When the bank wants an AI agent, write the rules it has to work inside.

Two worked cases, a regulatory crosswalk and a test plan

When the bank wants an AI agent, someone has to write the rules it works inside. This pack is for the BA who gets that job. Two worked cases at fictional Northgate Bank run through every file with matching IDs. A payment repair assistant is mostly rule-based and shows a complete specification. A complaints triage assistant reads unstructured letters, chooses its own tools and drafts a recommendation for a person to decide: a real agent. The guide helps you tell which you have.

£29including UK VAT (£24.17 + VAT)

File won't open or is missing? Email us and we'll put it right.

Written by Jennifer Brandt, working senior BA in UK financial services · 25 years in BA and consulting, including a Big Four firm.

Buy AI Agent Requirements Pack, £29

One payment, no subscription. Sold by Quantum Sphere Consulting Limited; your statutory rights apply.

Cover of the AI Agent Requirements Pack, version 1.4.

What’s inside

13 files: 8 Word, 3 Excel, a Start Here PDF and a licence file. Version 1.4.

Requirements for AI Agents guide (Word): read-this-first box, triage, specification, permissions, oversight and testing
Agent spec quick start (Word, one page): ten decisions to make before you write a specification, each pointing to the file and section
Two worked specifications (Word): payment repair with 25 requirements and 14 acceptance criteria; complaints triage with 10 requirements, 12 controls and 12 planned test cases
Specification template (Word): ten sections with guidance
Use-Case Triage workbook (Excel): hard stops, value and risk scoring, and an agent-fit check
Oversight and Controls workbook (Excel): 29 risks, 33 worked controls, each marked "prompt instruction only" or "enforced by the application", and an agent inventory
Evaluation workbook (Excel): 61 planned test cases, 14 red-team cases, a repeat-run gate and a daily canary
UK Regulatory Crosswalk (Word, one page): reading routes on PRA SS1/23, the FCA Consumer Duty, SM&CR and UK GDPR
Model-risk briefing note (Word, one page): what to tell your model-risk team
AI on the Assignment agreement (Word), a Start Here guide (PDF) and the licence

Who it’s for

BAs and senior BAs asked to specify, test or govern an AI agent, especially in regulated change.

How it fits

AI for BAs in Financial Services covers using AI yourself. This pack covers the other side: specifying an agent the firm will build or buy. It uses the same traceability habits as Starter and Complete. It is not a substitute for your firm’s AI, data or model-risk policies. Where your firm is a bank, ask whether SS1/23 model-risk principles apply to the agent you are specifying.

What it is not

Not legal, regulatory or model-risk advice, and not software. The regulatory pages are reading routes: check with your firm which rules apply. The test cases are a plan for you to run; no results are included. Northgate and its people are fictional.

Why the human checkpoint matters

In the Bank of England and FCA survey of 118 UK financial firms, 75% were already using AI, and 46% of those using or planning to use it reported only partial understanding of the AI they use. In the same survey, 55% of AI use cases had some automated decision-making; 24% of those were semi-autonomous and only 2% fully autonomous (p.4). 84% of firms had an accountable person for their AI framework, but most reported three or more accountable persons or bodies (p.5). Someone has to specify the human checkpoint. Bank of England and FCA, Artificial intelligence in UK financial services, survey of 118 firms, published 21 November 2024.

The FCA’s October 2026 publication, AI in financial services: good practice and areas for improvement, lists the risk of human review of AI outputs becoming a “box-ticking” exercise, rather than meaningful oversight, as an area for improvement.

Each pack is a ZIP of Word, Excel and PDF files. One person may adapt them for their own work; they cannot be redistributed.