The board paper below is written out in full, because a description of a good board paper is not much use to anybody.
Yours will not read like this one, and your numbers will not match these.
The counts here are the ones that follow from the model answers to Parts One to Five.
What the paper has to do is the same four things whatever your numbers are:
- say what it asks for,
- rest on a count you made yourself,
- give the Chief Technology Officer a way to be right, and
- stay inside what a person four weeks in post could be given.
1. The Board Paper
MERIDIAN GROUP PLC
BOARD RISK COMMITTEE
Meeting of 19 November 2026
Paper 7: Artificial intelligence: What the group uses, who is accountable for it, and what is asked of this Committee
Author | AI Governance Lead, Group Legal |
Sponsor | Priya Raghunathan, Group Head of Data Protection and Privacy |
Classification | Internal |
Position stated as at | 5 October 2026 |
This paper asks the Committee to | Note one thing, delegate one thing and request two things. It asks for no approval and no money. |
1. Summary
Meridian uses sixteen systems that make or materially shape a decision about a person using artificial intelligence, and through Talentflow it also builds and sells two of them to fifty-three external customers.
Nine sit in the top band of the classification at Annex B, and seven of those nine have no named individual who has accepted the risk that remains after the controls.
Work is in hand: An inventory exists for the first time, the Group AI Policy of February 2026 provides requirements to work from, and Group Legal and Privacy hold an enquiry from the Information Commissioner's Office, received 18 September, about automated decision-making at Meridian Financial Services.
The Committee is asked to name one executive accountable for artificial intelligence across the group, and to request two things back by its first meeting of 2027.
2. The count
AI systems evidenced across the nine entities, with a source against each row
16
Tagged as production AI on the IT asset register
4
In the top band of the classification at Annex B
9
Of those nine, with a named individual who has accepted the residual risk
2
Of those nine, with no named individual
7
The difference between sixteen and four is one of definition, not of fact. Four is what the IT asset register tags as production AI, and under that definition four is correct. Sixteen is every system that makes or materially shapes a decision about a person, wherever it is recorded, and several were found in finance and procurement records rather than technology records.
Both counts are honest answers to their own question. This paper works from the second, because it is the one that matches the group's risk.
Two of the nine carry a named acceptor: The Head of Credit Risk for Arbiter, through the model risk framework that already existed for prudential reasons, and the Head of Pricing for Pricewise Life.
On two of the seven that do not, the Group HR Director has declined to accept the residual risk in writing, on the ground that it is not hers to accept. She is right. It is a correct answer to a question put to the wrong person, and paragraph 5.2 is the remedy.
3. What is already in hand, and what it covers
The Group AI Policy, issued February 2026, sets seven requirements. Requirement 4.3 states that any new AI system must be approved by the Technology and Change Committee before deployment.
Four systems have gone into service since the policy was circulated on 11 February, and the Committee's minutes record no approval for any of them.
The policy has no named owner and no review date, and has not been communicated below director level.
Model validation runs annually for credit models because the Prudential Regulation Authority requires it. It covers Arbiter and reaches nothing outside Meridian Financial Services.
The ISO/IEC 27001 certificate is current to July 2028, and its scope line reads "the provision of shared IT infrastructure and service desk operations from the Leeds facility", so it does not reach Talentflow, Meridian Health Solutions or the Dubai estate.
That is a statement of scope and not a criticism of the certificate.
The Group Risk Register carries 47 open risks and none concerns artificial intelligence, models or automated decisions.
The Risk Appetite Statement of March 2026 is silent on all three, and none of the eleven audits in the FY2026/27 internal audit programme is on the subject.
On 2 September a generative customer service assistant issued goodwill credits totalling £4,180 to 23 customers in forty minutes, following a support email that manipulated its instructions.
The service desk found it in ninety five minutes and a change went in two days later, which is a good operational response. It was recorded as a service desk ticket, closed on 4 September, and has not appeared in any risk report.
4. What is not known
Three of the sixteen systems do not appear on the Record of Processing Activities.
