AI for report writing in a consultancy

Use it on assembly, never on facts

The safe way to use AI on a consultancy report is to treat it as an assembly tool: it builds the outline, turns your analysis into prose, checks the document against itself, and writes the executive summary last. It never supplies a fact, a figure or a conclusion.

That single rule — assembly yes, facts no — removes most of the ways report-writing with AI goes wrong, because the failure mode is not bad grammar. It is a confident sentence containing a number nobody can trace.

The order of operations that works

1. Build the evidence file first. Every figure, quotation and document reference, each with its source, in one place. This is human work and it is the foundation of everything after it.2. Outline before prose. Sections, the claim each section makes, the evidence supporting each claim. A model is a fast critic of an outline and an outline is cheap to redo.3. Draft sections from your own material. Feed the tool your findings and your evidence file, not a topic. Where the input is your analysis, the output cannot manufacture a conclusion you did not reach.4. Executive summary last. Written from the finished report, not from the brief. Most weak summaries are weak because they were written from the plan rather than the result.5. Consistency pass at the end. Same figure across sections, defined terms used consistently, cross-references that resolve, appendices that exist.

Where the numbers come from

A report is only as good as its weakest sourced claim. For Canadian data, Statistics Canada’s data tables are the default, and the Statistics Canada Open Licence permits reproduction and redistribution provided the information is reproduced accurately, its source is not misrepresented, and nothing implies that Statistics Canada endorses you or your use of it.

Those conditions are easy to breach accidentally by asking a model to "summarise the latest figures". A summarised statistic is a paraphrased statistic, and a paraphrased statistic has lost the geography, the reference period or the base that made it true. Type the number, cite the table, keep the reference period in the sentence.

Separate fact, assumption and opinion

The best short statement of report discipline in Canadian professional practice sits in an unlikely place: Engineers Canada’s guideline on the code of ethics. Its fourth principle requires registrants to provide professional statements that distinguish between facts, assumptions and opinions, and its interpretation is blunt — facts stated in professional documents must be supported by data or credited to a reliable source, and registrants should state what assumptions they are making in the absence of data.

It binds engineers, not consultants. But it describes precisely the discipline that stops a generated draft from doing damage, because the one thing a language model will not do unprompted is label its own uncertainty. If your report template carries three explicit registers — this is a fact, this is our assumption, this is our opinion — a drafted section that blurs them is visible on sight.

What goes into the tool, and what does not

Client documents are the client’s. Where a report draws on personal information, the Office of the Privacy Commissioner’s principles for generative AI expect a documented legal authority for the collection and use, and openness about it. Underneath, PIPEDA’s fair information principles limit collection and use to what the purpose requires.

In practice this means the interview transcripts get de-identified before they are summarised, the payroll extract does not go in at all, and the workbook containing customer records stays inside the client’s environment while you draft the commentary around it.

A worked example

A three-person consultancy is writing a 40-page operating review for a regional distributor. The evidence file is built first: 19 sourced figures, each with its origin recorded — the client’s own management accounts, two Statistics Canada tables with their reference periods, and the team’s time-and-motion observations.

The outline is drafted, argued over and cut from eleven sections to seven. Sections are then drafted from the evidence file and the team’s findings. The consistency pass catches three things: a headcount stated as 46 in one section and 48 in another, a defined term used loosely in the appendix, and a chart labelled with the wrong reference period.

The executive summary is written last, in ninety minutes, from the finished report. The recommendation in it was decided by the team in week two and has not changed since.

Questions we get asked

Can AI write the whole report if we give it everything?
It can produce something report-shaped. What it cannot do is decide what matters, and a report that treats every finding as equally important is a report nobody acts on. Prioritisation is the consultant’s job and it does not survive automation.

How do we stop invented citations?
By never asking for a citation. Ask the tool to write from the evidence file you supply and to flag any sentence where it needs a source you have not given it. If a reference appears that you did not provide, treat the whole section as suspect.

Does this actually save time?
It moves time. Outlining, first drafts and consistency checking compress; evidence-gathering and prioritisation do not. Firms that report no gain are usually the ones that skipped the evidence file and are now verifying prose line by line.