Research paper
Adaptive Diagnostic Methodology: Question-Led Context Extraction for Organizational Problem-Solving
Preprint, submitted to arXiv (cs.AI / cs.HC). Not yet peer reviewed.
Abstract
Practitioners in consulting, technology implementation, and organizational problem-solving often arrive at an engagement with the solution already chosen. When prescription comes before diagnosis, recommendations are shaped by what the practitioner assumes about the organization instead of what is true of it, a condition I term epistemic overreach. This paper introduces Adaptive Diagnostic Methodology (ADM), a question-led framework for extracting organizational context before anything is prescribed. ADM starts from the premise that question quality is the upstream variable of solution quality: a recommendation can be no more accurate than the context it was built on. The methodology defines a six-phase cycle (Inquiry, Extraction, Synthesis, Validation, Prescription, Adaptation), a six-category Question Taxonomy that moves a practitioner from surface symptoms to quantified root conditions, and a graduated completeness standard that says when a diagnostic record is sufficient to support prescription. The Synthesis and Validation phases produce structured context designed to feed AI implementation pipelines. Because ADM is defined by a structural condition and not by an industry, it applies to organizational diagnostics, technology scoping, messaging, employee onboarding, and sales conversations: anywhere a workflow prescribes to a person or organization from assumed context. I describe how I use ADM in my companies today, and I address its limitations, including respondent fatigue and the absence of controlled outcome evidence, along with the conditions under which it should not be used.
1. Introduction
Most people who have hired a consultant, a vendor, or an internal specialist have seen this pattern, though it is rarely named. The practitioner arrives with a solution in hand. The first meetings are nominally about understanding the client's situation, but the questions are shaped by the answer that is coming, and the recommendation that follows looks a great deal like the practitioner's existing product, framework, or specialty. What was presented as a diagnosis was closer to a sales process.
The cause is structural more than ethical. Practitioners build expertise by solving categories of problems, and that expertise becomes a cognitive prior: new situations look like ones they have seen before. Kahneman (2011) describes the human tendency to substitute fast, pattern-matched judgment for slower, effortful analysis, with System 1 thinking passing for deliberate reasoning. In professional services this appears as a habit of mapping new problems onto familiar solution templates. The resulting recommendations can be technically competent and still wrong for the organization in front of the practitioner.
The downstream costs are predictable. Implementations miss constraints that were never surfaced. Solutions treat symptoms. Organizations invest in change that does not hold, because it was designed around assumed needs. The practitioner moves on to the next engagement, and the client is left with the consequences.
This paper proposes Adaptive Diagnostic Methodology (ADM) as a formal alternative. ADM reverses the usual consulting sequence: questions come first, context is extracted instead of assumed, and the prescription waits until the diagnostic work is complete. My core thesis is that every good answer is found at the end of a great question. Solution quality cannot exceed question quality, because questions are how context enters the process.
ADM also has an AI dimension. Organizations increasingly use AI systems for decision support, content generation, project management, and analysis, and those systems are only as useful as the context they receive. The Synthesis and Validation phases of ADM produce structured outputs intended to give AI systems validated organizational context.
Section 2 places ADM within related work in consulting theory, organizational inquiry, and context engineering. Section 3 defines its design principles, Question Taxonomy, and six-phase cycle. Section 4 describes current use, works through five application domains, and sets out threats to validity. Section 5 discusses scope conditions, limitations, and implications for AI-augmented organizations. Section 6 identifies future directions, and Section 7 concludes.
2. Background and related work
2.1 Diagnostic and prescriptive paradigms
Management consulting has long carried a tension between diagnosis and prescription. The dominant commercial model, which I will call the expert model, treats the practitioner as a specialist who arrives with superior knowledge and recommends from patterns seen in earlier engagements. That model works well when a problem is well specified and the solution domain is bounded. It works poorly when the problem itself is unclear, contested, or embedded in organizational dynamics an outsider cannot see.
