Claims · committees · trades

the AI advised. your name is on the decision.

g-nosis makes AI safe to adopt in regulated firms. It keeps what you need to stand behind each decision: the advice, the authority it ran under, the challenge and, later, the outcome.

The paradox

the AI is astonishing. the payoff is missing.

seconds

for AI to read a claim file, draft the letter and recommend the payout.

83%

of firms say AI has not changed their productivity in three years. 87% say the same of their headcount.

NBER working paper 34836, revised Oct 2026: nearly 9,000 executives in seven countries.

Computers did this in the 1980s: everywhere except the productivity figures, until firms changed how the work got done.

01 · The mess

AI agents are the new spreadsheets.

Built or deployed in every team, duplicated, owned by no one, running on messy data and on know-how nobody wrote down. No one can tell whether the decisions AI shapes were right.

2/3

of large firms let staff build or deploy their own AI agents, in some form.

EY Responsible AI Pulse survey, Oct 2025.

2 of 31

In 2013 regulators told the biggest banks to bring their risk data, spreadsheets included, under control. Ten years on, two of 31 fully had.

Basel Committee, BCBS 239, January 2013; progress report, 28 November 2023.

That mess, not the models, is why AI has not yet transformed insurers and banks.

02 · The risk

AI took seconds. the complaint came years later.

The mess will cause incidents. A customer can take a complaint to the ombudsman years after the event. By then the model has changed, the handler has moved on and the notes are thin. Your firm still has to answer, and the FCA has told MPs that 'I did not understand it' is not a defence.

FCA evidence to the Treasury Committee, October 2025; HC 684, 20 January 2026, para 12.

150,000+

AI agents in use at the average Fortune 500 firm by 2028. One mistake repeats at that scale.

Gartner prediction, 28 Apr 2026.

46%

of the UK financial firms surveyed have only a partial understanding of the AI they use.

Bank of England and FCA, Artificial intelligence in UK financial services, Nov 2024: 118 firms.

£7.5bn

the FCA's estimate of the redress motor finance lenders will pay for undisclosed commission, on up to 12.1 million agreements going back to 2007.

FCA policy statement PS26/3, 30 Mar 2026. Parts of the scheme are suspended while a legal challenge is heard in February 2027.

03 · The answer

every decision, kept whole.

Your people, the AI agents that advise and act for them, your systems and data, and what your firm already knows meet at every decision. g-nosis is the harness that keeps it whole: who decided, on what advice, under whose authority and which rule, and what happened next.

the reason.

Why the exception was made. A reason is required to overrule the advice in the app, and agreeing with it is recorded too.

the replay.

The record as it stood, on any date: what was known, and when, under the rules in force that day.

the precedent.

Similar past decisions, acting as precedent where needed, each with its reason and how it turned out.

In the app

one screen per decision.

The facts, your rules, similar past cases and an AI reading of the relevant clauses. The person decides and signs it. Whoever proposed a decision cannot approve it in the app.

four controls on every decision.

why?

A reason is required to overrule the advice in the app.

two

Whoever proposed a decision cannot approve it in the app.

signed

A signed decision is replaced, not edited.

replay

Any decision replays as it stood, under the rules in force that day.

04 · The limits

you set the limits. AI agents act inside them. a person decides the rest.

Each AI agent works within limits a named person signs: what it may decide, up to how much, until when. Inside them it acts at once, so nobody waits. Anything outside goes to a person.

Any AI agent, any host

whoever built the AI agent, a named person answers for its decisions.

Copilot Studio, Agentforce or home-built: every AI agent's decisions land in one record. Each works to a mandate a person signed: what it may decide, up to what limit, until when. Keep Agent 365 for the AI agent; g-nosis reads its registry and keeps the decision.

Tested with NVIDIA OpenShell. OpenShell holds the boundary; g-nosis holds the decision. A mandate given in g-nosis compiles into OpenShell policy. OpenShell checks the action, not the amount; g-nosis checks the limit. g-nosis also checks its record against OpenShell's audit log, so a missing decision is found, not assumed.

The Mills Review, published by the FCA, July 2026, p.87: decision-making "shifts from point-in-time approval to pre-authorised, bounded and revocable mandates".

05 · What it adds up to

every decision teaches the next one.

Each new case arrives with similar past decisions, acting as precedent where needed, each with its reason. Your rules are written in plain words, so AI agents learn your firm's direction.

why did we decline the water damage claim in January, and was it right?

Ask why a year later, in plain English: the answer cites the record.

That's g-nosis

auditable. accountable. answered.

g-nosis is the harness that makes AI safe to adopt. Every decision a person or an AI agent makes is kept with its reason, who answers for it, the authority it ran under and its outcome. Any AI agent, any host, without slowing anyone down.

