AI Tools for Strategy Consultants: Desk Research and Market Sizing

For desk research, a strategy consultant can use Perplexity to find cited leads, ChatGPT or Claude to extract and compare claims, and a spreadsheet to calculate market size. Check every number against the original dated source, reconcile units and definitions, and show the arithmetic and assumptions separately before a client sees the estimate.

The most useful division of labour is simple: search tools help locate material, a language model helps organise it, and the consultant owns the evidence and model. A citation next to an AI answer is a route to inspection, not a warranty that the linked page contains the claimed figure. A spreadsheet cell with a formula is easier to challenge than a polished paragraph with an unexplained total.

Define the market before opening a research tool

Write the unit you intend to count in one sentence. Is the question about annual sales, the number of potential buyers, installed units, or spending on one service category? Then state the boundary: which customer segment, which period, which products, and whether revenue is gross or net. If the client says “market for outsourced maintenance”, ask whether internal maintenance departments are included. The answer changes the numerator before a single source is found.

A practical brief has four fields: decision, target segment, measurement and exclusions. An illustrative kitchen fitter evaluating a new aftercare service might ask: “Should we offer an annual inspection plan to previous installation customers?” The measure could be eligible households willing to buy one inspection per year, excluding commercial jobs and customers outside the service radius. That is narrower than all home-improvement spending and makes research queries answerable.

Keep three labels distinct. Total addressable market is the whole defined demand under an explicit assumption; serviceable market is the portion the business can actually serve; and obtainable market is the share it might plausibly win. These are modelling definitions, not facts a search engine can supply. If you need a broader introduction to research quality, read how to check AI research citations before you gather data.

The Complete Excel Course with AI Integration

From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.

By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner

Buy now

Assign Perplexity, a language model and a sheet different jobs

Perplexity's help centre says Pro Search runs several searches across articles, academic papers and forums, and that every answer includes direct links to the original sources. That makes it useful for discovering candidate evidence, especially when a term has several meanings. Ask for the source organisation, publication date, exact table or passage, denominator and direct URL. Open those links yourself. Discard any item that cannot be tied to a passage in the original material, even if several AI answers repeat it.

ChatGPT or Claude can turn the pages you have already checked into a comparison grid. Paste short excerpts or permitted documents and ask the model to preserve quotation context, note differences in scope, and mark missing values. If confidentiality terms prevent uploading client material, use public extracts only. Neither model should silently fill blanks. A useful instruction is “write unknown rather than infer”. The tool's name matters less than whether its output exposes the trace from each claim to a document.

For a broad first sweep, the deep research modes in ChatGPT and Gemini plan a multi-step search and return a longer report. OpenAI's help centre says ChatGPT shows a proposed research plan you can edit before the run starts, can be pointed at specific websites or uploaded files, and returns a report with citations or source links; how many runs you get depends on the plan. Google says Gemini Deep Research can browse up to hundreds of websites and can also draw on Gmail, Drive and Chat. Two habits follow. Edit the plan so it hunts for the metric and definition in your brief, and check which private sources are switched on before you run a client question, because a report that blends a client's Drive files with public pages is hard to cite cleanly.

Here is the kind of plan edit that matters (illustrative). The tool proposes: “1. Research the home maintenance market. 2. Identify key trends. 3. Summarise major players.” The consultant rewrites it to: “1. Find dated counts of owner-occupied homes with a kitchen fitted in the last ten years. 2. Find published prices for annual appliance or kitchen inspections, stating what each price includes. 3. List every source with its year, unit and definition. Do not estimate market size.” The second plan produces a shorter report, but nearly every line in it can go straight into the evidence ledger. The first produces trend commentary that has to be unpicked before any number can be used.

Use a spreadsheet for the arithmetic. Put inputs in named rows, cite the source and date in adjacent columns, and keep assumptions in a different colour or column. When the market has several price points, calculate by segment and sum them. A single “market size” cell typed by hand is a review obstacle. For a comparison of search-oriented and conversational tools, see Perplexity versus ChatGPT for business research.

Build an evidence ledger that survives a challenge

For every external number, make one ledger row with claim, unit, original source, publication date, page or table, access date, extraction note and status. Status can be “verified”, “scope mismatch”, “superseded” or “unverified”. If the source is a PDF, record the page number visible in the document rather than relying on the viewer's scrolling position. If a chart lacks underlying values, record that you estimated from a chart and do not report false precision.

Prefer a direct measure of your defined market. A supplier's investor presentation may describe a broad category because that makes its opportunity look larger. An industry report may use customer expenditure, while an annual report uses the supplier's recognised revenue. Both can be correct and still be incomparable. Ask: Who was counted? What was counted? For what period? At what point in the value chain? Does the source report a stock, such as installed units, or a flow, such as annual purchases?

Illustrative ledger row: “Eligible installed kitchens; 8,000; installations; internal job records; extracted on 12 October; jobs completed over six years; repeat customers removed; verified for our own customer base.” A second row might say “inspection uptake; 15%; assumption; no external source; test as 8%, 15% and 22%.” This is more honest than presenting the 15% as a published industry rate. The two rows have different evidence status even though both feed the same formula.