Edgar Schein identified this limitation in his work on process consultation (Schein, 1969). The expert model, he argued, assumes the client can diagnose its own problem accurately and describe it clearly, and in practice neither assumption reliably holds. The problem a client presents is often a symptom of a deeper systemic condition, and clients frequently do not know what they do not know. Schein proposed that the consultant's main job is to help the client build its own diagnostic capacity through a relational, iterative process.
Lean manufacturing offers a parallel discipline. The Toyota Production System's A3 method and the PDCA (Plan-Do-Check-Act) cycle attributed to Deming (Deming, 1986) both require a thorough problem definition before any corrective action is designed. The A3 format reserves a fixed share of the analysis for "current condition" before the practitioner reaches "countermeasures," and that ordering is the method. Womack and Jones (1996) document lean implementations that skipped diagnostic rigor and then optimized the wrong processes, achieving efficiency in areas that were never the constraint.
Medicine has the most formalized version of the principle. Evidence-based medicine's PICO framework (Population, Intervention, Comparison, Outcome) structures a clinical question before any search for evidence begins (Richardson et al., 1995). It was developed because physicians working from pattern recognition alone were selecting treatments without adequate problem framing. PICO does not generate answers. It generates precise questions, and those questions decide which evidence is relevant. A clinician who skips that framing builds a treatment plan from an assumed diagnosis, which is the medical form of epistemic overreach. A maxim with wide currency in medicine and consultative selling, and no single citable origin, states the standard bluntly, and this paper adopts it for knowledge work: prescription without diagnosis is malpractice. Khalsa and Illig formalize the same principle for client engagements as "diagnose before you prescribe" (Khalsa & Illig, 2008). No patient expects a prescription before the examination.
These traditions share a commitment: diagnosis precedes prescription, and diagnosis needs a method. ADM formalizes that commitment for knowledge work, including consulting, technology implementation, and organizational problem-solving, where A3, PDCA, and PICO were not designed to apply directly.
2.2 The Socratic method and organizational inquiry
The philosophical basis for question-led knowledge extraction is ancient; its organizational applications are recent. The Socratic method, systematic questioning to expose assumptions, test beliefs, and reach more reliable conclusions, maps onto diagnostic practice. A practitioner who asks why a problem exists, then asks what produced that condition, is practicing a form of Socratic inquiry adapted to organizations.
Cooperrider and Srivastva's (1987) Appreciative Inquiry (AI) is one influential adaptation. It changes the diagnostic question from "what is broken?" to "what is working, and what conditions produced it?" Its methodological insight, that the questions an organization asks shape the future it builds, anticipates ADM's central claim. The way a practitioner frames an inquiry decides what becomes visible.
Peter Senge's work on organizational learning adds further grounding (Senge, 1990). His mental models, the tacit assumptions and generalizations that shape how members of an organization see the world, describe the mechanism through which epistemic overreach operates. Practitioners bring mental models formed by prior experience, and without an explicit diagnostic process those models filter out disconfirming information. Senge's remedy, surfacing and testing mental models through structured dialogue, is a direct precursor to ADM's Inquiry and Extraction phases.
ADM differs from these traditions in scope and specificity. Appreciative Inquiry was designed for organizational transformation, and its question categories are broad and forward-looking. Senge's framework is mainly descriptive. ADM addresses a narrower and more common situation, a practitioner working with an organization or an individual to identify a problem and design a solution, and for that it needs actionable question categories, a defined phase sequence, and explicit integration with AI implementation pipelines.
2.3 Context engineering and AI integration
Large language models have added a new dimension to the diagnostic problem. An AI system has no access to organizational reality; it receives what it is given. If that context is shaped by assumption, the output inherits the distortion, and inaccurate inputs come back as confident, well-written outputs that do not fit the organization's actual situation.
The term context engineering, the deliberate construction of the information environment an AI system works in, was proposed by Tobi Lütke and popularized by Andrej Karpathy in mid-2025 (Lütke, 2025; Karpathy, 2025). Karpathy describes it as the art and science of filling the context window with just the right information for the next step. Output quality depends on context quality as well as model capability, which changes the first question of an AI implementation. Before choosing a model, an organization has to ask what context it is feeding the model and whether that context is accurate.