We blend data science, the g-nosis fabric and deep technical expertise, alongside your existing infrastructure.

The fabric

who decides, who proposes, what they draw on.

People

a named person answers for every call.

They decide with the facts, your rules and similar past cases in front of them. A reason is required to overrule the advice in the app.

AI agents

every AI agent works to a mandate a person signed.

They read documents, draft rules and fill in forms, and ask before they change anything. Outside its mandate, a person signs it off.

Systems

your records stay in your systems.

g-nosis reads Microsoft 365 and any system with a web API, and looks records up when a decision needs them.

Knowledge

your rules and your past decisions.

Rules in your own words that change only when a person adopts them, and signed decisions kept as precedent.

Worked decisions

the same five steps, wherever decisions get made.

The question, the options, the challenge, the call, the outcome. A claim, a committee's call and a trade are kept the same way.

General insurance · claims handling

a water damage claim, £9,000 reserved, loss dated in December.

01
the question

The claim arrives with the policy terms, the reserve and the nearest past claims attached.

02
the options

Settle at £8,400 on three matching precedents. Decline on the exclusion. Refer for inspection.

03
the challenge

The handler pushes back on settling: it looks like a gradual leak, which the policy excludes, and this book sees them every winter.

04
the call

Referred for inspection, not settled. The handler’s reason is recorded against the exclusion and signed as the call is made.

05
the outcome

Inspection confirms the leak and the claim is declined on the exclusion. The outcome is recorded against the decision, so the next handler sees how it turned out.

Without it

The reason for a decline lives in one handler’s notes. When the complaint comes years later, there is little to answer with.

Asset management · investment committee

cut the property allocation, or hold it.

01
the question

The papers, the exposure and the last three related calls reach the room before the meeting does.

02
the options

Hold at 12%. Cut to 7%. Cut to zero and re-enter later.

03
the challenge

Risk pushes back on the full exit: liquidity in the vehicle is worse than the paper assumes, so the terms change.

04
the call

Cut from 12% to 7%, conditional on a six-month liquidity review, with the risk challenge and the reason for the call recorded.

05
the outcome

Recorded twelve months on, against what the room expected, not against what everyone now remembers.

Without it

Five years later someone asks why. Today that answer is a search through minutes and mailboxes, and the person who knew has left.

Asset management · portfolio implementation

an index change forces a rebalance, and the desk has to choose how to trade it.

01
the question

The desk sees the constraint, the tracking error it is defending, and the last comparable action.

02
the options

Trade in full at the close. Stage it over three days. Seek an exemption from the constraint.

03
the challenge

Compliance blocks the full size pre-trade. The desk records why it wanted the size, not just that it was stopped.

04
the call

Staged over three days, with the compliance block and the desk’s reason recorded.

05
the outcome

The execution result is recorded against the reason for staging, not only against the benchmark.

Without it

Pre-trade blocks and overrides are logged, but rarely in a form anyone can search a year later.

Illustrative examples

Not model governance

models get tested. decisions get questioned.

The model's file

the model, tested.

  • Accuracy tests
  • Bias checks
  • Drift monitoring
  • Version history
The decision's file

the decision, kept.

  • What the AI advised
  • Who challenged it, and why
  • The call, and the named person who made it
  • The rule in force, and the outcome

The model is rarely what goes wrong. The mess around it is.

The founding pilot

start with claims. ninety days to a record you can question.

No new committees, no change to who decides, no re-platforming. Before week one: your security review, DPIA and data-feed approvals; acceptance tests agreed before you sign. The record is yours, whatever you decide next.

How the pilot runs

Weeks 1 to 2 · map

Every AI agent, model and spreadsheet in one decision area, and who answers for each.

Weeks 3 to 10 · capture

Every decision recorded as it is made. We sit alongside your team, then hand it over.

Weeks 11 to 13 · proof

Precedent at work, the evidence pack your supervisor would ask for, an outcome review, and a report for your risk committee.

The founders

one of us put AI to work in investment management. the other built graph systems for enterprises.

Sam Dunkley

Sam Dunkley

Co-founder

Six years at Vanguard, ultimately its fintech strategist; a CFA charterholder; previously PwC analytics and MBDA.

Previously · Vanguard · PwC · MBDA
Chris Hobson

Chris Hobson

Co-founder

Twenty years building for demanding rooms: seven at AT&T as lead engineer, eight at Expero leading the development practice for enterprise graph-technology clients.

Previously · AT&T · Expero

bring us one decision you have to explain.

A claim referred above authority, a complaint upheld, a trade outside its limit, a committee's call, a decision an AI agent made that nobody can answer for. We will show you the record it would have left.

45 minutesYour decision, not a canned demo