AI Market Research for Consultants: Build a Checked Evidence File

Define the decision before choosing the research questions

Start with a decision statement that contains an action, a scope, and a deadline. A question such as whether maintenance software is growing is too broad to direct an engagement. A better question asks whether a vendor should fund a limited sales experiment for independent U.S. food-processing plants during the next planning cycle. The narrower statement identifies what the client can decide and which evidence may affect that choice.

Clarify the unit of analysis. A plant, company, establishment, and corporate parent are different counts. A vendor that sells to sites needs site-level economics; a vendor that contracts centrally needs account-level buying information. If a source counts establishments while your revenue model counts customer organizations, create a mapping assumption rather than treating the two as interchangeable.

Specify the product boundary. Maintenance scheduling, inventory management, enterprise asset management, and industrial analytics may overlap without representing the same buying decision. Write the client's proposed offer in plain language and list its included functions. Exclude adjacent products until you have a reason to include them. Broad market labels can make an opportunity appear larger by combining customer needs the client cannot serve.

Clarify the geographic and customer restrictions. The hypothetical client can support U.S. customers, works in English, and initially serves independently managed facilities. Those boundaries affect reachable demand, implementation requirements, and sales economics. A worldwide market report may provide context, but it cannot directly establish the client's initial addressable base.

Identify the alternatives to market entry. The client could improve retention in its current segment, partner with a specialist, run a smaller discovery effort, or postpone investment. Research should compare the proposed entry with a realistic alternative, because an attractive market can still be a poor use of this client's limited resources.

Agree on the minimum evidence needed for a decision. The client might require a defensible customer count, several documented buyer interviews, a competitor capability comparison, and a plausible route to a first sale. Those are local engagement criteria. They do not transform incomplete research into certainty, but they prevent a team from calling a collection of summaries a completed market assessment.

Construct a research map with explicit evidence classes

Separate research into market definition, customer population, buyer problem, competitive alternatives, economics, and entry constraints. Each class needs different sources. A public statistical table may help estimate the population. It will not establish whether plant managers find the problem urgent or have permission to buy the product.

For each research question, write the preferred source and a fallback. Original government tables, company disclosures, product documentation, customer interviews, and direct observations answer different questions. A secondary article can point you toward evidence, but copying its number without locating the original source leaves definitions and caveats unresolved.

Record the source date and the period described. An article published this year may quote a survey conducted several years earlier. A product page accessed today can establish what the vendor currently advertises, but it may not show what existed during the competitor's reported sales period. Research freshness follows the claim, not simply the page's visible date.

Distinguish four labels in the evidence file: observed fact, source estimate, consultant assumption, and unresolved question. An observed fact might be the number in a specified public table. A source estimate might be a published forecast. A consultant assumption might be the share of facilities with a qualifying maintenance problem. An unresolved question might concern purchasing authority. Keep those labels attached when the content moves into slides.

Track independent evidence, not source count alone. Five articles repeating one company's press release do not create five independent confirmations. A market estimate distributed through multiple websites can appear strongly corroborated while resting on the same proprietary model. Identify the underlying origin and avoid presenting repetition as triangulation.

Add a field for decision relevance. A source may be accurate yet unhelpful. A global spending forecast may belong in a context appendix, while an interview revealing that sites cannot buy without central approval changes the sales experiment. This field encourages the team to spend review time on claims that could alter the recommendation.

Build an evidence ledger that survives the move into a deck

Use one row per claim rather than one row per webpage. Include the claim identifier, exact research question, proposed statement, source URL, document title, publication date, observation period, locator, unit, definition, evidence class, and limitations. A long page can support several distinct claims with different caveats, so a page-level citation alone is insufficient.

Capture a short permitted excerpt or a precise locator alongside your own paraphrase. The excerpt helps reviewers see what the source actually says, while the paraphrase expresses its relevance to the client. Avoid saving an entire copyrighted report simply because a tool can retrieve it. Source access and reuse rights should be handled according to the engagement and the provider's terms.

Attach the ledger identifier to the draft slide. A slide note reading C-017 is useful when C-017 points to the original table and its definition. A list of links at the end of the deck makes review harder because the reader must guess which source supports which sentence. Where practical, include a concise citation near the claim as well.