Van Clief and McDermott's (2026) Interpretable Context Methodology (ICM) addresses the structural side of that question. ICM defines a folder-based architecture for organizing context so that AI agents can navigate and interpret it, and it treats the context container as a design artifact that shapes agent behavior. ICM gives a rigorous answer to how context should be organized.
By design, ICM does not address how context should be extracted in the first place. It assumes accurate context exists and defines how to structure and transmit it. ADM addresses the prior question: how a practitioner extracts accurate context from an organization where that context is often implicit, contested, or hidden behind assumption. Extraction comes before structure, and no context architecture can produce reliable AI output from inaccurate inputs.
Kahneman's (2011) framework explains why explicit extraction protocols are needed. System 1 thinking (fast, associative, pattern-matched) dominates under time pressure and expertise, and experienced practitioners are especially susceptible because their expertise makes pattern-matching feel reliable. ADM is a System 2 intervention. It slows diagnosis down, requires questions to be formulated explicitly, and makes the practitioner engage with the specifics of the current situation before applying any pattern.
That grounding needs a qualification immediately. Fast pattern-matching is sometimes reliable. Kahneman and Klein (2009), reconciling two research traditions that had reached opposite conclusions, agree that intuitive expertise can be trusted only in high-validity environments: settings with regular, reasonably rapid, and unambiguous feedback, such as firefighting, chess, or anesthesiology. Where feedback is delayed, sparse, or confounded, the intuition that serves an expert elsewhere has no reliable basis. Most settings ADM addresses are low-validity in this sense. Outcomes arrive weeks or months later and depend on many variables, so a diagnosis is rarely followed by a clean signal that it was right. ADM therefore does not claim that structured inquiry beats expert intuition in general. Its claim is narrower: in low-validity environments, where intuition lacks the feedback that would make it dependable, deliberate extraction of context is the safer default. In high-validity environments the delay ADM introduces may not be worth its cost, a boundary Section 5.2 returns to.
2.4 Problem framing and the ceiling on solutions
ADM's central thesis, that solution quality is bounded by question quality, rests on the literature on problem framing and problem construction. Schön (1983) distinguishes problem setting from problem solving and argues that professionals rarely meet well-formed problems; they face indeterminate situations and must first name what they will treat as the problem and frame the context in which they attend to it. Rittel and Webber (1973), introducing wicked problems, go further: such problems have no definitive formulation, and understanding the problem is bound up with the idea one has for solving it, so formulating the problem is, in effect, the problem. If framing sets the solution space, the questions that produce the framing set an upper bound on any solution that follows.
Empirical work on problem construction points the same way. In studies of creative performance, how actively a person explores and constructs a problem before committing to a solution predicts the rated quality and originality of the result (Getzels & Csikszentmihalyi, 1976; Mumford et al., 1994; Reiter-Palmon & Robinson, 2009). Separately, self-generated elaboration (asking someone why a claim holds, as opposed to handing them the explanation) deepens processing of the material. The benefit grows with the responder's prior knowledge and can reverse when a low-knowledge responder generates a wrong explanation (Pressley et al., 1987; Dunlosky et al., 2013). That boundary supports ADM's structure: the practitioner supplies domain knowledge while the questions extract the situational context the practitioner lacks, which is the combination the elaboration literature finds most productive.
The defensible form of the thesis follows. Question quality is a necessary constraint on solution quality, not a sufficient cause of a good solution. Whether this holds in general, outside creative-task laboratories, is not settled, and reviews of behavioral interventions at scale find effects considerably smaller than laboratory work implies (DellaVigna & Linos, 2022). ADM's claim is bounded accordingly. Better questions raise the ceiling on the answer, and they do not guarantee the answer reaches it.
