The easiest way to make an AI programme look active is to collect a long list of use cases. The easiest way to make it underperform is to fund those use cases without redesigning the work around them.
For investment firms, a practical AI transformation should begin with operating domains, data and accountability—not with a generic chatbot or an agent searching for a problem.
1. Choose an operating domain
An operating domain is a connected set of workflows with a shared outcome, such as:
- fund operations and administrator oversight;
- investor onboarding and servicing;
- investment reporting and client communications;
- research and knowledge management;
- finance, expense and management reporting.
Working at domain level allows the firm to understand upstream data, downstream decisions and repeated controls.
2. Map the work as it really happens
Documented procedures often omit the emails, spreadsheets, judgement calls and informal workarounds that keep a process running.
Map:
- triggers and deadlines;
- source systems, files and documents;
- manual preparation and rekeying;
- business rules and tolerances;
- exceptions and escalation paths;
- review and approval authority;
- evidence and retention requirements;
- volume, cycle time and recurring pain.
This reveals where conventional automation is sufficient and where AI may add value.
3. Prioritise by value, feasibility and control
Each candidate use case should be assessed on three dimensions.
Value
- How much skilled time is consumed?
- Does delay affect clients, risk or decision-making?
- Will the capability be reused across workflows?
Feasibility
- Are the inputs accessible and sufficiently consistent?
- Can output quality be evaluated?
- Can the solution integrate into the existing workflow?
Control
- What happens if the output is wrong?
- Can material actions remain subject to review?
- Is the data permitted for the proposed model and environment?
- Can decisions and changes be traced?
A high-value use case with unclear data rights or no workable review mechanism is not ready for implementation.
4. Fix the operating foundation
AI transformation often exposes problems that already existed: duplicated data, unclear document versions, undocumented adjustments, ambiguous responsibility, and controls performed only through individual memory.
The right response is not to hide those problems behind a model. Establish approved data sources, permissions, workflow states, exception categories and accountable owners first.
5. Prototype with realistic work
A proof of concept should process a realistic sample of the workflow—not a carefully selected demonstration file.
Evaluation should include:
- normal cases;
- missing or inconsistent inputs;
- unusual document layouts;
- historical examples of material exceptions;
- data the system is not permitted to use;
- cases that must be escalated to a specialist.
Define acceptance criteria before the test. Useful measures may include preparation time, exception detection, extraction accuracy, reviewer correction rate, turnaround time and percentage of work requiring escalation.
6. Design human review explicitly
“Human in the loop” is too vague to be a control. The operating design should state:
- who reviews;
- what they review;
- when review is mandatory;
- which threshold triggers escalation;
- what evidence is retained;
- who has final authority;
- how corrections improve future performance.
For low-risk drafting, review may be lightweight. For investor, accounting, regulatory or payment-related outputs, approval and segregation of duties may need to be much stronger.
7. Move from tool to operating model
Production adoption requires more than model performance. The firm needs:
- a named business owner;
- operational support and incident handling;
- version and change control;
- permission and data-handling rules;
- evaluation after model or workflow changes;
- training and user guidance;
- a route for exceptions and feedback;
- clear vendor and concentration-risk decisions.
NIST’s AI Risk Management Framework organises AI risk activity around Govern, Map, Measure and Manage. The principle is highly relevant: trustworthy AI requires an ongoing management process rather than a one-time technical approval.
8. Build reusable patterns
The long-term advantage comes from reusable components:
- approved data connectors;
- document-classification methods;
- permission patterns;
- evaluation datasets;
- exception and review components;
- audit and evidence structures;
- prompt, rule and workflow versioning.
These patterns reduce the cost and risk of each subsequent use case. They also prevent every team from creating its own ungoverned AI process.
A phased roadmap
Phase 1 — Discover
Select a domain, map workflows, identify data and controls, and prioritise use cases.
Phase 2 — Prove
Prototype one bounded workflow with realistic inputs, measurable acceptance criteria and explicit human review.
Phase 3 — Implement
Integrate with approved systems and permissions; document responsibilities, controls and operating procedures.
Phase 4 — Operate
Monitor usage, quality, exceptions, incidents and business value. Improve through controlled releases.
Phase 5 — Scale
Reuse common data, evaluation and governance patterns across the next operating domain.
The operating principle
An investment firm should not ask, “Where can we add an AI agent?” It should ask, “Which operating domain should work differently, and what combination of data, rules, automation, AI and human judgement will produce a controlled result?”
Sources and further reading
This article is general information and is not legal, accounting, regulatory or investment advice.