AI Consulting Report Writing: From Findings to a Client-Ready Draft

Build a findings log first, give every finding an ID and an evidence tag, let the AI draft only from that log, and run a separate number-check pass before anything reaches the client. Write the executive summary last, from the finished sections. The AI shapes sentences; the log decides what is true.

Invented facts rarely look invented. They look like a tidy rounding ("about a quarter" for 22%), a swapped word ("average" for "median"), or a sensible-sounding benchmark nobody supplied. A reader who trusts the report finds them months later, in front of the client's finance director. The workflow below is built so each of those gets caught while it is still cheap.

Follow me on Instagram@sagnikteaches

Start with a findings log the AI cannot add to

The log is a table you build before any drafting, with one row per finding. If you have interview notes to synthesise first, the tutorial on AI prompts for management consultants: interviews, synthesis, slides covers that step. Columns that earn their place are the ID, the finding in one sentence, the evidence type, the source, the figure with its denominator and period, and your confidence. An illustration, for a review of order dispatch at a 40-person distribution business, using invented figures:

Connect on LinkedInSagnik Bhattacharya

Three habits make the log work. Write the denominator and the period into the figure column every time ("405 of 1,840 orders in March", never "22%"), because that is what gets lost first. Record the statistic ("median", not "time"). And log the weak findings too, with their weakness stated; a finding you can call low-confidence on page one will not embarrass you on page twelve. Then freeze the log. From this point the AI may use it but never add to it, and any new fact has to be logged, with a source, before it can appear in the report.

Subscribe on YouTube@sagnikteaches

Evidence tags: no tag, no sentence

An evidence tag is the ID of the finding that supports a sentence, in square brackets, which stays in the working draft until the final read. The rule is blunt: every sentence that states a fact, figure or quote carries a tag, and a sentence without one is either cut, turned into a question, or labelled as your judgement. Two further tags help. [A-1] marks an assumption (a number you estimated, with the reasoning written beside it), and [J] marks professional judgement, which the client is paying for but which should never be mistaken for data.

Here is the difference in practice. Untagged: "Dispatch is slow because stock records are unreliable." Tagged: "The median order took 31 hours to dispatch against a 24-hour promise [F-02, F-05]. Nine of the 12 people we interviewed named stock accuracy as the main delay [F-03], and 405 of March's 1,840 orders (22%) needed a manual stock check [F-01]. Our judgement is that stock accuracy is the largest single cause [J]." The second version tells the client what is measured, what was said, and what we think, in that order. It is also longer, and that is fine.

From log to outline: three messages, not eight findings

Ask the assistant to group the log into themes, and make it show which findings support each theme and which complicate it. The complications matter, because a report that lists only supporting evidence is the one a sceptical client picks apart.

Below is my findings log. Each row has an ID, a finding, an evidence type, a source, a figure and a confidence level.
Group the findings into three or four themes for a client report. For each theme give:
(1) a one-sentence message the evidence supports;
(2) the IDs that support it;
(3) the IDs that weaken or complicate it.
List any finding that fits no theme. Use only the IDs and wording in the log. Do not add findings, figures or context of your own.

A reply to that prompt, abridged and illustrative:

Theme A: Stock accuracy drives most avoidable delay. Supports: F-01, F-03, F-04. Complicates: F-06 (cut-off timing is a separate cause).
Theme B: The 24-hour promise is not built into the working day. Supports: F-02, F-05, F-06, F-08. Complicates: none.
Theme C: Returns run as an informal side process. Supports: F-07 only. Note: one low-confidence finding; this may be an observation, not a theme.
Unassigned: none.

The assistant did two useful things here: it separated cut-off timing from stock accuracy, and it admitted that theme C rests on one weak finding. Your edits are the other half. A weaker run of the same prompt might call theme C "a significant issue", which the evidence would not support, or fold F-06 into theme A because both concern delay. Cut any such adjective, restore the separation, and decide yourself whether returns deserve a section or a line in an appendix.

Drafting one section at a time from tagged findings

Draft a section at a time, giving the assistant only the findings that belong to it. A whole report in one go invites the model to fill gaps with plausible material; a 250-word section from three findings leaves it no room.

Write the section "Stock accuracy" for the client report: about 250 words, plain business English, no adjectives of degree.
Use only findings F-01, F-03 and F-04 from the log.
After every sentence that states a fact or a figure, add its ID in square brackets, for example [F-01].
If you want to say something no ID supports, write it as a question to me instead.
Keep the statistic words (median, of, per cent of what) exactly as the log has them.

What comes back (abridged, illustrative):

In March, 405 of 1,840 orders (22%) needed a manual stock check before dispatch [F-01]. Nine of the 12 people we interviewed named stock accuracy as the main cause of delay [F-03]. In the warehouse we found two separate stock spreadsheets, one kept by the pickers and one by the dispatch supervisor, and the two did not always agree [F-04].
Question for you: do you want a figure for the time a manual check adds to each order? The log does not hold one.