2.5 Adjacent question-led methods
Several established methods already pair structured questioning with professional work, and ADM's claim to novelty has to be stated against them. SPIN Selling (Rackham, 1988), derived from analysis of roughly 35,000 sales calls, sequences Situation, Problem, Implication, and Need-payoff questions and is the closest existing taxonomy to ADM's. The two differ in where they end. SPIN's questions are built to move a buyer toward the seller's solution, finishing with questions that lead the buyer to voice the payoff of a predetermined product. ADM extracts the situation whatever solution, if any, follows, and has nothing equivalent to the Need-payoff move. Jobs-to-be-Done (Christensen et al., 2016) and behavior-first interviewing (Fitzpatrick, 2013) scope customer discovery for a product decision; contextual inquiry (Beyer & Holtzblatt, 1998) extracts user context to inform a design; and the Question Formulation Technique (Rothstein & Santana, 2011) shows, in a classroom setting, that question generation can itself become a repeatable procedure. Each is tied to a domain (a sale, a product, an interface, a classroom) or describes extraction without making it a gated precondition to action.
The diagnose-before-prescribe ethos is already present in these methods and in the consulting canon, so ADM does not claim it. What ADM contributes is a combination of three properties that no earlier method holds together: a portable question taxonomy that is not tied to a single outcome; phase gates that make extraction a precondition to prescription, so a skilled practitioner cannot skip it; and outputs structured for direct use by AI systems. The third property is missing from every method above for a simple reason. All of them predate large language models as the downstream executor of a diagnosis.
3. Adaptive Diagnostic Methodology
3.1 Design principles
ADM rests on six principles. Each is an operational constraint that separates ADM practice from conventional consulting behavior.
Principle 1: Diagnose before prescribing. No recommendation is formed before the diagnostic cycle is complete. This directly opposes the expert model. A practitioner who feels a solution forming during the early phases should note it privately and keep asking questions. The solution may be correct; what changes is when it is introduced.
Principle 2: Questions precede solutions. This is a sequencing rule. The first deliverable of an ADM engagement is a question set, and it is reviewed, refined, and executed before the practitioner synthesizes any recommendation.
Principle 3: Context is extracted, not assumed. The practitioner's prior experience is a source of hypotheses, and it does not substitute for inquiry. What worked elsewhere suggests questions to ask here.
Principle 4: Adaptation is part of the cycle. Priorities shift, and new information changes the problem definition. ADM ends every cycle with an Adaptation phase that feeds the next Inquiry phase. It runs after successful cycles as well as failed ones.
Principle 5: Reality over perception. Organizations present themselves as they believe themselves to be, and that is not always how they are. The question categories are designed to surface disconfirming evidence: constraints nobody has stated, priorities that conflict with the stated ones, patterns that persist despite intervention. The methodology treats epistemic overreach as something to be actively countered.
Principle 6: Practitioner neutrality during diagnosis. During Inquiry and Extraction the practitioner does not share opinions, signal a preferred solution, or affirm the client's assumptions. The aim is to hear what is true, which may differ from what the client believes the practitioner expects to hear. Neutrality is lifted during Synthesis and Prescription.
3.2 Question Taxonomy
ADM sorts diagnostic questions into six categories, each with its own epistemic function. All six are used across an engagement, and their relative weight changes by phase.
| Question Type | Purpose | Example |
|---|---|---|
| Context Questions | Establish the current state without normative framing | "Walk me through what a typical week looks like for this team." |
| Root Questions | Identify the conditions that generate the observed problems | "How long has this been happening, and what changed around that time?" |
| Constraint Questions | Map the boundaries of the solution space before designing within it | "What has already been tried, and why didn't it hold?" |
| Magnitude Questions | Quantify the size, frequency, and cost of the problem, including the cost of leaving it unaddressed | "How often does this happen in a month, and what does each occurrence cost in hours or dollars?" |
| Priority Questions | Surface what matters in practice, which can differ from what is nominally important | "If you could only fix one thing in the next 90 days, what would it be?" |
| Validation Questions | Establish criteria for knowing whether the right problem was solved | "What would success look like in a way that would be hard to argue with?" |
Most engagements begin with Context Questions. They are open and descriptive so the client can tell the story without the practitioner imposing a frame. Root Questions follow once the surface narrative is established, and they push below symptoms toward causes. Constraint Questions are often overlooked, and many implementations fail because the constraint map was incomplete even though the solution was sound. Priority Questions expose the gap between stated and actual priorities, which is frequently where the most important diagnostic information sits. Validation Questions close the loop by defining success before the prescription is written, so the solution can be tested against a clear standard.