Facial recognition at the Leeds operations centre went into service in June 2026 as a variation to an existing contract, so it was never treated as a new deployment.
A workforce analytics platform has two years of purchase orders and no entry in any technology record.
A generative tool is in daily use by a support team of twelve people in Kraków, on a corporate card, in a location that appears in no group compliance document.
None of this is hidden. Each was recorded correctly by the process that recorded it, and no process held the whole.
5. What the Committee is asked to do
5.1 Note | the count at paragraph 2, and that sixteen and four differ by definition. The paper taken to the Technology and Change Committee on 9 September is accurate against the register. This paper is accurate against the group. |
5.2 Delegate. | Name one executive accountable for the group's use and supply of artificial intelligence. Two seats can hold it. The Group Chief Technology Officer, who owns the technical environment, wrote the Group AI Policy and already chairs the monthly committee where these decisions are taken. Or the Group General Counsel, who owns the policy framework and the regulator correspondence. Either is workable. Neither being named is what produced the seven empty cells at paragraph 2. |
5.3 Request | that Group Risk open one entry on the Group Risk Register for the group's use and supply of artificial intelligence, owned by the executive named under 5.2, and that the next review of the Risk Appetite Statement say whether this sits inside technology risk, conduct risk or both. The Register holds 47 entries. This paper asks for one. |
5.4 Request | that the executive named under 5.2 bring back to this Committee, at its first meeting of 2027, a named individual against each of the nine systems at Annex B, or the reason there is not one. |
6. Cost
This paper asks for no money in FY2026/27. The ninety day plan at Annex C runs on the post already funded in cost centre GL-004.
Two lines do need money and neither is requested here. The two standards referenced have to be purchased, and Group Legal will absorb that.
Independent assurance cannot be supplied by this function for itself, and is recorded at Annex C as a known limitation with one route out of it: A proposal to Internal Audit before the FY2027/28 programme is set.
Certification to ISO/IEC 42001 was quoted in July 2026 at between £46,000 and £71,000 for two entities. It is not requested.
There is not yet a management system to certify, and a certificate bought ahead of one would evidence nothing.
7. When the next paper comes
One page into this Committee's papers each quarter, beginning with its next meeting, and a full paper annually.
The terms of reference at Annex D extend the mandate of the Technology and Change Committee, so they go to that Committee, which is where the authority to change them sits, and not to this one.
Annexes.
A: The inventory, sixteen rows, with the source of each.
B: The risk classification.
C: The ninety day plan.
D: Draft terms of reference.
Annexes are for the record and members are not expected to read them.
What this paper does.
The first paragraph is four sentences and the fourth is the ask, which is the only structure week 19 gives you for an opening. The strongest sentence in it is a count, because a count cannot be answered by agreeing with it.
It says the Chief Technology Officer is right, in his own terms, in paragraph 2 and again in 5.1, and then moves on without a second sentence about it. It offers his seat first among the two candidates for the executive owner, which costs nothing and changes the whole reading of the paper.
It puts the September incident in the evidence section rather than the summary, so it lands as a fact rather than as an argument.
It asks for nothing that this Committee cannot give, and it says out loud that the one thing it cannot give will be taken somewhere else. And it names the next paper, which is what turns an event into oversight.
Two small things a former regulator will notice. It fills its two sides and no more, and everything else is in an annex it tells her not to bother with.
And it says the date on which its facts were checked, which is 5 October, five weeks before the papers are due, because a count without a date is not a count.
What a weaker paper does.
It opens by explaining what artificial intelligence is.
It says the group is exposed to regulatory action, which is a prediction and a hostage. It says the September paper was wrong or misleading, which makes the subject a person.
It asks for a budget, a committee and a head of function in month two. It leaves the empty cells out because they looked unfinished. Every one of those is understandable, and every one of them costs the next twelve months.
2. The Programme Plan
One page, covering the first ninety days. This is the plan, not a description of it.
Where the function sits, and what that home costs.
