Whose Formula Is It Now? Guardrails for AI Agents Working Inside UK Finance Workbooks
Photo: Infoacrosseverything, CC BY 4.0, via Wikimedia Commons
A familiar file with a new attack surface
A workbook that has survived a decade of month-ends can feel like furniture. It is, in practice, a small application: formulas, links, macros, lookups and a set of conventions that only two people fully understand. Give an AI agent permission to open it and you have not simply added a helper. You have added a user with no institutional memory, no sense of which column is sacred, and the ability to write to hundreds of cells in seconds.
That is the governance question now landing on the desks of UK finance and operations teams. The assistant is often bought at department level, trialled on a real file, and judged on the quality of its summary. The controls that would make the trial safe are rarely part of the purchase. This review is deliberately practical: it sets out what to test, what to document and what to switch off before the first prompt is sent.
Draw the permission map first
Begin with access, not capability. Ask three questions of every agent: which files can it see, which parts of those files can it change, and under whose identity does it act? A named service account with a documented owner is preferable to a personal login borrowed from a leaver. Scope should be limited to folders and, where the tool allows, to specific sheets or ranges.
Distinguish reading from writing. Summarising a board pack is a different risk from recalculating a forecast. If the agent can only read, say so in the register; if it can write, require an approval step. Check sharing links, external connections and any route by which the assistant could reach a file outside the intended folder. Under UK GDPR, the permission map is also a data protection artefact, not merely an IT convenience.
Provenance: can the agent show its working?
A summary is only useful if it can be traced. Test the agent with a direct question — which sheet, which range, which timestamp produced this figure? — and compare its answer with the workbook. Agents can invent a plausible range as easily as a plausible number. Require cell-level references in outputs where the tool supports them, and treat unsourced figures as drafts.
Pay attention to the difference between a retrieved value and a derived one. A total that the agent has calculated should carry its method. A blank cell and a zero are not the same thing, and a good test is to plant both in a copy and see whether the agent conflates them.
Rollback is a control, not a comfort
Version history is often assumed to be a safety net until it is needed. Before an agent writes to a live file, confirm that restore points exist, that they cover the relevant period, and that someone has rehearsed a restore on a copy. Shared workbooks with co-authoring can complicate this: a rollback may remove a colleague's legitimate edits along with the agent's mistake.
Name the person authorised to trigger a rollback and set a time limit for doing so. A file that can be restored within an hour is a managed risk; one that can only be restored after a call to the vendor is not.
Audit trails and the human in the loop
Every write by an agent should be attributable to a time, a user and, where possible, a prompt. Not every organisation needs to retain full conversation logs, but the changes themselves should be reconstructable. Reconcile after each session: compare the agent's stated changes with the file's actual change history. Segregation of duties matters here as much as in any finance process. The person who approves the agent's access should not be the only person reviewing its output. For management accounts and anything feeding statutory records under the Companies Act, the review should be explicit and recorded.
What to switch off before the first prompt
Some features are better disabled than supervised. Macros and external data connections, links to other workbooks, automatic refresh from production systems, email or send actions, and any auto-apply setting that writes suggestions without confirmation should be off by default in a first trial. So should access to columns containing personal data unless the trial genuinely requires it.
If the agent can reach a live shared file, prefer a copy. The cheapest guardrail in the review is a sandbox that nobody can mistake for the real thing.
A short test protocol
Use an anonymised copy of a real month-end workbook, not a toy example. Run a standard set of prompts: summarise the variances, explain the largest movement, list the exceptions, recalculate a total. Add a few adversarial prompts that ask for a figure that does not exist, or that invite the agent to overwrite a protected cell. Compare every output with a manual check. Record failures as well as successes. A trial that produces a clean report but no test log has not been governed; it has been demonstrated.
The documentation pack
Keep a one-page register per agent: the file, its owner, the data classification, the permissions granted, the approver, the test results, the rollback contact and the next review date. That single page is the difference between a controlled pilot and an unlogged experiment.
The spreadsheet is not going away. For many UK firms it remains the most honest record of how the business actually runs. Giving an agent the keys is a reasonable ambition, provided the door has a lock, the changes have a trail and someone knows how to put the room back as it was.