When a source offers a range, carry the range into the model. Do not average two reports merely because they disagree. First check whether one counts buyers and the other counts transactions, or whether one uses a calendar year and another a financial year. If the difference remains, show two scenarios with their source definitions. A disagreement can reveal the client's actual decision boundary: perhaps the opportunity is attractive only under the broader definition.

From 8,000 past customers to a revenue range

The main worked example uses illustrative figures: a kitchen fitter with 8,000 unique past installation customers in its own records. A review of addresses and service capacity suggests that 6,000 are still serviceable. The firm is considering an annual inspection at $120, with a separate repair quote if work is needed. The consultant has no validated uptake survey, so models three uptake assumptions rather than claiming to know demand.

At 8% uptake, 6,000 × 0.08 = 480 inspections and 480 × $120 = $57,600 annual inspection revenue. At 15%, the result is 900 inspections and $108,000. At 22%, it is 1,320 and $158,400. These are illustrative planning scenarios, not forecasts. They omit churn, discounts, repeat visits, marketing cost and repairs. The first number to verify is 6,000: duplicate customers, moved households and inaccessible jobs can shrink it.

Capacity creates a second boundary. Suppose one technician can complete three inspections a day over 180 available days: 540 inspections. If only one technician is allocated, the 15% scenario cannot be delivered without more capacity. The serviceable obtainable volume for that staffing plan is at most 540, or $64,800 before costs. If the work takes longer than expected, the cap falls again. A market estimate without a delivery constraint can be numerically tidy and commercially useless.

Finally, run a break-even check using internal costs rather than web figures. If the illustrative contribution after travel and labour is $45 per inspection and the launch needs $18,000 of annual fixed effort, the plan needs 400 inspections to cover that effort. At 480 inspections the headroom is 80; at 540 capacity the upper headroom is 140. This turns an abstract market size into a pilot question: can the firm obtain at least 400 paid bookings without exceeding the assumed unit cost?

The model should show the calculation chain in full: customer records → eligible contacts → likely uptake → fulfilment cap → revenue → contribution. Each arrow needs either a source or an explicit assumption. A reader can then alter the 15% uptake or the 180-day capacity and see which conclusion changes. If AI writes the explanatory paragraph, compare every figure in it with the sheet before it leaves the practice.

Use prompts that demand provenance and missing values

Do not ask “How big is this market?” and copy the first round number. Ask the research assistant to find original candidate sources and state how each source defines the market. A copyable prompt is below; replace the bracketed terms with the actual brief. The answer is a research queue, not the estimate.

Find original, dated sources for [defined market] in [period].
For each candidate, give the exact metric, unit, population,
publication date, table or page, and direct source URL.
Separate primary sources from commentary. Do not infer missing
values. Flag a number if its scope differs from my definition.

An illustrative output might put “12 million households” next to a source that actually counts all households, while the brief asks for recently fitted kitchens. The repair is not to ask AI for a more persuasive citation. Mark the row “scope mismatch”, retain it only as an upper bound if useful, and seek a direct installed-base measure. Five minutes spent on that check prevents a bad client slide later.

For extraction, provide a short passage and ask for a narrow table: “Copy the stated year, count, denominator and caveat. Use null for missing data. Do not calculate.” In a sample run (illustrative), the model turned “up to 2.4 million units” into “2.4 million annual sales”. The words “up to” and “units” matter; the latter might be cumulative installed stock. Compare the table back to the passage, correct it and preserve the source wording in the ledger.

For a client-facing note, instruct the model to write a paragraph that distinguishes verified facts, internal records and model assumptions. If it produces “the market will generate $108,000”, revise to “at 15% assumed uptake, the serviceable customer base implies $108,000 of annual gross inspection revenue before capacity and costs”. The cautious sentence is more useful because it tells the client what to test. The same discipline applies when using AI to research a price.

Run six checks before accepting any market number

1. Source check: can you open the original source and point to the exact sentence, row or chart?
2. Date check: is the measurement period current enough for the decision, and is a later edition available?
3. Definition check: are the population, geography of service, product scope and value-chain level aligned with the brief?
4. Unit check: do counts, currency, percentages and time periods match throughout the formula?
5. Arithmetic check: can another person reproduce the result from the displayed inputs, including rounding?
6. Sensitivity check: which one or two uncertain inputs can reverse the decision?

A small numerical mistake illustrates the unit check. Suppose a source reports 250,000 annual purchases and a model assigns a $40 monthly subscription to each. Multiplying 250,000 × $40 gives a monthly-equivalent figure only if all purchases are active subscriptions for the entire month. Multiplying it by 12 to claim $120 million annually assumes every purchase remains active for a full year. If the source counts transactions rather than retained subscribers, the model overstates the result. This is an illustrative error, and the correction is to obtain active customer months or state a retention assumption.