Record data changes. If you filter a population, convert units, adjust for geography, or multiply by a price assumption, show each step. The ledger should reveal where the original source ends and your modeling begins. An AI-generated market-size paragraph often hides these boundaries by presenting the final result as though it were directly published.

Keep contradictions visible. Suppose one source counts 2,000 organizations while another counts 3,500 facilities. The figures may describe different units, not a disagreement. If they truly differ under the same definition, record both and investigate. Do not ask AI to average them unless averaging has a defensible analytical purpose.

Assign review status: not checked, checked against original, definition unresolved, calculation checked, approved for recommendation, or context only. This is an editorial workflow, not a truth score. A checked estimate remains an estimate. The status should tell the next analyst what work has been completed and what they must still examine.

Work a bottom-up sizing example without disguising assumptions

Consider the fictional vendor Fieldline. Its consultant constructs a synthetic planning population of 2,400 U.S. facilities that fit an initial segment definition. This count is invented for the exercise. It is not evidence about the food-processing industry. Assume that 35% have the qualifying workflow problem and that 60% of those fit Fieldline's deployment requirements. The resulting serviceable population is 504 facilities: 2,400 multiplied by 0.35 multiplied by 0.60.

At an assumed annual subscription of $12,000 per facility, that synthetic serviceable revenue pool is $6,048,000. The calculation is easy to reproduce, but its interpretation matters. It is a modeled pool under specified assumptions, not a revenue forecast, contracted demand, or a claim that the vendor can capture the full amount.

Now introduce sales reach. Suppose the client can identify and contact 120 qualifying facilities during a test period. If an assumed 20% agree to discovery, that gives 24 conversations. If an assumed 25% of those conversations become paid pilots, that gives six pilots. At an assumed $3,000 pilot fee, test revenue would be $18,000. Every percentage remains an assumption until the experiment supplies observations.

The distinction between a market pool and a sales funnel prevents a common overstatement. Saying the vendor has a $6 million near-term opportunity would collapse all reachable, qualified, interested, and won stages. The decision may instead concern whether six potential pilots would generate enough learning to justify the test's cost.

Build sensitivity around the uncertain filters. If problem prevalence is 20% and deployment fit is 50%, the same population yields 240 serviceable facilities and a $2.88 million modeled pool. If prevalence is 50% and fit is 70%, it yields 840 facilities and a $10.08 million pool. These are constructed scenarios, not confidence intervals or statistically estimated probabilities.

Show which evidence could narrow the assumptions. Interviews and deployment reviews may test problem prevalence and fit. A sales experiment may test response and conversion. A different source may revise the population definition. This makes the research actionable: the next study targets the assumption most capable of changing the decision rather than adding another impressive-looking market forecast.

Check competitor claims at the level of customer choice

Build competitor profiles around the client's proposed buying situation. Include the alternative used today, a specialist competitor, a broader platform, and the option of doing nothing. Customers do not always compare products within the category a market report uses. A spreadsheet and an internal maintenance coordinator may be the actual alternative to a new subscription.

Distinguish advertised features from verified delivery. A product page can support the statement that a vendor advertises mobile work orders. It does not prove usability, adoption, implementation speed, or business impact. Record whether a capability comes from documentation, a demonstration, a customer reference, or an unverified sales claim. Keep those distinctions in the recommendation.

Verify pricing conditions. Public starting prices may exclude implementation, minimum seats, integrations, support levels, or contract terms. If the information is unavailable, mark it unavailable and explain which comparison remains incomplete. AI should not supply an industry-normal implementation fee and present it as the competitor's actual charge.

Compare workflows rather than feature counts. For Fieldline, the relevant workflow might be opening a maintenance request, assigning work, recording completion, and exporting an audit trail. A platform with many unrelated features can still be weaker for that workflow, while a simple product may lack a required permission control. Define the test before interpreting the feature list.

Keep a claim matrix with source locators and observation dates. If a competitor changes its documentation during the engagement, update the affected claim and record the change. A screenshot can preserve what you observed, subject to appropriate use, but it should not substitute for a precise description of what was tested.

Avoid claiming a competitive advantage from missing evidence. If you cannot find whether a rival supports offline use, write that the public material reviewed did not resolve offline capability. Saying the rival lacks offline use goes further. The distinction protects the client from building a positioning message around a gap that disappears during the first sales conversation.