Magnitude Questions turn the narrative into measurement. A diagnosis without numbers is an opinion: a practitioner who knows that invoicing is painful, but not how many invoices, hours, or errors a month are involved, cannot rank it against any other problem. Magnitude Questions attach time, frequency, volume, cost, and failure rate to the problems the other categories surface, and they extend to stakes, meaning what the condition costs now and what it will cost if it is never addressed. Two disciplines govern the category. First, best-effort estimates are acceptable and expected. A prospect rarely has exact figures at hand, and stalling the session to chase precision defeats its purpose, so estimates are recorded as estimates and flagged for verification. Second, an inability to estimate is itself diagnostic data. An organization that cannot approximate its own cycle time, error rate, or cost per incident has revealed a measurement gap, and that gap often sits closer to the root condition than the symptom that opened the conversation.
3.3 The ADM cycle
ADM runs as a six-phase cycle. Each phase has defined inputs, a process, and an output that feeds the next. The cycle is meant to repeat: the Adaptation phase resets it with updated context.
| Phase | Input | Process | Output |
|---|---|---|---|
| 1. Inquiry | Engagement scope, practitioner hypotheses | Develop question set across all six taxonomy categories; sequence for the diagnostic session | Structured question set, ordered by category and dependency |
| 2. Extraction | Question set, access to organizational stakeholders | Execute diagnostic sessions; document responses verbatim before interpretation; run the gap-detection loop against completeness criteria | Raw context record: unfiltered, unedited organizational input, with explicit gap register |
| 3. Synthesis | Raw context record | Identify patterns, contradictions, and gaps; map stated problems to Root and Constraint findings; rank problem candidates by extracted Magnitude data | Diagnostic summary with problem candidates ranked by measured impact, with confidence levels |
| 4. Validation | Diagnostic summary | Return findings to stakeholders for accuracy confirmation; surface any corrections or additions | Validated problem statement, agreed upon by practitioner and client |
| 5. Prescription | Validated problem statement, constraint map | Design solution within confirmed constraints; sequence implementation against stated priorities | Recommendation document with implementation sequence and success criteria |
| 6. Adaptation | Implementation results, new organizational inputs | Compare outcomes to Validation criteria; identify what changed; reset Inquiry scope for next cycle | Updated context record, refined question set for next cycle |
Three parts of the cycle need more explanation.
Extraction uses verbatim documentation as a discipline. A practitioner who paraphrases in real time adds interpretation before the raw record is complete, so ADM requires the first pass to capture what was said. Interpretation waits for Synthesis.
Validation is the phase most often skipped in practice, and I consider it the most important. Taking a summary of what the practitioner heard back to stakeholders, before any recommendation is written, catches misinterpretation early. It also moves ownership of the problem: a client who has confirmed the problem statement is more likely to engage with the solution.
The separation of Prescription and Adaptation is deliberate. Many consulting frameworks treat adaptation as error recovery, something that happens when a recommendation fails. In ADM, Adaptation follows every implementation cycle whatever the outcome, because success also changes the organization's situation and a changed situation needs new inquiry.
3.4 Diagnostic completeness: a phased definition of done
A method has to say when the diagnosis is finished. Without a completion standard, practitioners fail in two opposite ways: they stop too early, which brings back the assumption-filling the method exists to prevent, or they never stop, and rigor turns into paralysis. ADM handles this with a phased definition of done, three completeness levels, each producing a usable artifact and each gating entry to a later phase.
Level 1: Narrative completeness. Every taxonomy category has at least a first-pass response in the client's own words. No category is empty, and the account could be read back to the client without objection. Level 1 is the minimum for closing an initial diagnostic session, and its artifact, a structured problem narrative, is enough to decide whether the engagement should continue.