Another illustrative failure is double counting channels. A craft brewery supplier might see 1,200 independent outlets in a directory and 400 wholesale accounts in its CRM. Adding them to 1,600 assumes no outlet appears in both lists. Match names and addresses first. If 250 overlap, the combined unique count is 1,350 before applying any eligibility rule. Ask AI to propose matching candidates if the records are messy, but have a person confirm ambiguous matches; a chain with two similar trading names can still be two buyers.

A third failure is mixing prices. A published retail price of $80 may include distributor and retailer margins while a supplier would realise $46. Using the retail price with supplier units inflates supplier revenue. In an illustrative 10,000-unit scenario, the difference is $800,000 versus $460,000. Record which price belongs to which value-chain step, then use the number relevant to the client's revenue. If margins are unknown, show both but label them.

Use a two-person review for consequential estimates. One person reproduces the sheet from source rows without looking at the final answer; the other challenges the market boundary and the least secure assumption. If both reviewers use the same AI summary as evidence, there is no independent check. A second AI answer also does not create an independent source. The review must reach the original material or the firm's own records.

Turn uncertainty into a short research plan

Desk research rarely resolves buyer willingness, implementation friction or a local firm's real share. Rank unknowns by their power to change the recommendation. In the kitchen example, eligibility and uptake dominate; another decimal place on the advertised price would not help much. A small outreach pilot can test uptake: offer the inspection to 100 eligible past customers, record delivered messages, replies, paid bookings and reasons for declining. If 9 book, that is a signal from that sample, not proof that the entire base will book at 9%.

For an illustrative care agency considering a training product, the problem might be whether 300 potential employers each buy one course or whether a few central buyers purchase for many staff. The model should use accounts and seats as separate units. A desk search can locate organisation counts; only interviews or a pilot can validate procurement behaviour. Build those questions into the research plan instead of adding AI-generated certainty.

For an illustrative childcare nursery expansion, a headline count of children may omit available places, opening hours and parents' willingness to travel. A better worksheet distinguishes eligible families, competing capacity and the nursery's own permitted places. These are not interchangeable. The consultant can use AI to compile candidate documents and compare terminology, but should not infer enrolments from a population count alone.

Make the pilot's denominator explicit. If 100 invitations produce 9 bookings but 20 messages bounce, the result is 9% of selected records or 11.25% of delivered messages. Neither number is automatically representative: the sample may favour recent customers or those with email addresses. Report both denominators and the selection rule. A second wave can deliberately include older customers to see whether the apparent demand survives a different cohort. The AI assistant can format that comparison, but it cannot decide whether the sample bias is acceptable for the investment at stake.

Keep a small rejection log as well as the evidence ledger. If you dismiss a large market report because it includes product categories outside the brief, note its title and the exact mismatch. Otherwise a reviewer may find the same prominent figure, assume it was overlooked, and reopen the argument late in the project. A three-line rejection note can save more time than another round of searching: “Source counted all maintenance contracts; our brief concerns annual inspections only; no reliable split published.” This also guards against a language model reintroducing the discarded figure in a later draft.

If the client asks for a single headline, choose a rounded range that reflects uncertainty. In the kitchen illustration, presenting “$57,600–$158,400” can look more precise than the assumptions justify. “Roughly $58,000–$158,000 of gross annual inspection revenue across the modelled uptake cases, before capacity” is clearer. Better still, put the capacity-limited upper case alongside it and explain why staffing limits delivery. Preserve the exact spreadsheet values for checking, while the narrative makes the scope and uncertainty readable.

Finish with a one-page evidence note: “Defined market”, “best supported range”, “inputs by source”, “assumptions”, “capacity limit”, “decision threshold” and “next test”. Attach the calculation sheet and ledger. The client should be able to ask “What if uptake is half that?” and get a new answer in seconds. That is a more valuable deliverable than a single impressive figure whose trail disappears when the AI chat closes.

Further reads

• How to Do Competitor Research With AI in One Afternoon — Collect comparable competitor evidence without mistaking claims for demand.
• How to Test a New Service Idea With AI Before You Launch It — Use a small market estimate to frame a real demand test.
• The AI Tool Stack a One-Person Consultancy Actually Needs — Choose a lean tool set for a solo consulting practice.
• How to Write a One-Page AI Strategy for Your Business — Turn a research finding into an operating decision.
• How Consultants Use AI for Client Research Before Discovery Calls — Prepare for interviews that can challenge desk assumptions.
• Best AI for Consulting Research: Market Scans With Source Checks — A source-check routine for consultants: prompt chain, source log, three verification tiers and how ChatGPT, Claude and Gemini research modes differ.
• Best AI for Consulting Work: From Interview Notes to Findings Deck — Transcribe, code themes, build an evidence table and draft the findings slides, with consent steps, prompts, sample output and a worked 12-interview example.
• ChatGPT Prompts for Strategy Consultants: Issue Trees and Hypotheses — Prompts for an MECE issue tree, testable hypotheses and interview guides before a workshop, with sample outputs and the checks a consultant must run.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Perplexity help centre, What is Pro Search?; OpenAI help centre, Deep research in ChatGPT; Google, Gemini Deep Research overview (all checked 4 November 2026).