Map Clients, Referrers, and Buying Questions

3.1 Identify Prospective Clients, Referrers, Internal Sponsors, Evaluators, and Procurement Teams 

Start with roles in a particular buying decision. “Chief financial officer” is a job title; deciding whether to outsource the monthly close is a task. The task tells you which questions to investigate and which facts an answer should establish.

Use the following roles to identify the people whose questions deserve attention. These are research categories, not a requirement to find five separate individuals for every engagement.

An owner may perform several roles. An HR director may identify the need, investigate lawyers, and recommend a firm to the CEO. A trusted accountant may introduce a consultant and then have no further involvement. Include people who will use or implement the advice when their requirements affect selection, even if they do not approve the expenditure.

For each relevant role, record the immediate decision, the person’s starting knowledge, the constraints that could rule a provider out, and the evidence needed to proceed. Avoid adding personal characteristics that do not change the question or the service requirement.

Consulting example

A buyer considering an acquisition wants commercial due diligence on a distributor. The investment lead wants to understand demand and competitive position. An operating executive wants to know whether the consultant understands distribution economics. Deal counsel, acting as a referrer, wants confidence that an introduction matches the requested scope. Each can ask a different question about the same consulting practice.

Interviewing only the investment lead would leave the operating executive’s evaluation questions and counsel’s referral questions unexplored. Interviewing only the consultant’s partners would describe the firm’s view of those questions. Mark that difference in the research plan.

Recruit people with relevant experience of selecting or referring to the service, including people who chose another provider where access is appropriate. GOV.UK’s user research guidance recommends actual or likely users and warns that recruitment choices introduce bias. Apply that principle by varying the perspectives you seek, rather than relying entirely on enthusiastic existing clients. 

3.2 Gather Real Questions From Interviews, Proposals, Intake Discussions, and Business Development 

Begin with information the firm already has permission to use: intake notes, proposal clarification requests, business development records, referral introductions, and questions raised during pitches. Include won, lost, declined, and unresolved opportunities. Look for the wording used before the client adopted your service terminology.

Extract the underlying question and the circumstances that made it matter. “Can you take this off my plate?” might concern transaction processing, oversight of an internal team, or responsibility for the entire finance function. Preserve that ambiguity until evidence clarifies it.

Conduct interviews around a recent decision

Ask someone to reconstruct a specific selection or referral. GOV.UK’s interview guidance favors open, neutral questions and concrete experiences over general accounts of how a process ought to work. The following discussion guide applies that approach to choosing professional services. 

1. What was happening when you first considered outside help, and what did you call the problem?
2. Who became involved, and what did each person need to understand or approve?
3. Which people, websites, tools, or other sources did you consult? What did you ask first?
4. Which firms or professionals were suggested? What did you check about them next?
5. What made you continue with a provider, rule one out, or postpone the decision?
6. What could you not establish from the available information, and what did you need to ask the provider directly?

Then ask specifically whether an AI assistant played any part. If it did, ask which tool, what the person was trying to accomplish, what they remember asking, and what they did with the answer. A voluntary, appropriately redacted excerpt can improve accuracy. A recollection should remain labeled as a recollection.

Keep hypothetical discussions separate. “What might you ask an assistant next time?” can generate useful test ideas, but it does not document past behavior. Similarly, a question raised in a sales call is evidence of a client’s information need, not proof that the client entered it into ChatGPT.

Preserve the origin of each question

Record two things independently: where the question came from and how its wording has changed. The origin might be an observed AI prompt, a client question outside AI, a referrer’s recollection, or a team hypothesis. The test wording might be verbatim, reconstructed, or adapted. One label cannot express both distinctions.

For example, an intake note could record “Will you run payroll too?” The team might group it under “boundaries of outsourced accounting” and write a fuller prompt about monthly close and payroll responsibilities. Retain the connection between those three records. GOV.UK’s analysis guidance similarly separates what was seen or heard from interpretation and subsequent action. 

Keep a source reference and date where appropriate, with sensitive material restricted to the people authorized to use it. The working prompt library should contain only information approved for its intended use. Remove unnecessary identities, transaction details, allegations, and personal information; changing a name alone may leave a situation identifiable. Chapter 4 establishes the firm’s formal rules for this work.

Review the emerging themes with both a practitioner and someone who handles inquiries. Ask what is missing, which distinctions were lost in paraphrasing, and whether staff assumptions are being presented as client evidence. AI can help suggest groupings from material approved for that use, but proposed questions remain hypotheses until supported.