Group Legal, inside the Data Protection and Privacy team. What that home gives: assessment machinery that already runs, an organisation already in the habit of consulting somebody before a system goes live, and a reporting line through the Group General Counsel.
What it takes away is scope. Talentflow's position as a provider, the Leeds sorting machine as a regulated product, and Claimsight's accuracy are none of them privacy questions, and each has to have its jurisdiction argued before the work can start.
The cover built for that is a standing monthly item on the Technology and Change Committee and a quarterly line into the Board Risk Committee. Neither is a new forum.
What it costs, and who pays.
Line | FY2026/27 | Whose budget |
|---|---|---|
My time | The post as already funded | GL-004, Group Legal |
The two standards | Low hundreds of pounds | GL-004, absorbed |
External counsel, two or three times, on Talentflow's provider position | Within the existing external counsel line | Group General Counsel |
Training and literacy | Nothing this year. Month ten, inside the mandatory training cycle that already runs | Group HR |
Tooling | Nothing | Not applicable |
Running the controls | The largest line on this page, and not mine | The business owner of each system |
Deliberately not requested this year: ISO/IEC 42001 certification, quoted July 2026 at £46,000 to £71,000; any headcount; any tool.
Who owns what.
Eight standing jobs, three columns.
Job | Does it | Oversees it | Assures it |
|---|---|---|---|
The AI policy | Office of the Group CTO | Executive named by the Committee | Nobody |
The inventory | AI Governance Lead | Priya Raghunathan | Nobody |
The risk register entry | Deborah Lange | Board Risk Committee | Elaine Prosser, not yet scheduled |
Assessments | AI Governance Lead | Priya Raghunathan | Nobody |
Supplier reviews | Meera Sandhu, with AI Governance Lead | Meera Sandhu | Stuart Fenwick, security only |
Incidents | Rob Merrick | Stuart Fenwick | Nobody |
Training and literacy | AI Governance Lead, with Fiona Hargreaves | Fiona Hargreaves | Nobody |
The report upward | AI Governance Lead | Sponsor | Nobody |
Two findings, written down rather than tidied away.
Supplier reassessment at renewal has no owner on any row, which decides whether the supplier questionnaire was worth building.
And my own name appears in the first column five times and the third column never, which is correct, because I cannot assure my own work.
The route out of that is one AI audit in the FY2027/28 internal audit programme, proposed with a risk, a rationale and a scope before the programme is refreshed.
What this function does not own.
It does not run the models, it does not own information security, and it does not own the supplier relationship, only the assessment inside it.
The rhythm.
Weekly, thirty minutes, inside the privacy team's existing slot: clears the queue. Monthly, standing item on the Technology and Change Committee, second Wednesday, a body that already meets and already has a quorum: Decides.
Quarterly, one page into the Board Risk Committee papers: Reports. Annually, alongside the review the privacy team already runs: changes the machine.
Nothing on this page has a calendar of its own, which is what carries it through a bad quarter.
Three measures.
Headline | Guardrail beside it | ||||
|---|---|---|---|---|---|
Coverage | Proportion of the inventory assessed and classified | Systems discovered after they went live | |||
Cycle time | Median elapsed days from request to decision | Decisions reopened, or approved with conditions nobody checked | |||
Decisions | AI items decided at the monthly committee | Proportion of AI items noted rather than decided |
One measure is not offered. Harm avoided leaves no data, and refusing to invent a number for it is what makes the other three believable.
What does not happen in the first ninety days. A policy rewritten from scratch, which arrives in month four. A tool purchase. A maturity assessment, which produces a score rather than a decision.
3. The two answers
The first question. "If one of these systems were producing a bad outcome for a customer today, who in this organisation would know, and how would they know?"
A model at 145 words.
For five of the sixteen, somebody would know. Complaints catches what a customer notices and writes in about. The service desk catches what breaks. Model validation catches Arbiter, because the Prudential Regulation Authority requires it and Credit Risk built the function years ago.
For the other eleven, the honest answer is whoever happens to be in the ticket queue that morning.
