The first impressive AI-generated report can make the investment decision look easy. Notes become a coherent narrative in minutes. A difficult paragraph reads more clearly. The management pack appears closer to completion.
The real test comes later, when somebody asks whether the narrative is correct, whether the figures agree with the approved records and whether the explanation is supported. Those questions determine whether the speed has commercial value.
I see a useful role for AI in reporting when management defines the work around it. The tool can assist with preparation. An accountable person must still own the report's conclusions and the decisions made from them.
Choose the part of the report that benefits from assistance
“Use AI for reporting” is too broad an instruction. A report contains different tasks: assembling data, calculating results, identifying exceptions, interpreting causes, drafting commentary and approving the final document.
These tasks do not carry the same consequence or require the same evidence. Turning approved notes into a clearer draft can be relatively straightforward to review. Asking a tool to infer the reason for an unexplained cost increase is more demanding. Allowing its output to trigger a payment or contractual instruction introduces a different level of authority.
Start with a defined task and specify its boundaries. Identify the information the tool may use, the form of the output and the decisions it is not authorised to make. The person responsible for the final report should help design these boundaries before the pilot begins.
NIST's Generative AI Profile is a cross-sector resource that complements its voluntary AI Risk Management Framework. It provides a useful reference for considering risks across the use of generative AI. The operating arrangement still needs to be designed for the actual business task.
Give the draft a controlled set of evidence
The report should begin with approved records and a clear reporting period. If the input contains inconsistent dates, duplicated transactions or unclear status definitions, a fluent summary may conceal the weakness rather than resolve it.
Provide the relevant figures together with their definitions and known limitations. Distinguish approved facts from provisional estimates and unresolved questions. Preserve the source references so that a reviewer can trace a material statement back to the record supporting it.
For a monthly project report, this might mean supplying the reconciled cost position, current programme, approved changes and a separate list of matters awaiting decision. The instruction should require the draft to preserve those distinctions. A proposed variation must not become an approved instruction through the way a sentence is written.
Sensitive business information belongs only in an environment approved for that use. Management should establish the permitted data, access and handling arrangements with the appropriate technical and legal input before expanding beyond suitable non-confidential material.
Review the claim, the calculation and the implication
A review should examine more than grammar. It needs to test three different aspects of the output.
First, establish whether the statement is supported. A supplier delay recorded in the programme may be a fact. A claim that the delay resulted from poor supplier performance may require evidence the report does not contain.
Second, check the numbers independently against the approved records. Recalculate material totals, percentages and comparisons. Confirm the units, currency and period. A calculation can be arithmetically correct while comparing the wrong populations.
Third, assess the management implication. An increase in expenditure may reflect additional authorised scope rather than deterioration in delivery. A lower cost may reflect deferred work rather than improved efficiency. The reviewer needs enough subject knowledge to recognise those differences.
Consider a draft that describes a project as under budget because cash paid is below the approved budget. That statement overlooks unpaid commitments and remaining work. The correction requires understanding the project's financial position, not simply improving the wording.
The reviewer should be able to reject an unsupported conclusion and return it for correction. A sign-off step becomes weak if the person signing lacks either the information or the authority to challenge the draft.
Measure the complete reporting effort
A pilot should compare the time and quality of the complete workflow. Include preparation, checking, correction and approval, as well as generation.
If drafting falls from two hours to twenty minutes but review requires two additional hours, the business case needs reconsideration. Conversely, a modest reduction in drafting time may be worthwhile if it also improves consistency and makes exceptions easier to identify.
Record the types of error encountered and their consequences. Repeated problems with figures suggest a different response from repeated problems with tone. The task may need better inputs, a narrower instruction, a different tool or a decision to retain a manual step.
Use realistic acceptance criteria before expanding the process. These should cover the accuracy required, traceability of material claims, handling of missing information and the responsible approval. Passing a few polished demonstrations is weaker evidence than reliable performance across ordinary and difficult reporting periods.
The report that reaches a director or client should have a clear owner and a defensible basis. AI can help that person prepare the work more efficiently. Its value is established when the whole process produces dependable information with less effort, and management remains able to explain how the conclusions were reached.