who carries the risk and how you get paid.
An example, invented for illustration
Consider an IT managed services company with fifty desk staff. For seven years its ticket system and remote management tool have logged every password reset, new user setup and printer fix: the request, each command run, the result and whether the ticket was reopened. Joined by ticket number and cleaned of client names, user names and any credentials, five thousand complete runs of three task types go into a sample for the owner to review. Logs from clients who own their systems stay out.
Questions owners ask about this
What is the difference between a trajectory and a workflow?
A workflow describes how a task should be done. A trajectory records one time it was actually done, step by step, including the detours. Labs use both, and a set that pairs them is stronger.
Do failed attempts have value?
Yes. A run that went wrong, with a record of how it was fixed, shows a model how to recover. Mark the outcome so a buyer can tell success from failure.
Can we license logs from our internal AI assistant?
Possibly. Check the terms of the model provider behind it and remove anything staff or customers typed that identifies them. The accepted or rejected verdict on each run adds value.
Didn't find your question? Ask us directly.
Keep reading
Related kinds of records
- Project and task history: Jira, Asana, Linear, Monday, Trello, ClickUp and Airtable comments and status changes, not task titles.
- Code and pull request history: The commit, the review and the failed fix before the working one, with secrets and borrowed code removed.
- Evaluation datasets: Exam questions, graded work and checked test cases from real jobs, kept private so they can measure a model.
Industries that keep these records
- IT services and MSPs: Tickets, runbooks and alert-to-fix histories; client secrets and client contracts set the limits.