Sector
AI consultancy for financial services
Model risk, Consumer Duty, and a supervisor asking how the decision was reached. The blocker is explainability and data control, not capability.
The constraint
Why financial services AI projects stall
Model governance predates generative AI, and does not bend for it. A model that influences a customer outcome needs an owner, documented validation, ongoing monitoring and an explanation second line can challenge.
Generative models are awkward against all of it. Output varies between identical runs, reasoning cannot be inspected, and vendors change the model without notice. That last one hurts most: a validated model that silently becomes a different model has invalidated its own validation.
Then the data. Customer records, transaction histories and unpublished positions make an internal assistant useful. Nobody wants them leaving a controlled environment.
Where it works
Where AI earns its place
Ordered roughly by how quickly they get approved. The first is usually the right first project: it is the least contentious.
Policy, procedure and regulatory retrieval
Answers from the handbook, policy and regulatory material, cited to source. Low risk, high volume, measurable. A good first build.
Complaint and correspondence triage
Classify, route and summarise complaints, flagging vulnerability and Consumer Duty issues. The machine reads; a person judges.
Document and KYC extraction
Structured extraction from statements, certificates, applications and letters. Measurable against ground truth, so defensible to validate.
Suitability and file review support
Surface files for human review against defined criteria. A prioritisation aid, not an assessment. Your second line will care about the difference.
The gate
What assurance will ask for
We produce it during delivery. Cheaper, and more likely to be accurate.
- Model documentation covering purpose, data, limitations and monitoring
- Version pinning, so a validated model does not silently change
- Logged inputs and outputs that can reconstruct any individual decision
- A clear statement of where the model does and does not influence customer outcomes
- Human oversight that is resourced realistically
- Data residency and retention position, written for second line and audit
Illustrative scenario. A composite of how an engagement typically runs, not a specific client.
Illustrative: a mid-size lender
Advisers waste time searching policy documents for current criteria. The material is internal, non-personal and changes often, and superseded versions circulate.
Classification shows no personal data in the corpus, so a private UK deployment is enough; self-hosting is unnecessary. The build: retrieval with mandatory citation, an effective date on every document, and an owner accountable for retiring old versions. That last change does most of the work.
The evidence pack covers version pinning, the evaluation set, measured accuracy, and a written position on why the system does not directly influence customer outcomes. That document gets second line approval quickly.
Questions from this sector
Can generative AI ever be used where it affects customer outcomes?
Yes, but the bar rises steeply and the architecture changes: version pinning, full logging, validation against a held-out set, drift monitoring and a real human decision point. Most firms do better starting where the model supports staff, proving the governance there, then moving closer to the customer.
How do we validate a model whose outputs vary between runs?
Validate the system as a whole: a held-out evaluation set with reviewed answers, performance measured with variance reported, pinned versions so what you validated is what you run, and monitoring for degradation. A different approach, not an impossible one.
Does our data leave the UK?
Not if the architecture is designed so it cannot. We set the data boundary before choosing components, and write a residency and processing position your second line and auditors can rely on.
Next step
Start with a 30-minute call
No deck. Tell us the problem. We will say whether AI is the right tool, what it would take and roughly what it would cost.