We have a worked example from six weeks ago. A generative assistant paid out £4,180 to 23 customers in forty minutes. It was found in ninety five minutes, fixed in two days, and closed as a service desk ticket. Operationally that is a good response. It never reached a risk report, because nothing in our process turns an AI failure into an incident.
One named executive and one route that treats these as incidents would change the answer to your question for all sixteen.
The second question. "Suppose this Committee agrees with every word of it and still does not fund it. What happens then?"
A model at 147 words.
Then very little changes, because the plan asks for no funding this year. It is built to run on the post that is already budgeted.
A decline changes sequence rather than direction. I would keep the inventory and the register entry current, take one system all the way through to a named acceptor so there is a worked example rather than a process document, and bring a funded proposal back with the FY2027/28 budget, with a year of evidence behind it instead of an argument.
What would not get done, and I would rather say it plainly than discover it later: Supplier reassessment at renewal, the provider documentation Talentflow needs, and independent assurance, which I cannot provide for my own work.
I would record what was decided, by whom and on what date, and attach it to the file. Then I would help make the decision work.
What these two answers do.
The first refuses the reassuring version, gives a number, and lands on a real event rather than a hypothetical one.
The second treats a decline as a legitimate decision taken by people entitled to take it, which is what week 19 teaches, and it names what will be dropped, which is the thing that proves capacity has been thought about. The last two sentences are the record and the sentence after it.
A governance lead who writes the note and withdraws is no use to anybody.
4. The rubric
Score yourself on each criterion. Be strict.
Criterion one: The ask
Score | Description |
|---|---|
Yes | The paper names what it wants using note, approve, delegate or request, and the ask appears in the first paragraph. Every item is something the Board Risk Committee can actually do under its own terms of reference. The ask is proportionate to four weeks of tenure: one named executive, one register entry, one report back, and no money |
Nearly | The verbs are used but one ask sits outside the Committee's powers, usually a budget approval or a decision about a named system. Or the ask is proportionate but buried on page two |
Not yet | The paper explains artificial intelligence, or asks for a framework, a committee, a budget and a headcount, or ends without a sentence beginning "the Committee is asked to" |
Criterion two: The evidence underneath it
Score | Description | ||
|---|---|---|---|
Yes | The paper rests on counts the learner made: systems evidenced, systems in the top band, systems with a named acceptor and systems without. The empty cells are on the page rather than tidied away. The existing furniture is described by what it covers and what it does not, including the certificate scope line. The September incident appears as evidence rather than as an argument | ||
Nearly | The counts are there but the acceptor column is missing, so the paper describes risk without describing who is carrying it. Or the unknowns are mentioned in general terms rather than named | ||
Not yet | Assertions in place of counts. Or a reassuring paper, where the gaps were left out because they looked unfinished |
Criterion three: The colleague, and the plan
Score | Description |
|---|---|
Yes | The four is described as defensible under its own definition, in one sentence, with no second sentence about it. The Chief Technology Officer is offered as a candidate for the executive owner rather than treated as the obstacle. The plan names what the home costs, attaches every meeting to a diary that already exists, puts a name in every cell including the learner's own where that is honest, and gives three programme measures each with a guardrail. The answer about a decline is calm and names what will be dropped |
Nearly | The September paper is handled fairly but the plan invents new forums, or the measures are activity counts with no guardrail beside them, or no measure of the programme is distinguished from a measure of a model |
Not yet | The paper says or implies that the Committee was misled. Or the plan is a wish list of everything that will exist by day ninety. Or the answer about a decline reads as a warning about what will go wrong without the money |
What you now hold
Three artefacts, made on a real organisation, from evidence that did not agree with itself: An inventory with sources, a risk classification with names against it, and a board paper that asks for something a board can give.
That is the file the last lesson of the programme told you to send. Not a polished version. The version you have.
Submitting this completes the Certified AI Governance Expert Foundation.
Part Six Capstone: Model Answer and Rubric
Read this against what you submitted. The point is not the score. The point is the gap.