Using AI to update the instructions people follow

A returned wheelchair and a missing receipt.

In July 2023, a man returned a wheelchair borrowed for a family member at a Hospital Authority hospital. He had written confirmation of the return, but the payment counter requested the original deposit receipt. The hospital eventually refunded him without it in September. [1 · §§1, 6–7]

The Ombudsman found that the hospital’s records and the documents presented were sufficient to verify the relevant details. The counter staff had followed the accounting procedure then in place. [1 · §§8, 15]

That distinction interests me. A service can be inconvenient even when a member of staff follows the instruction correctly. Improving it requires a decision about the rule and a way to carry that decision into everyday work.

The investigation, announced in September 2025, recorded HA’s agreement that borrowers would generally no longer need the deposit receipt for refunds under the new arrangements. The recommendations also addressed staff training, borrower information and consistent implementation. [2 · Announcement] [1 · §§16, 19]

My proposal starts after an organisation has approved a change like that. It concerns the work of updating instructions, rather than deciding whether a particular refund should be paid.

Start with the instruction that changes someone’s next action.

Imagine an approved change says a receipt is no longer a prerequisite for a routine return. The old counter card still says to send the borrower away if they have forgotten it.

A document search could find the word “receipt” in both versions. The useful comparison is more specific: one permits the checks to begin, while the other prevents them. I would ask the system to identify that difference and show the two passages beside each other.

Then I would ask what the replacement must preserve. Removing one document requirement may leave identity checks, return checks and exceptions in place. A shorter instruction still needs to carry those conditions.

For each instruction, I would record who acts, what triggers the action and which conditions apply. That makes it possible to compare differently worded versions. It also gives a reviewer something more useful than a warning that two paragraphs look different.

The inventory needs attention too. A leaflet that was never supplied to the review cannot be described as checked. I would keep “this wording agrees” separate from “all relevant copies have been reviewed”.

Find the conflict. Keep the conditions.

The example uses an invented equipment-loan service with one approved policy change. It shows three different review problems: an old requirement, an overbroad rewrite and a missing copy.

Build a review pack around one approved change.

I would begin with a small, named set of materials: the authorised policy, the relevant staff instructions and the customer-facing explanation. Record their owners, dates and locations. Include translations where they are used, with a reviewer able to check each language.

AI would propose a comparison of the actions and conditions in those materials. Each finding would carry the source passage, the instruction it conflicts with and a draft correction. A reviewer could accept, edit or dismiss the finding and record why.

A source with an unclear date or authority would need clarification. The most recently edited file may still be an unapproved draft. The service owner would identify which rule governs the review.

I would keep the release record simple. The approved wording, the person responsible for each copy and confirmation that the replacement is available belong together. Updating a master file does not establish that a printed card has been removed or an online page has changed.

For a physical notice, confirmation might be a dated check at the relevant location. For a webpage, it might be the published page captured after the change. The system would preserve that evidence beside the approved correction.

The communication output needs to match its reader.

The service owner needs the full comparison. A colleague at the counter needs a short instruction and the route for exceptions. A borrower needs to know what to bring and where to go.

I would work from the same approved change to produce those different explanations. An interactive page can show a team the decision points. A short film or presentation can walk through the changed process. The operational check then confirms that the materials in use carry the agreed meaning.

Test the accuracy of the change, and the effort it takes.

The first pilot could use service documents alone. There is no need to bring patient records into a test of instruction consistency.

I would compare the system against a careful manual review and a straightforward keyword search. Include a paraphrased old rule, a genuine exception and a document that was deliberately left out of the supplied set.

• Does it find the instruction that conflicts?Check missed conflicts and false alarms. A pile of weak warnings creates another review queue.
• Does the proposed correction preserve the rule?Check that it keeps the right audience, conditions and effective date. A fluent sentence can still change the meaning.
• Does it account for the copies it has not seen?Keep missing material visible. A clean comparison of supplied documents cannot establish complete coverage.
• Does the accepted change reach the point of use?Check the released copy and include the reviewer’s correction time in the effort comparison.

No time saving has been measured in this study. If review and correction take longer than the existing method, the design needs to change.

I would also test the finished instructions with the person who has to act on them. Give them a routine case and an exception, then ask what they would do next. That is a useful final check on both the system and the explanation it produces.