Level 2: Measured completeness. The leading problem candidates carry Magnitude data (time, frequency, volume, cost, and failure rate), with estimates flagged as estimates. Prior attempts are documented along with why each did not hold, and priorities are forced-ranked. Level 2 closes Extraction and gates entry to Synthesis, since ranking problem candidates requires numbers to rank them by. Its artifact is a quantified diagnostic record.
Level 3: Validated completeness. The client has confirmed or corrected the diagnostic summary, success criteria are stated in falsifiable terms, and every remaining unknown is recorded as an explicit gap. Level 3 closes Validation and is the only gate to Prescription. No recommendation is written against a record below Level 3.
Within each level, a taxonomy category is satisfied when four criteria hold. The responses describe specific instances. Whatever can be quantified has been, at least by estimate. At least one follow-up question has probed for disconfirming evidence. Anything still unknown is written into the gap register. These criteria drive the gap-detection loop that runs throughout Extraction: after each response, the practitioner checks the record against the criteria for the active category, and each unmet criterion produces the next question. A category closes when its criteria are met or when the client cannot answer, and an inability to answer is recorded as diagnostic data, since what an organization cannot say about itself is often more informative than what it can.
The gap-detection loop is what separates an adaptive diagnostic from a scripted intake form. A form asks a fixed set of questions and accepts whatever comes back. The loop asks whatever the completeness criteria require next, so the depth of inquiry follows the ambiguity of the situation and not the length of a script. It also gives the levels their stepped structure. At each level the practitioner can stop, ship an artifact, and continue upward, instead of carrying an undivided mass of discovery that ends only when time runs out.
3.5 ADM outputs and AI-ready context
Each phase of the ADM cycle produces a specific type of structured content, and each type corresponds to a class of context that AI systems need in order to operate accurately. I hold that an organization deploying AI for knowledge work needs all six present and verified before AI output can be trusted.
| ADM Phase Output | Context Type Produced |
|---|---|
| Context Question responses | Current-state and organizational identity |
| Root Question responses | Historical patterns and causal context |
| Constraint Question responses | Operational constraints and prior attempts |
| Magnitude Question responses | Quantified baselines, impact data, and cost of inaction |
| Priority Question responses | Objectives and decision-weight context |
| Validated problem statement | Success definition and evaluation criteria |
An organization that runs ADM correctly therefore does not need a separate exercise to create AI context. The diagnostic process produces validated, categorized organizational context as a direct output, ready for a machine-readable context architecture without further translation.
The consequence for AI use is straightforward. An AI system given ADM output works from organizational input that stakeholders have checked. A system given context assembled without a diagnostic process works from whatever the organization believes about itself, which may or may not match its situation. To the extent output quality depends on input accuracy, it depends on what the system received as much as on the model.
4. Applications
ADM is defined by a structural condition. It applies wherever a workflow begins from assumptions about a person or an organization whose actual context could be extracted by asking. The cases below were chosen because they make the mechanism easy to see; they do not mark its limits. Consulting is where the pattern is most visible, and the same mechanism governs technology scoping, messaging, onboarding, and sales conversations, each of which prescribes something (a system, a message, a first-week experience, a pitch) from an assumed model of the person on the other side. Substitute the diagnostic subject and the six phases hold: a client, a new hire, a prospect, and a target customer are each a context to be extracted before anything is prescribed to them.
4.1 Current use
I use ADM in my companies, principally Dovito Business Solutions, in three settings. With consulting clients, it structures the conversations that establish context before any recommendation is made. In Dovito's own delivery, internal AI agents run ADM before design work begins, so that a project is specified from extracted context before anything is built. In client-facing communication, beginning with the company website (dovito.com), the taxonomy supplies the questions Dovito asks prospects and clients at each stage of the relationship.
These uses are the practical basis for the method described here. I have not yet measured their outcomes in a controlled way, and I make no claim about results that such measurement would need to support (see Section 4.7).
4.2 Organizational diagnostic
Consider a professional services firm that hires a business consultant to address declining client retention. A prescriptive approach would audit the firm's onboarding process, compare it to industry benchmarks, and recommend improvements. That may be right, but it assumes the problem is onboarding, and the assumption may not survive a rigorous diagnostic.