Read it as an editor would. The structure is right, the question at the end is exactly what the prompt invited, and the tags are in place. But "and the two did not always agree" is not in F-04. The log says the sheets exist, not that they conflict. The model inferred it, because two spreadsheets and a stock problem tell a tidy story. The tag makes the problem findable: the sentence cites F-04 and F-04 does not say that. Either go back to the warehouse notes and, if the disagreement was observed, log it as a new finding with its source, or cut the clause. Embellishments like this one are a common kind of invention in AI-drafted reports, and a tag-by-tag read is the dependable way to find them.

Headings that tell the reader the finding

Once a section is verified, rewrite its heading so that it carries the finding. "Stock management" tells a busy reader nothing; "22% of March orders waited on a manual stock check" tells them what to look for in the paragraphs below it. Ask the assistant to propose headings from each section's verified text, then check that each one traces to a tag. Keep them under about 15 words. The test is that a client who reads only the headings can follow the argument: stock records first, then the cut-off, then returns. Two more before-and-afters from the same report: "Order timing" becomes "Orders placed after 15:00 wait for the next day's pick"; "Process documentation" becomes "The procedure promises 24 hours, and nothing in the day is organised to deliver it".

The number verification pass, line by line

Treat this as a separate job from drafting, done after the sections are complete, ideally in a fresh chat so the assistant is not defending its own draft. It has six steps.

1. Ask the assistant to extract every number, percentage, date, period and quantity word ("most", "about a third", "double") from the draft, one per row, with the sentence it came from.
2. Match each row to a log ID and figure. No match means "unsupported".
3. Recompute every percentage and difference in a spreadsheet, not in the chat.
4. Check what each number counts: the denominator, the period, the statistic (mean or median) and the unit.
5. Check that parts add up and that categories you add together do not overlap.
6. Compare the numbers in the recommendations and the executive summary with the body, one by one.

Here is the register from an illustrative first draft of the dispatch report, about 900 words, with the problems the pass found:

Eight problems in 900 words, and not one of them looks wrong on a read-through. The 22% versus "about a quarter" slip makes the finding sound worse than the data; the invented benchmark makes the client's performance look poor by comparison with nothing. The arithmetic row also shows why step 3 happens in a spreadsheet: 7 divided by 31 is 22.6%, which rounds to 23%, and the draft's 8 hours and 26% were internally consistent but built on the wrong subtraction.

Finish with the reverse check. Ask the assistant which log figures appear nowhere in the draft. A figure that is in the log but missing from the report is sometimes the most important thing you forgot to say.

When interviews and data disagree

Disagreement is often the most useful result in a project. Suppose four of the 12 interviewees, all managers, said orders go out "the same day", while the March export gives a median of 31 hours. Log both (call the interviews F-09) and report both plainly: "Managers described same-day dispatch [F-09]; the order export shows a median of 31 hours [F-02]. We have not established why the two differ." An assistant's instinct is to smooth the gap with a sentence such as "perceptions differ from reality", which is a judgement dressed as a finding. If you hold that view, write it yourself and tag it [J].

To catch conflicts you have missed, run a one-line prompt before drafting: "List every pair of findings in the log that point in different directions, with the IDs and the reason." The reply is a short list of things to explain to the client, or to chase before the report is final.

Recommendations that cite their findings

A recommendation is a claim about the future, so it needs a different kind of evidence: the findings that justify it, an effect you can defend, and a label wherever the effect is an estimate. Let the assistant draft options, but make the format force the connection:

Two things in that table protect you. R-3 says plainly that it rests on a weak finding, and R-1's number is labelled as an assumption with the test that would settle it. The assistant will happily write "this will reduce dispatch time by 25%". Ask it for the range and the assumption behind it instead, and put your own judgement in the status column. The tutorial on writing proposals with AI in under an hour shows how the same scoping habits carry into the follow-on work.

The executive summary is written last

The summary comes last because the numbers in it change while you verify the body, and a summary written first goes stale without anyone noticing. The rule for it is the mirror of the evidence tags: the summary may contain only claims that appear in the body with the same figures, and every recommendation in the body appears in it once. Give the assistant the finished, verified sections and nothing else.

Below are the verified sections of the report and the recommendations table. Write an executive summary of no more than 180 words for a managing director who will read nothing else.
Order: the headline finding with its key figure; the two or three facts that support it; the recommendations in one line each; the single decision we need from the client.
Use only figures that appear in the sections, written exactly as they appear there. Keep the [A-1] and [J] labels on anything that is an estimate or a judgement.

Compare the two versions. In the illustrative first draft, before the verification pass, the summary said: "Orders take over a day to dispatch because a quarter wait on a manual stock check. Fixing stock accuracy will cut dispatch times by a quarter, and most staff agree." After the pass and the rewrite, the summary reads: "The median order takes 31 hours to dispatch against a 24-hour promise. In March, 405 of 1,840 orders (22%) needed a manual stock check, and nine of the 12 people interviewed named stock accuracy as the main delay. We recommend a single stock count, a later cut-off for stocked lines and daily returns processing. The first change could remove about half the manual checks (an estimate, to be tested in a two-week trial). We need a decision on the trial by the end of the month."

