AI agents can do the work. But can companies govern them?

The promise and the problem

AI agents that can read contracts, route claims and answer customer questions are starting to appear in banking and insurance ^[1]. The appeal is clear: they can work around the clock, handle volume and learn from feedback. The difficulty is working out how to let them act without losing control.

The question is not whether agents can do the job. It is whether the organisation can explain what they did, and prove it ^[2].

Differentiated governance: not all tasks carry the same risk

The article argues that the way to govern agentic AI is to differentiate ^[1]. Not every task needs the same level of human oversight. A system that classifies incoming post can be allowed to run autonomously. A system that decides whether to grant a loan cannot.

That distinction is sensible. It reflects what regulators already expect for high-risk AI systems: that decisions about people, money or safety must be subject to human review ^[3]. The practical challenge is drawing the line in each part of the business and making sure the agent respects it.

Say a bank's agent is allowed to approve standard overdraft requests up to a certain amount, but must flag anything over that threshold for a human. The agent needs to know the rule, the human needs to see the agent's reasoning, and the organisation needs a record of both. If the threshold changes, or the model is updated, someone has to check that the new version still follows the rule.

The evidence question

When an agent acts, three questions follow. What did it do? On whose instruction? And which version of the model was running?

If the agent's actions are logged by the same system that runs it, and the logs can be edited by whoever operates that system, the record is easy to challenge. An external auditor, a supervisor or a customer may want to see proof that the log has not been changed since the event.

That is the evidence question, and it matters more as agents become more autonomous ^[4].

What this means for you

If your organisation is looking at agentic AI, the questions to ask are practical:

• Which tasks can the agent handle on its own, and which need a human decision?
• How will you show an auditor or a customer what the agent did on a given day?
• If the agent's reasoning or the model version is questioned later, do you have a record that cannot be quietly rewritten?
• When the model is updated, how will you show which version was running when?

Differentiated governance is the goal. Provable records are how you get there.

Sealing what the agent does

Here, the thing to seal is each instruction the agent receives, each tool call it makes and each result it returns. Sealing takes a fingerprint of the record, not the data itself, so the instruction and the result never leave your systems.

The fingerprint is sent to WitniumChain through the REST API, the SDK or the MCP server, and you get a witness ID back. When someone questions what the agent did, you share the record and the witness ID. Anyone you share it with can calculate the fingerprint again and check that it matches the sealed one. Change a single character in the record and it will not match.

What that shows is that the record of what the agent did, and when, has not changed since it was sealed. It does not authorise the action or show that the action was correct. It does not replace access control or monitoring. What it does is give you a record that can be checked by someone who does not have to trust you.

The content stays with you. The first 2,000 witnesses a month are free.

How it would work

1. The agent receives an instruction and your system seals a fingerprint of it.
2. The agent calls a tool, and the call and the result are sealed as they happen.
3. Later, if the action is questioned, you share the record. The person checking it calculates the fingerprint and compares it with the sealed one.

What to watch

Article 26 of the AI Act will require users of high-risk AI systems to have human oversight and to monitor the system's operation. For systems under Annex III, that applies from 2 December 2027 ^[5]. The requirement to inform employees before high-risk AI is used at work applies from the same date ^[6].

For agents that make decisions about people, the right to an explanation under Article 86 will apply from 2 December 2027 as well ^[7]. The person affected has the right to know the AI system's role and the main elements of the decision. Being able to show what the agent did, with which model version, will matter when that question is asked.