When noting repetition, distinguish separate buying situations from several exchanges about one engagement. Six emails from one prospect can show that an issue required sustained clarification; they do not establish that six independent buyers raised it.

Finally, describe the limits of the research honestly. Qualitative interviews can reveal a range of needs and reasons; a small, deliberately selected sample does not establish market percentages. HM Treasury’s evaluation guidance makes this distinction explicit. “Three interviewees raised this issue” is a description of your evidence, not an estimate of how often all buyers ask it. 

3.3 Map Problem Definition, Referral Validation, Shortlisting, Due Diligence, and Engagement Selection

Assign each question to the decision it helps someone make. The five categories below provide a workable starting structure. Use them as categories rather than requiring every buyer to follow a fixed sequence.

A referral can arrive before the client has named the problem clearly. A discussion about fees can reveal a scope mismatch and reopen the shortlist. Different participants can also be working on different decisions at the same time. Let the map reflect that.

For each question family, write one sentence describing what an accurate, helpful answer would accomplish. At problem definition, that may be a clear explanation of service options without naming any firm. At shortlisting, it may be a set of plausible providers to investigate. During due diligence, acknowledging that public evidence is insufficient can be the most useful result.

This step prevents a common mistake: assuming that the firm should appear in every answer. An educational question does not automatically call for a recommendation. Nor should a firm receive credit for a flattering answer that attributes unsupported qualifications or services to it.

Law firm example

An HR director receives the name of an employment lawyer from an accountant. The initial question concerns whether the lawyer advises employers in the relevant jurisdiction. A later question concerns experience conducting workplace investigations. Before appointment, the organization needs direct answers about the proposed team, conflicts, availability, and engagement terms.

These questions should not collapse into “best employment lawyers.” They involve a named professional, a particular client perspective, a specific service, and facts requiring different kinds of verification. A public biography may support a statement about practice experience; it cannot, by itself, establish that the lawyer has cleared conflicts or accepted the matter.

Add an information requirement beside each question: what could responsibly be established from public sources, and what needs direct contact? This identifies material the firm may eventually need to clarify. It does not mean every question needs its own web page or that confidential engagement information should become public.

3.4 Distinguish General Discovery Questions From Questions Naming a Firm or Professional

Record whether the question already supplies a provider’s name. This changes what the test can tell you.

• General discovery: “Which firms provide outsourced accounting support for a manufacturer like ours?” The assistant has not been given your firm’s name.
• Firm validation: “Does [Firm] provide the monthly accounting support we need?” The name is supplied; the question concerns representation and suitability.
• Professional validation: “What relevant experience does [Professional] at [Firm] have?” The test concerns an individual and the relationship between that person’s experience and the current firm.
• Named comparison: “How do [Firm A] and [Firm B] compare for this requirement?” The alternatives are supplied; the task is evaluating the available evidence.

Keep these classes separate when interpreting results. A named-firm answer cannot establish that the assistant would have discovered the firm independently. A professional’s relevant experience also needs accurate attribution; work at a previous employer should not silently become an engagement delivered by the current practice.

Preserve realistic referral context. “Our accountant suggested [Firm]” supplies a reason the user is investigating a name. It does not establish that the referral is correct. Avoid adding the firm’s preferred conclusions, such as calling it the leading specialist, unless that wording is part of a real question you are explicitly studying.

Record supplied material separately from named entities. “Assess [Firm]” differs from “Read this page and assess [Firm].” Anthropic documents Claude’s ability to retrieve a supplied URL. The methodological implication is straightforward: such a test can examine interpretation of supplied material, but it cannot establish whether the assistant would have found that material without the link. 

Also preserve meaningful conversation context. OpenAI documents that ChatGPT can rewrite a request into targeted searches and use location information and, when enabled, relevant saved memories. A bare service phrase and a question embedded in a detailed buying conversation are therefore different test conditions. This documentation concerns ChatGPT; do not assume identical behavior across providers. 

If a question is a follow-up, retain enough of the preceding exchange to identify its meaning and label it as such. Save independent first questions separately. Chapter 6 will specify how to control and record these conditions in a monitoring program.

For example, “Which firms can help with that?” depends on the earlier description of the problem and the assistant’s intervening answer. If the earlier answer already introduced your firm, the follow-up is no longer an independent discovery test. Preserve the relevant exchange, including the answer that supplied the name, instead of saving only the last sentence.

3.5 Segment Questions by Practice, Industry, Geography, Urgency, and Client Circumstances

Add context when it could change the service required, the plausible providers, or the evidence needed to assess fit. Use the following dimensions as prompts for judgment, not as fields that must all appear in every question.