Every figure in the second version has an ID behind it, the estimate is labelled, and the last sentence asks for something. For the slide version of the same material, see AI for consulting work: from interview notes to findings deck.

What you paste into the AI, and what stays out

The log is the only material the assistant needs, and you can write it so it carries no names. Replace the client with "the client", interviewees with numbers, and anything identifying (a supplier, a product line, a person's role if it points to one person) with a neutral label. Keep the raw interview transcripts, contracts and anything marked confidential out of the chat entirely. Then check two things before you start: what your engagement letter or the client's own rules say about third-party tools, and what your plan's data terms say about retention and training. Rules differ by tool and by plan, so read the terms for the plan you actually pay for instead of assuming. If a client bars the use of AI tools on its data, the workflow still works by hand; only the drafting step changes. The tutorial on reusing past client work with AI without leaking client data covers the related problem of old material.

Which parts the AI drafts and which parts you write

Ten checks before the report goes to the client

1. Every factual sentence has a tag; search the draft for sentences with none.
2. Every number is in the register and matched to the log.
3. Each recommendation names its supporting findings and labels its estimates.
4. The executive summary matches the body, figure by figure.
5. Names, product names and the client's own terms are spelled as the client spells them.
6. Quotes from interviewees are anonymised unless you agreed otherwise.
7. Periods are explicit ("March", not "recently"), and tables say what, when and from where.
8. A colleague or the client contact reads the summary cold and says what they would do next. If they cannot, rewrite it.
9. You have read the whole report aloud once, cutting hedges and adjectives as you go.
10. The working tags are removed or turned into appendix references, and the findings log is attached so the client can trace any claim.

None of this is complicated, and all of it can be skipped under deadline pressure, which is why it is worth writing down and sticking to. A report where every claim traces to evidence is also easier to defend when the client pushes back, because you can open the log and show where each sentence came from.

Questions consultants ask about AI report drafting

Should I tell the client I used AI to draft the report?

Check your engagement letter and the client's own rules first; some require disclosure and some bar certain tools outright. Where nothing applies, a one-line note in the method section is honest and costs little, provided every figure was verified by you. Never present AI-drafted analysis as work you did not check.

Which assistant is best for drafting reports?

The method matters more than the model. Any assistant that accepts a pasted table or uploaded spreadsheet and keeps a long conversation straight will do. Test two or three with the same findings log and count the unsupported claims in each draft; the one that adds fewest wins, whatever its marketing says.

How do I stop it inventing benchmarks and citations?

Do not ask for them. The instruction to use only the log, the rule that every factual sentence carries an ID, and a verification pass that flags anything untagged catch most invented context. If you need an outside benchmark, find it yourself, log it with its source and date, and then it can be used.

How long does the whole workflow take?

For a report of ten to fifteen pages from a log of 30 to 50 findings, allow roughly a day of focused work. The log takes longest, drafting the sections an hour or two, the number pass about an hour, and the summary and final read another hour. Treat that as a guide, not a promise.

Further reads

• AI ESG Report Drafting for Sustainability Consultants — The same discipline where claims face outside scrutiny.
• Claude vs ChatGPT for Consultants: Which to Pay For and How to Test Both — One client task given to two assistants.
• Use of AI in Consulting: Where a Small Firm Should Start — Where a small firm should start with AI.
• How Much Time Can AI Save a Solo Consultant Each Week? — Realistic weekly time savings for one consultant.
• Best AI Tools for Management Consultants: Seven From Research to Decks — Seven AI tools for independent consultants from research to decks, with a confidentiality tier rule, costs and a worked two-week diagnostic.
• Consulting Pricing Models After AI: Fixed Fee, Retainer or Value? — Which pricing model survives AI-cut hours? Worked fixed-fee and retainer sums, a value-pricing test, scope limits and wording for telling clients.
• AI Brand Audit: A Template Branding Studios Can Sell — A ten-day brand audit for a music school, step by step, with the 12-page client deliverable a branding studio can reuse and sell.
• Consulting Engagement Letter: Scope, Fees and AI-Use Wording — A clause-by-clause engagement letter for independent consultants: scope, fees, change control, confidentiality and plain AI-use wording you can adapt.
• Consulting Utilisation Rate: Track It Weekly and Price Projects With AI — Billable versus available hours, bench time, a weekly AI roll-up from timesheets and a worked project price checked against your rate card.
• Using AI in Environmental Consulting: Draft Site Report Sections — Prepare site-report sections from controlled notes and lab results, with a traceable check of every number and conclusion.
• AI for Testing Psychologists: Report Drafts Without Pasting Raw Scores — How a testing psychologist can draft assessment report sections with AI from de-identified summaries, and what must never go into a chatbot.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: the worked example uses invented, illustrative figures for a generic distribution business; no external statistics are cited.