Agent logs, UK view
Can I sell my AI agent's logs? Yes, the runs you own, with three checks
Updated
Yes, for the runs your company owns and has cleared: one buyer licenses qualifying agent trajectories for $1–2M a year. Not the personal data inside them, and not where your model provider's usage rules forbid training another model on its outputs.
Ask about my agent logs 6 of the 19 agent platforms we read write that they will not train on what you send. Tell us where your traces sit and we will write back about yours, as a licence, a closing company's sale, or an introduction through your adviser.
What a buyer pays for agent logs
micro1 says every agent task produces a trajectory, valuable for showing how real work gets done, and that it licenses qualifying trajectories for $1–2M annually, with the company's approval. Its programme looks for established companies with 30+ full-time employees and substantial historical operational data.
Source: Who takes AI agent traces and trajectories
Magma says it packages an agent's interaction data for AI labs and pays a revenue share as it licenses, with the company keeping control of the scope. Ooak Data turns real company data into training environments for agents, transforming names, dates and proprietary content.
Whose logs are they under UK law?
If your agents work on your customers' data under their instructions, the processor rules apply. Article 28 of the UK GDPR says a processor handles personal data only on documented instructions from the controller, and the ICO's guidance adds that a processor acting outside those instructions, or for its own purposes, becomes a controller for that processing, with every duty that brings.
Source: UK data law on passing on agent logs
If the logs are your own company's, the people in them may never have been told their words could train an AI model. UK GDPR Article 5(1)(b) bars processing personal data further in a way incompatible with the purpose it was collected for, and Article 8A, in force in full since 5 February 2026, sets the test for a new purpose: any link between the two purposes, the context and your relationship with the person, the nature of the data, the likely consequences for them, and safeguards such as encryption or pseudonymisation.
What your model provider's usage rules allow
Most of what an agent log contains was written by a hosted model, and the providers' own usage documents limit what may train on it. OpenAI's usage policies, revised on 21 October 2026 with effect from 20 November 2026, say its models and outputs must not be used to prepare or supply training material for models intended as alternatives to or substitutes for OpenAI's; its terms of use for the UK and Europe already bar using output to develop models that compete with OpenAI. Anthropic's usage policy forbids using outputs to train an AI model without prior authorization from Anthropic, and xAI's acceptable use policy forbids using any output to develop machine learning models that compete with it.
Source: Model providers' usage rules on training with output
What to take out before a buyer sees a sample
- Customers' runs held as a processor, unless the customer instructs the licence in writing.
- Names, contact details and anything else that identifies a person, unless the reuse passes the Article 8A test.
- Runs made on a model whose usage rules forbid training a buyer's model, until the provider agrees.
- API keys, credentials and secrets the agent saw in its tool calls.
micro1 says personal and identifying information is removed or transformed before a company's data is shared with third parties, but the decision on what may leave is yours first.
A note for your CTO or founder
"Our agents' runs may be worth a recurring licence to an AI lab. Before anyone prices them we'd need to check which runs are ours, what our model provider's usage rules allow and what personal data comes out. Worth half an hour?"
Tell us what your agents do
Which agents you run, on which model and for whom: say so on the enquiry form, and see what your trace store keeps on LangSmith data retention.