Use interviews to investigate hypotheses and preserve disagreement

Interview research should begin with a hypothesis register, not a script designed to confirm the proposed offer. Fieldline may hypothesize that maintenance supervisors lose time reconciling paper records and would value a mobile system. Ask for recent examples, the current process, consequences, attempted fixes, and the buying route. A general question about whether AI sounds useful is unlikely to reveal the operational decision.

Separate what the participant said from your interpretation. The note might state that a supervisor spent two afternoons preparing a record export last month. Your interpretation might be that reporting creates recurring effort. A further hypothesis might be that Fieldline's product could reduce that effort. These are different evidence levels and should appear as separate fields.

Use permitted recording and transcription practices defined by the client and engagement. Approval to interview someone does not automatically authorize uploading a recording into an unapproved service. If the source is confidential, preserve access restrictions in the research workflow. Where recording is unsuitable, structured notes with participant review may be more appropriate.

Ask AI to propose coding labels or organize approved notes, then verify the excerpts that support each theme. A theme can become too broad when the system compresses several distinct complaints into one category. Reporting burden caused by missing data differs from reporting burden caused by an awkward export. The product implications are different.

Preserve negative and contradictory cases. If one site has no urgent problem because its existing system works well, that matters. If central procurement controls all purchases, site-level enthusiasm may not translate into sales. Count those cases in the analysis rather than using only quotes that strengthen the pitch.

Describe the sample honestly. Several interviews can reveal mechanisms, objections, and language. They do not establish representative national prevalence without an appropriate sampling design. Avoid turning eight exploratory interviews into a statement that most U.S. plants want the product. Their proper role may be to improve the next experiment and revise the qualification criteria.

Diagnose false precision, retrieval gaps, and recommendation drift

False precision appears when a modeled opportunity is reported to the nearest dollar even though its assumptions are uncertain. The Fieldline calculation can be shown exactly in the working model, while the executive narrative describes a roughly $6 million synthetic serviceable pool under the base assumptions. The exact arithmetic should not imply exact knowledge of demand.

Retrieval gaps appear when the AI system returns only sources that are easy to access or resemble the query wording. An inaccessible original report, a source behind registration, or a differently named category may be omitted. Keep a research log of searches, exclusions, and unresolved sources. A blank result is not proof that evidence does not exist.

Source substitution occurs when a system cites a page that mentions the topic without supporting the claimed number. Open the cited source and locate the actual statement. If the link is a press-release summary of another document, follow the trail. If the original cannot be examined, label the limitation and decide whether the claim can carry the recommendation.

Recommendation drift occurs when the draft strengthens its conclusion during editing. A cautious evidence table may become a slide saying the client should enter immediately. Compare the final recommendation with the evidence and the agreed decision criteria. Ask which assumptions changed, who approved the change, and what new information supports the stronger language.

Include a challenge case in review: an impressive but irrelevant market forecast, a competitor capability unsupported by its cited page, an interview excerpt that contradicts the leading theme, and a unit mismatch. A workflow that fails these checks should be narrowed or repaired before the client receives the deck. Passing them establishes performance on that local test only.

The consultant task experiment found uneven AI effects across selected tasks. Its specific firm and older tool do not benchmark this research process. The practical implication here is to evaluate your own retrieval, interpretation, and recommendation steps rather than importing a general productivity claim.

Set acceptance criteria around the decision-bearing claims

Agree on acceptance criteria before reviewing the final presentation. Every numerical claim used to recommend investment should have an original source or a clearly labeled assumption, an appropriate definition, and a reproducible calculation. Every competitor claim used in positioning should state the evidence type and unresolved conditions. These criteria make the deliverable inspectable.

Require reconciliation between the narrative, tables, and model. If the deck says 504 facilities while the spreadsheet's updated scenario says 420, the package is not ready. Check totals, units, dates, currency, scenario names, and filters across artifacts. AI can help identify differences, but the engagement owner remains responsible for resolving them.

Ask a reviewer to reproduce the central sizing without the author explaining it orally. They should locate the source population, follow each filter, and identify the price assumption. Failure often indicates documentation gaps rather than a difficult calculation. Repair those gaps before using the model to justify spending.