• Practice and service: Specify the work being considered. Commercial diligence, financial diligence, and operational diligence should not be treated as interchangeable requests.
• Industry: Identify sector conditions that affect the expertise needed. Keep a sector label only when it helps explain the work or the client’s selection criteria.
• Geography: Distinguish the client’s location, the location of the work, and any relevant jurisdiction. A nearby office alone does not establish suitability.
• Urgency: Capture a real deadline or timing constraint when it affects the engagement. Public information may leave current availability unresolved.
• Client circumstances: Include material facts such as organization size, existing staff, systems, project stage, budget constraints, or the side of a dispute the person occupies.

Separate requirements that could rule a firm out from preferences that might favor one suitable provider over another. Then identify which requirements are established and which the buyer still needs to clarify. Do not invent detailed client circumstances simply to make your firm seem like an unusually close match.

Accounting example

“We need accounting help” could lead to very different questions. An owner with late monthly accounts may need bookkeeping capacity, a better close process, or oversight from an experienced controller. A business seeking an independent financial statement audit is asking for a different service. If the chosen practice does not offer audits, that distinction belongs in the question map.

Preserve the buyer’s early uncertainty. One prompt can ask what kind of help is appropriate; another can ask which providers offer the clarified service. Writing every prompt in the firm’s precise technical terminology would leave the first situation unexamined.

Engineering consulting example

A property owner investigating a structural assessment may care about the building type, project location, required deliverable, and who will take professional responsibility. A prompt mentioning only the closest engineering office would omit those distinctions. Record the actual requirements without assuming that proximity proves competence or authorization.

Where clients use more than one language, investigate the terms they actually use in each. Have someone familiar with both the language and the service review proposed wording. Label unvalidated translations as hypotheses rather than treating them as evidence of local client questions.

Avoid generating every possible combination of service, industry, location, and client size. Create a separate question family when a difference changes the decision or the evidence needed. Treat simple wording variations as variants within a family.

Prioritize the families using three judgments: relevance to the agreed program, consequence for the client’s decision, and strength of the evidence that the question matters. Use plain labels such as initial, next, and later, with a short reason. A rarely encountered but consequential service misunderstanding can deserve immediate attention.

Keep research confidence distinct from priority. Mark unvalidated questions as needing research, whatever their priority. An important hypothesis may warrant investigation; a well-documented question may fall outside the current scope. Do not lower a question’s priority merely because the firm lacks a persuasive public answer. That gap may be precisely what the program needs to understand.

3.6 Build Representative Test Prompts, Including Situations in Which the Firm Is Unsuitable

Here, representative means covering the important roles, decisions, and circumstances identified by your research. It does not mean that the library statistically represents all questions asked in AI tools. The initial library is a deliberate selection for learning.

Choose a set small enough for the operator and professional reviewer to examine carefully. There is no mandatory prompt count in this playbook. Begin with the most consequential question families and add prompts where they expose a distinct information need, referral situation, or suitability boundary.

Write in language a prospective client or referrer could plausibly use. Preserve short or incomplete questions when they reflect the evidence; retain more detailed scenarios when buyers actually supply that context. Do not improve every question into an elaborate instruction that quietly supplies the firm’s preferred positioning.

The nine prompts below illustrate a starting structure across three professions. They are hypothetical templates, not a proposed universal monitoring panel. Use the examples relevant to your scope and replace bracketed placeholders with verified names and locations before testing. The identifiers can link each prompt to its question family.

Consulting example: choosing a due diligence consultant

C1 — General discovery and shortlisting. “We are considering acquiring a US industrial distributor. Which consulting firms could assess customer concentration, demand, and competitive position as part of commercial due diligence?”

C2 — Referral validation. “Our deal counsel suggested [Consulting Firm] for commercial due diligence on an industrial distributor. What publicly available evidence is relevant to evaluating that referral, and what should we verify directly?”

C3 — Service boundary. “We need quality-of-earnings analysis and tax structuring for an acquisition. Is [Consulting Firm] an appropriate provider to investigate?”

For C3, assume the firm has verified internally that it offers commercial diligence and does not provide the two requested services. A useful answer should avoid assigning those services to the firm. If public evidence does not resolve its scope, an answer that identifies the uncertainty and recommends verification may be appropriate.

Accounting example: evaluating outsourced support

A1 — Problem definition. “Our monthly accounts are late, and the owner still supervises the bookkeeping team. What type of outsourced accounting support should we investigate?”

