Human-in-the-loop telemetry.
A behavioural assurance layer for firms using AI in regulated decisions. It shows whether people genuinely review AI recommendations or rubber-stamp them, and turns that into evidence for risk teams and regulators.
The question
When AI recommends and a person approves, can a firm prove the person really reviewed it?
Built with
Lightweight web SDK, browser extension and event API; behavioural analytics engine; append-only audit trail.
Status
In development
Most AI in financial services still has a person in the loop: the AI recommends, a colleague reviews and approves. That review is what firms point to as oversight.
Under volume and time pressure it can shrink to a click. A handler accepts an AI-drafted refund in two seconds without opening the evidence, and the system records a human review. If the AI was wrong, nobody can show it was checked.
Regulators are raising the bar. The FCA’s Mills Review, published in July 2026, maps a five-level spectrum of AI autonomy and recommends clarifying how the Consumer Duty and the Senior Managers regime apply as decisions spread across systems. Firms need oversight they can evidence, not assume.
Sentinel is an independent oversight layer. It doesn’t manage cases or replace workflows: it sits alongside the tools colleagues already use and looks at how AI recommendations are reviewed.
It captures passive signals, such as whether the evidence was opened, how long key documents were read relative to their length, and whether the AI’s draft or figures were changed. From these it produces a Meaningful Oversight Score for each queue and team, with alerts when review starts to look like rubber-stamping.
Sentinel captures signals in three ways, without replacing core systems: a lightweight web SDK for in-house screens, a browser extension for third-party case tools, and an event API for AI recommendations and outcomes.
Behind them sit four analytics engines: engagement and vigilance, anchoring and automation bias, seeded ground-truth checks, and Consumer Duty and vulnerability assurance.
Time on case alone is easy to game and says little, so Sentinel measures what was opened, read and changed, adjusted for how complex the case was.
Seeded test cases, synthetic cases with known errors, are mixed into live queues so they are handled like real work, and never reach customers. That gives an objective measure of whether reviews catch what they should, rather than relying on how long someone spent on a case.
Sentinel is in development. The working prototype shows the full loop: a rushed approval flagged within seconds as low-engagement review, a careful one recognised, and both rolled into team trends and an exportable audit log.
The pilot will set real baselines, and we’ll publish what it shows, including what doesn’t work.
Sentinel assures the review process, not individual people. Reviewer identity is pseudonymised and reporting defaults to team level, so it shows whether oversight is working without becoming a tool for monitoring staff.
Every review event is written to an append-only audit trail, so risk teams can show how oversight was evidenced and when. Real-time prompts to handlers are optional and set by the firm.
Sentinel was developed through FinTech Scotland’s FRIL Responsible AI innovation call and is now being co-developed with a UK financial services partner in a four-to-six-month pilot: map the process, set up passive capture, calibrate baselines, then validate the supervisory reporting with risk and compliance teams.
It applies wherever AI recommends and people approve: complaint handling and redress, credit decisions and, as AI agents take on more of the work, the people supervising them.
Sentinel came out of FinTech Scotland’s FRIL Responsible AI innovation call, where ClearSky is working with a UK financial services partner on meaningful human oversight of AI.
As AI moves from assistant to delegate, ‘a human approved it’ stops being enough. Sentinel proves the review really happened.
Most tools audit the algorithm. Sentinel audits the human and AI decision loop: whether the person approving an AI recommendation genuinely engaged with the evidence. It complements model risk monitoring rather than replacing it.
It’s designed to assure the process, not to performance-manage people. Reviewer identity is pseudonymised, reporting defaults to team level, and the firm controls who sees what.
No core system is replaced. Sentinel captures signals alongside the case tools colleagues already use, and a pilot starts with one process.
Speak directly with our team in Edinburgh or Glasgow about your platform, data or AI initiative.
See what else we’re building and testing in Future Labs, our self-funded R&D.