An ADM engagement begins differently. Context Questions establish the full picture: which clients are leaving, at what point in the relationship, and what their exit communications say. Root Questions go deeper, asking when the decline began, what else changed at that time, and how long-retained clients describe their experience compared with those who left. Constraint Questions map the organization's limits, including what has already been tried, how much capacity for change each team has, and what resources any solution can draw on.
In a case like this, the process might show that retention tracks a change in account-management staffing more closely than it tracks onboarding. The diagnostic would redirect the engagement. Without it, the consultant could deliver a sound onboarding improvement with no effect on the actual problem.
4.3 Technology implementation
Software implementations depend heavily on requirements work. Field research identifies requirements engineering practices as a clear contributor to project success (Hofmann & Lehner, 2001), and defects traced to the requirements phase are among the most expensive to fix, with post-delivery fixes costing up to one hundred times more than fixes made during requirements and design (Boehm & Basili, 2001). A system built to solve a problem that was never the constraint is the costliest requirements defect of all. Applied at the scoping stage, ADM changes the opening of a technology project from requirements gathering, which assumes the problem is known, to context extraction, which does not.
The taxonomy is especially useful here. Constraint Questions bring out technical, organizational, and political constraints that people rarely volunteer and that often decide the outcome. A company may ask for an AI-assisted customer communication tool, and a Constraint Question session may show that the real bottleneck is response time on complex inquiries, which calls for a different architecture than a general communication assistant. Priority Questions separate the features that matter to the person approving the budget from those that matter to the people doing the work, and the two sets often differ.
In a technology context, Extraction and Synthesis produce a validated scope document: a structured account of the problem, the constraint map, and the success criteria. That document is a context package that can be handed to an AI development environment as it stands.
4.4 Marketing and messaging strategy
Marketers face the same assumption problem as consultants. Copy and campaigns are built on what the practitioner believes the audience wants to hear, often based on what similar audiences wanted elsewhere, and the specific customer's decision process, objections, and language frequently go uninvestigated.
Applied to messaging, ADM treats the target customer as a context to be extracted. Context Questions establish how customers already talk about the problem the product solves, capturing their language in place of the company's preferred terms. Root Questions probe the conditions under which customers feel the problem most acutely, which informs timing and placement. Priority Questions separate the concerns that decide a purchase from those that are raised but do not block it.
The resulting customer-context document comes before any creative work. Its value is in how a message gets recognized: customers recognize a message that uses the words they already use for the problem, and extracted context captures those words. A message built on assumed context may still land, but its author cannot know which words the customer was waiting to hear.
A completed diagnostic record also surfaces more distinct, addressable problems than a narrow intake does. Questioning across all six categories, at increasing depth, exposes conditions a single presenting complaint would never reveal, and each one is a candidate that some solution could address. I am not claiming the method sells more work. A practitioner who extracts the whole picture holds more true statements about the client's situation, and some describe problems the client did not open with. Whether any of them becomes follow-on work is a separate question the methodology does not settle.
The diagnostic record can also shape the message itself, where a question becomes a way of holding attention. A prospect who reads a question naming their exact situation ("Still stitching your scheduling tool to your payment system by hand at month end?") answers it internally and is drawn into the problem before any solution appears. The persuasion literature offers two mechanisms. A question makes salient a specific gap between what the reader knows and what they want resolved, and an open gap motivates its own closure (Loewenstein, 1994). A question about the reader's own circumstances also induces self-referencing, which deepens engagement with what follows (Burnkrant & Unnava, 1995). This device sits downstream of ADM. The hook works only when it is built from language the diagnostic surfaced, because a question the reader recognizes as their own opens a real gap and a generic one opens none.