A2 — Named comparison. “We are considering [Accounting Firm A] and [Accounting Firm B] for monthly close and ongoing finance oversight at a manufacturing business. What can we establish about their services, and what should we ask each firm?”

A3 — Service boundary. “We need an independent audit of our company’s financial statements. Is [Accounting Firm] an appropriate provider to investigate?”

For A3, assume the selected firm has confirmed that it does not offer audit services. For A1, a good answer may explain options and ask clarifying questions without naming providers. For A2, the answer should preserve differences in scope and identify facts that require confirmation.

Law firm example: finding employment counsel

L1 — General discovery and shortlisting. “Which law firms should an employer investigate for workplace investigation advice in [Jurisdiction]? What experience would be relevant to assessing them?”

L2 — Professional validation. “Our HR director was referred to [Professional] at [Law Firm] for employment advice to employers in [Jurisdiction]. What can we verify about the relevant practice and experience?”

L3 — Client boundary. “I am an employee seeking representation in a dismissal dispute in [Jurisdiction]. Is [Law Firm] a suitable place to seek help?”

For L3, assume the practice has confirmed that it represents employers and does not accept employee-side matters. This is a service-boundary test. It does not determine whether any particular matter would pass the firm’s intake and conflicts procedures.

Define what the answer needs to get right

Keep the assessment criteria in the research record, separate from the prompt submitted to the assistant. Record the client’s decision, the relevant verified facts, material limitations, and matters that remain unknown. Use those criteria to evaluate an answer without prescribing a flattering result or a particular list position.

Distinguish a verified exclusion from missing public evidence. A sparse biography does not prove lack of experience. An omitted service description does not prove the firm never provides that service. In the examples above, the exclusions are explicit hypothetical facts; in your program, verify them with the responsible practitioner.

Include situations with insufficient information as well as clear matches and clear exclusions. A request that omits the relevant jurisdiction or the required service may call for clarification. Do not automatically treat a clarifying question, a qualified answer, or a correct exclusion as a failure.

Finally, review the prompts before running them. Remove accidental clues that reveal the desired conclusion. Keep any deliberately leading or supplied-source questions labeled for the specific issue they investigate. Do not keep rewriting discovery prompts until they produce an attractive answer and then present the revised set as evidence of general visibility.

Practical Output: Create the Question Map and Initial Prompt Library

Use one working question map with a row for each question family. Keep the full prompt wording in linked records so the map remains readable. The following extract illustrates the structure; all entries are hypothetical, and each firm would select only the rows relevant to its program. In the accounting illustration, the initial program emphasizes service clarity; named comparisons are queued for a later pass.

For each family, also record its material circumstances, source references and dates, whether the evidence concerns actual AI use, the basis for its priority, and the person responsible for keeping it current. Preserve distinct subquestions if grouping them would hide different service requirements.

Each prompt record should contain the following:

• Identity: prompt ID, question-family ID, exact wording, and version date.
• Origin and context: evidence source; whether wording is verbatim, reconstructed, or adapted; audience; decision; named entities; and any supplied material or preceding conversation.
• Assessment: what a useful answer should establish, verified service boundaries, supporting facts, and matters requiring direct confirmation.
• Management: priority, owner, and whether the wording is approved for use in the planned AI tests.

Worked prompt record: A3

Family: A-Q3. Audience and decision: finance lead checking provider suitability. Prompt: use the exact A3 wording above, with the selected firm’s name. Origin: team hypothesis in this illustration; no actual AI use has been observed. Context: standalone question, firm named, no supplied URL.

Verified fact: the responsible practitioner has confirmed that the firm does not offer audit services. Record who confirmed this and when. Answer requirements: do not attribute audit services to the firm; distinguish established facts from unresolved public evidence. Uncertainty: direct verification may be appropriate if available sources do not settle the question. Priority: initial. Owner and version: enter the designated operator and review date.

The confirmed service boundary and the hypothesis about what buyers ask are separate entries. Evidence supporting one does not automatically support the other.

Review the map with the practice lead and the person responsible for intake or business development. They should be able to trace each priority prompt to evidence or an explicit hypothesis, recognize the client situation, and explain what an accurate answer would accomplish.

The chapter is complete when the map covers the important decisions within scope, the library includes meaningful fit boundaries, and the team understands what its evidence can and cannot establish. Carry both outputs forward. Chapter 4 establishes the rules for using and publishing information; Chapter 5 uses the approved prompts to audit what AI systems actually say.