The situation in most financial-services teams
The appetite for AI is rarely the problem. Operations teams can see where analysts spend hours collating documents, servicing teams know which customer queries are repetitive, and someone has usually built a promising prototype. What stops the work is the question every second-line function asks: what exactly can this system do, who approved it, and how would we know if it went wrong?
If engineering cannot answer that in writing, the pilot stalls in review. If engineering answers it by claiming the system is “compliant”, the pilot fails later and more expensively. The useful middle ground is an engineer who designs the controls into the system, produces the evidence your reviewers need, and leaves the compliance judgement with the people accountable for it.
I designed and built performance-critical components of an AI agent stack for regulated financial services as Principal AI Architect at a UK fintech. This page describes how I apply that work to a scoped pilot for your firm.
What the work involves
Every engagement starts from one workflow, not a platform. Typical candidates are analyst support (preparing a KYC or due-diligence file for a person to review), servicing (drafting responses to routine customer queries for an agent to approve), fraud and alerts triage (summarising the evidence behind an alert, not deciding it), and internal knowledge retrieval over policies and procedures.
For that workflow I do four things.
- Tier every AI step. I use the published Tiered Governance Model: T0 read-only (summarise, classify, search), T1 advise (the system drafts, a person approves), T2 act under supervision (actions are logged, attributed and reversible), T3 act autonomously. First pilots almost always sit at T0 or T1. The tier determines the controls, not the other way round.
- Trace the data. Each source is mapped to its classification, its lawful basis as your data protection owner has recorded it, and the environment it may be processed in. Personal data that does not need to reach a model does not.
- Make approvals mechanical. Where a person must approve, the system cannot proceed without that approval, and the approval is recorded. I follow the Substrate Pattern: permissions and approval boundaries are enforced in the runtime, not stated in a prompt.
- Evaluate in terms reviewers recognise. An agreed set of real or representative cases, expected outputs, and failure categories that matter to your firm, such as a missed adverse-media hit or an unsupported statement to a customer.
The signature deliverable, illustrated
You receive the working pilot, plus a control register your reviewers can read without reading code. Illustrative example, not taken from a client:
| Step | Tier | Control | Evidence produced | Approver |
|---|---|---|---|---|
| Retrieve customer file and screening results | T0 | Read-only service account scoped to case | Retrieval log with document IDs | Data owner (one-off) |
| Draft due-diligence summary | T1 | Every claim linked to a source document | Draft plus citations | Analyst, per case |
| Flag missing documents to customer | T1 | Template-bound message; no free-text send | Approved message and sender | Team lead, per message |
| Update case status | T2 (later phase) | Reversible within 24 hours; daily sample review | Action log with reviewer | Operations risk owner |
How acceptance is judged
Acceptance is agreed in the scope before the build. It normally has three parts: the pilot meets the evaluation thresholds your reviewers set on the agreed case set; every control in the register can be demonstrated live; and the evidence pack has been reviewed by the named risk and compliance owners. Their decision to approve, revise or stop is theirs. My acceptance criterion is that they have what they need to make it.
Ownership and handover
The workflow belongs to a named business owner from the start, and the system to a named engineering owner who pairs with me during the build. At handover they hold the code, the runbook, the evaluation harness and the control register. If the pilot proceeds, those owners decide whether to scale it, and any tier change goes back through your own approval process.
When to choose something else
If the need is turning an agreed AI policy into staff workflow controls across operations, see AI enablement for regulated operations. If the pilot is an agent that must take actions in your systems, governed AI agent implementation goes deeper on that build. If you want senior capacity inside your own team, use the contract AI engineer route. Other sectors are compared on the industries overview.