How Should Line Managers Set Boundaries for Staff Using Approved AI Tools?
Line managers should turn the firm’s AI policy into a short, task-level agreement: which approved tool may be used, what information may enter it, what output must be checked, and when a person must take over. They should ask staff to demonstrate the workflow before independent use and revisit the boundary when the task or tool changes.
Key takeaways
- Approval of a tool does not approve every use of it.
- Define the task, data boundary, checker and escalation route together.
- Use a real work example to confirm staff can apply the boundary.
- Review the agreement after tool or workflow changes.
A team may have access to an approved AI assistant but still face uncertain day-to-day choices. A manager needs to convert general policy into decisions staff can use during a live task. The boundary should be documented alongside the normal workflow and agreed with relevant control owners.
Where the boundary is needed
Permission to open a tool says little about whether a person may paste a customer record, draft an answer or act on a recommendation. List the tasks the team performs and identify the points where a mistaken output or inappropriate data entry would matter. The manager should use the firm’s existing classification, access and approval rules.
Use existing supervision methods
Managers already use authorised task lists, quality checks and escalation paths for new processes. Adapt those methods: record the approved tool and purpose; specify permitted information; name the person who checks the output against source records; and state which cases require specialist review. A short example is more useful than a vague instruction to “use AI responsibly”.
Where AI helps
Within an approved environment, AI can help draft internal summaries or organise questions for a colleague. Its fluent answer is still a draft. Staff should compare material facts with the source and keep a record where the workflow requires one. The manager decides whether the task is suitable; the tool does not grant permission.
Make the boundary usable
Ask each user to work through one representative case and explain when they would stop. Capture the result in a team instruction that names the tool version or service, task, permitted inputs, required review and escalation contact. Recheck it when the tool or workflow changes. This is an editorial practice recommendation, not a prescribed regulatory checklist.
Example
A London investment operations team is given an approved assistant for drafting internal reconciliation notes. The manager permits redacted exception descriptions, bars customer identifiers under the firm’s policy, and requires an analyst to compare each draft with the trade record. A mismatch goes to the existing exception queue. The team practises one case before using the tool independently.
FAQs
-
Can a manager authorise any use of an approved tool?
No. The manager works within the firm’s tool, data and task approvals; new use cases go through the relevant owner.
-
Does every output need the same review?
No. Set review in proportion to the task and the consequences of an error, using the firm’s control framework.
-
What should be recorded?
Record the permitted task, input boundary, review step and escalation route in the team’s normal instructions.
AI Learning for Finance Teams
AI learning modules designed to develop practical AI capability through short, facilitated modules built around real financial services scenarios
