Agentic accountability
Who is answerable when nobody approved it?
Your agents already decide what a customer can afford, confirm who is on the line, agree terms, and route hardship cases — on live accounts, with nobody approving each call.
When one of those decisions is challenged, can you prove what it did? Most firms can’t — and have no reason to realise it until someone asks.
The shift
None of it crosses a desk any more. That was the point — and it is the problem. The approval you took out was not just a control. It was the evidence: a name, a timestamp, a person who could be asked.
Take the approver out and the liability stays exactly where it was. The proof walks out with them.
The gap
Your logs will not cover you
Two reasons, and both are already true in most firms.
Your tools record the wrong thing
The standard nearly every monitoring tool is built on records which tool was called — not what it reached, and not whether it read something, changed it, or deleted it. It was written for model calls. The agent layer, the part that touches your customers, simply is not in it.
Your evidence is already being deleted
Observability is billed by the gigabyte, so logs are kept for two to four weeks. Complaints arrive for years. By the time someone asks, the record is usually already gone.
Six months after an agent acts, most firms could not reconstruct what it did on a single account — and have no reason to realise it until the question lands.
The standard
“A useful answer is not enough if the basis for it cannot be reconstructed.”
Who actually asks
Dates and sources, not opinion.
The FCA sets out where this is going
The Mills Review, commissioned in January and reported on 6 July 2026, carries seven priority recommendations on AI in retail financial services. It describes a spectrum of human involvement, from Operator at one end to Observer at the other, where people watch outcomes and the system runs inside limits agreed beforehand, and advises firms to place their use cases somewhere on it. It frames permissions, monitoring, auditability and escalation as the conditions that make deployment in a regulated environment possible, not as safeguards added afterwards. And it accounts for why so many deployments stop short: a firm that cannot say who is answerable for a loss caused by an agent tends to insist on human confirmation at every step.
The Ombudsman asks for clarity
Responding to that review, the Financial Ombudsman Service said it would welcome clearer FCA expectations on record keeping, escalation to a human, and handling disputes where no human is involved. Its decisions bind firms in law, it took in roughly 214,600 complaints last year, and it determines cases on the evidence put in front of it. It also notes that AI complaints are currently rare, which is worth reading carefully: the mechanism is being prepared before the volume arrives.
The person on the hook
There will be no separate Senior Manager Function for AI. Accountability sits inside the regime that already exists, which means a named individual is answerable for what an automated system did on their watch. The Treasury Select Committee has asked the FCA for practical guidance by the end of 2026 on what assurance senior managers are expected to show.
APRA told Australian regulated entities in April 2026 that it expects risk and internal audit functions to hold tooling capable of independently assessing AI systems, agentic workflows included, and reported that most do not. Australia's automated decision transparency obligations begin on 10 December 2026. California's automated decision-making rules have been in force since 1 January 2026, covering decisions on providing or refusing credit, with a right for consumers to information about how the technology was used.
Almost every rule says keep records. Almost none asks whether those records would satisfy somebody who has no reason to take your word for it.
The fix
Two ways I can help
Not more dashboards. A record that still answers the question when it lands — and a clear view of where you stand today, before it does.
Evidence readiness assessment
A register of every agent and tool you run, what each one reaches, and a plain statement of what your current records can and cannot establish.
What you get
- A register of the estate — every agent, every tool, what each can reach and what it does when it gets there.
- A plain statement of what your records can and cannot establish, case by case, with the gaps named rather than glossed.
Price by the size of the tool catalogue
- Up to 50 tools£5,000
- 50 to 100 tools£9,000
- 100 to 150 tools£18,000
Larger estates scoped individually.
Price can rise above the band where tools are undocumented or records are spread across several systems. Either way, you get the figure before the work begins.
Start a conversation →The evidence engine
Your existing agent records, turned into a sealed account of what happened, which of your own rules held, and what the evidence cannot settle.
Four properties
- It carries nothing of yours. No prompts, no replies, no tool arguments, no outputs — tested, not merely intended. The record can be handed to an auditor, a client or a regulator without the underlying data travelling with it.
- Anyone can check it. A short standalone script reproduces the seal, with no access to my software and no need to trust me.
- It comes out the same every time. Identical input and configuration produce an identical file, so two parties in disagreement can each run it and arrive at the same place.
- It is tiny. A million agent actions compresses to under five megabytes, because none of the content is in there. Retention stops being a budget question.
Price licensed annually, by estate size
Get in touch and I will give you a figure.
Start a conversation →
About
Luke
LHC TechnologiesI came to this sideways. I spent years paying attention to human behaviour, for reasons that have nothing to do with technology, and the first thing I built was an app for understanding how people's patterns hold or break over time, to help them become more consistent.
Working on it, I kept returning to one question. If you can describe a pattern of behaviour precisely enough to record it, could you do the same for a system that acts on its own? That turned out to be the harder problem. Once software takes an action nobody approved, what it did becomes a matter of who you believe.
Four months later I was accepted into the SETsquared incubator to build the answer. That is what LHC Technologies does.
A record only I can explain is not evidence. So the one I produce can be checked by anyone, without my software and without taking my word for it. That is the whole point.
Start a conversation
Let’s talk about what your records can prove.
luke@lhc-technologies.co.ukThe leads come from conversations. This is the one way in.
Based in Bristol. Part of the SETsquared incubator.