Require a missing-evidence register. For Fieldline, unresolved buying authority, integration requirements, and pilot conversion may deserve named owners and a next research action. A polished final report can openly contain unknowns when it explains how they limit the decision. Concealing them makes the recommendation harder to trust.

Define severity. A broken appendix link might require correction without changing the recommendation. A source that counts companies where the model assumes facilities can undermine the sizing. A fabricated customer quote can invalidate the research artifact. Severity should reflect decision impact and professional obligations, not only the number of edits.

Have the client sponsor review whether the answer fits the decision they actually face. An evidence-perfect national market report may miss a local channel constraint. Acceptance includes analytical integrity and usefulness for the scoped question. Avoid declaring success solely because the report contains a predetermined number of slides or citations.

Turn the findings into a testable sales experiment

The research file should help the client choose what to learn next. For Fieldline, the first test could investigate whether an identified supervisor can bring the purchasing decision to the person who controls the budget. That question deserves its own observation because user enthusiasm and buying authority are different stages. Record the current approval route for each contacted facility and the reason a conversation stops.

Choose a consistent qualification definition before outreach. A qualifying facility might have the defined maintenance workflow, a responsible owner willing to discuss recent examples, a feasible deployment route, and an identifiable purchasing process. These are constructed criteria for the exercise. If the team changes them during the test, preserve the original definition and report the change instead of comparing incompatible conversion rates.

Use a simple contact log with facility identifier, segment, contact role, qualification disposition, documented objection, next action, and outcome date. Keep personal details within the approved client system. The research summary can use aggregated or appropriately de-identified observations where suitable. AI may help classify permitted notes, but the consultant should inspect the records behind any claimed pattern.

Distinguish a nonresponse from a rejection. A facility that never answers does not establish that the problem lacks importance. A respondent who describes a problem but cannot authorize a purchase reveals a channel issue. A qualified buyer who rejects the price reveals another issue. Compressing these outcomes into one failed-conversion category loses the information needed to revise the offer.

Predefine the decision at the end of the test. The sponsor might continue if observed discovery conversations reveal a consistent problem and a viable buying route, revise if the problem exists but integration blocks adoption, or stop if the current offer cannot meet the defined workflow. Specify who will make the decision and which evidence they will examine. Avoid using an arbitrary success score that conceals the underlying observations.

Finally, connect the experiment back to the model. Replace an assumption only with evidence that measures the same thing. A response rate from a warm referral list should not become the assumed response rate for cold national outreach. A paid pilot is not a converted annual subscription. Preserve those distinctions so the next planning model becomes better informed without claiming more certainty than the test provides.

Measure the whole research process and communicate the limits

Record source discovery, original-source retrieval, ledger entry, calculation, interview coding, drafting, review, and revision separately. A tool that reduces summary writing may increase source checking. Compare total effort on a consistent task, and note whether the assisted process changes the breadth or quality of the research.

Use a synthetic pilot comparison to understand the measurement. Suppose an unaided evidence file requires 90 minutes of searching, 80 minutes of source checking, 60 minutes of modeling, and 70 minutes of drafting and review, totaling 300 minutes. The assisted version might use 50, 100, 60, and 55 minutes, totaling 265. The observed difference in this illustration is 35 minutes, not the 40-minute search saving alone.

Track unsupported claims discovered, unresolved sources, calculation defects, and recommendation changes after review. A report that takes less time but requires withdrawal of its main recommendation has not demonstrated useful efficiency. Define the quality floor before interpreting elapsed time.

In the client handoff, explain how AI contributed and what the team checked. For example, AI helped organize public-source notes and draft competitor summaries; analysts opened the original pages, checked the market definitions, reproduced the sizing, and reviewed the recommendation. The explanation gives the client a concrete account of the process without pretending that AI use itself establishes quality.

Deliver the evidence ledger and model with the deck according to the engagement's access and confidentiality terms. Keep internal working material, sensitive interview details, and client-ready evidence distinct. Explain which files the client may update and which assumptions require new research rather than a simple refresh.

State the recommendation conditionally where appropriate. Fieldline could fund a bounded discovery and sales test if the client accepts the cost and can resolve integration requirements. That is different from endorsing a full market launch. The recommendation should include the stopping condition and the next evidence that would justify expansion.

Build research judgment through a four-artifact learning plan

An experienced consultant can practice this method with a market they already understand. Choose a narrow customer decision and create a source ledger before using AI to draft anything. The first artifact is a definition sheet that distinguishes customer, establishment, site, product category, geography, and time period. Review it with someone who knows the market.

The second artifact is a sizing model with every assumption labeled. Build a base case and two scenarios without assigning unsupported probabilities. Change one filter at a time and observe how the decision changes. Explain which assumption contributes the most uncertainty and what evidence could narrow it.

The third artifact is a competitor workflow comparison. Use documented capabilities, clearly marked unanswered questions, and a test the client can understand. Compare the same workflow across alternatives. This teaches more than collecting every feature a vendor advertises.

The fourth artifact is a recommendation memo that includes a contradictory finding. Ask a colleague to challenge the memo using only your evidence file. If they can identify a missing premise, revise the recommendation or obtain the missing evidence. The goal is to make the argument stronger through inspection, not to defend the first draft.

Use AI after each artifact has a clear purpose. Compare whether it helps discover questions, organize notes, or express a checked conclusion. Preserve examples where its output was useful and where you rejected it. A portfolio of such decisions demonstrates professional judgment more credibly than a collection of generic AI certificates.

Choose one research habit to carry into the next engagement. You might require a denominator beside every percentage or attach the original-source locator to each recommendation-bearing slide. Put the habit into your review checklist and observe whether it changes the team's work. That small operating change is concrete enough to test and useful even if the preferred research tool changes.

Compare your options

Build a market-claim evidence ledger

Choose one client decision and a permitted research example. Connect the recommendation to original evidence, explicit sizing assumptions, conflicting findings, and the test that could change your advice.

A reflection tool for planning your next step, rather than a predictive score.

1. Define the client action before collecting market summaries.
2. Build the evidence ledger from original sources.
3. Label each sizing input as observed, estimated, assumed, or unresolved.
4. Challenge competitor and interview conclusions with conflicting evidence.
5. Review the recommendation against the agreed decision criteria.

Fieldline's completed synthetic worksheet begins with a limited market-entry experiment, not a full launch. Its planning population is 2,400 facilities; qualifying-problem and deployment-fit assumptions produce 504 serviceable facilities. At an assumed $12,000 annual price, the modeled pool is $6,048,000. A reachable test list of 120 facilities, assumed discovery conversion of 20%, and assumed pilot conversion of 25% yield six hypothetical pilots.

The evidence register identifies buying authority and integration requirements as unresolved. The recommendation is to fund a bounded discovery test only after the sponsor approves its cost and access conditions. Expansion depends on observed qualification, conversion, and delivery effort. The calculation teaches a planning method; it does not estimate an actual market or promise sales.

Frequently asked questions

Search research and audience discussions informed these questions. Some wording has been adapted to the task covered by this guide.

Sources and evidence

These guides were drafted with AI assistance and checked against the linked sources. Examples labeled hypothetical illustrate a decision; they are not participant stories or measured results.

Studies answer specific questions about particular people, tasks, and tools. Their dates and limits matter. A result about task assistance is not a forecast for an entire profession.

1. Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality

Where the audience questions came from

These linked discussions and comments supplied questions and wording. They do not establish the truth of the replies, verified professional identities, or how common a concern is.

1. Public discussion supplying a question: Can AI replace the human judgment in market research?
2. Public discussion supplying a question: What does validating an AI research response involve?
3. Public discussion supplying a question: How should I assess an unfamiliar market-research report provider?
4. Public discussion supplying a question: How should I structure a market assessment of a peer company?
5. Public discussion supplying a question: Should a consultant use a large expert network or a smaller provider?
6. Public discussion supplying a question: How should I compare the cost of expert-research providers?
7. Public discussion supplying a question: How do I find data that can test a specific market hypothesis?
8. Public discussion supplying a question: Can AI help draft a market-assessment report?
9. Public discussion supplying a question: Should an early diagnostic change the research engagement scope?
10. Public discussion supplying a question: When does a research follow-up become new scope?
11. Public discussion supplying a question: What AI training is useful for a market-research consultant?
12. Public discussion supplying a question: How can I obtain useful industry evidence without buying every report?