The effect is conditional, and I claim no more than the mechanism. The same research shows the device reversing when it is overused or when its persuasive intent becomes obvious; attention then moves from the problem to the persuader, and the question backfires (Petty et al., 1981; Ahluwalia & Burnkrant, 2004). The related question-behavior effect, in which asking someone about an intention shifts their later behavior, is small on average and turns negative in face-to-face settings while holding in anonymous one-to-many channels (Spangenberg et al., 2016). That contrast marks the boundary. The question-as-hook belongs in broadcast copy and not in the live diagnostic interview, where the same phrasing can read as manipulation. ADM's role is upstream, supplying the extracted, validated language that decides whether a hook addresses a gap the reader actually has.
4.5 Employee onboarding
Onboarding is a diagnostic problem that is rarely treated as one. The organization prescribes a first week, a training sequence, and a set of introductions from an assumed model of what the new hire needs, while the new hire pieces together the role, the norms, and the informal system by observation and scattered questions. The socialization literature has long described organizational entry as a two-way information problem. Newcomers seek technical, referent, normative, performance-feedback, and social information (Morrison, 1993), using tactics that range from overt questions to quiet monitoring (Miller & Jablin, 1991), and organizations that attend to compliance, clarification, culture, and connection onboard more successfully (Bauer, 2010). This literature describes the extraction but does not turn it into a procedure. Applied to onboarding, ADM treats the incoming hire as a context to be extracted before the plan is written. Context and Root Questions establish what the person already knows and how they work, Constraint Questions find what would block them, and Priority Questions establish what a successful first ninety days would contain from the hire's point of view. The output is a per-hire context record that the same downstream systems can act on.
4.6 Sales scripting
A sales script is a prescription issued before the diagnosis, the pattern ADM reverses. Scripts written from an assumed model of the prospect ask questions built to steer toward a known offer, the move SPIN (Rackham, 1988) formalizes and the one Section 2.5 distinguishes ADM from. ADM does not turn the sales conversation into an interrogation. It reorders the conversation so the Question Taxonomy runs before the pitch takes shape. The script becomes adaptive: its branches follow what each answer leaves unresolved, and its eventual recommendation rests on the prospect's stated constraints and priorities. This is the gap-detection discipline of Section 3.4 applied to a commercial conversation, and it carries the same limit discussed below. A prospect can be over-questioned, so the method depends on stopping when the picture is complete, even if the script has more lines.
4.7 Threats to validity
ADM's diagnostic quality depends on conditions that are not always present. I note five limitations.
First, ADM requires an organization that is willing to be questioned. Some clients have already decided on a solution and want validation. ADM cannot extract accurate context from a client unwilling to engage openly, so practitioners should assess willingness at the outset and should not apply ADM where it will meet structural resistance.
Second, question quality depends on practitioner skill. The taxonomy gives categories and examples, but writing questions that surface root conditions, without leading the client toward the practitioner's preferred diagnosis, takes experience and self-awareness. A poorly run Inquiry phase can produce context biased toward the practitioner, which defeats the method's purpose.
Third, ADM's findings do not transfer from one organization to the next. Two organizations with similar surface problems can have entirely different root conditions, and ADM's value comes from not assuming otherwise. It cannot produce the generalizable playbooks some organizations prefer; each engagement produces findings specific to its context.
Fourth, the claim that ADM improves outcomes is currently an argument from mechanism. No controlled study shows that structured diagnostic protocols produce better decisions than expert judgment in knowledge work, and the nearest adjacent evidence is mixed: controlled trials of cognitive-forcing and debiasing strategies in clinical reasoning showed little or no reduction in diagnostic error (Sherbino et al., 2014; Norman et al., 2017). ADM differs from those interventions in one respect. They ask practitioners to reason more carefully with information they already hold, while ADM changes what information is available by extracting context the practitioner does not yet have. Whether that difference yields measurably better outcomes is an open empirical question. This paper argues that the mechanism is sound and worth testing; it does not claim the result.
Fifth, questioning has a natural limit, and past it the method degrades the context it exists to gather. Survey and interview research is consistent here. As the number of questions grows, people move from considered answers to merely adequate ones, agreeing without reflecting, skipping, or flattening their responses, a drift the survey literature calls satisficing (Krosnick, 1991; Herzog & Bachman, 1981; Galesic & Bosnjak, 2009). A diagnostic that mistakes a long question list for thoroughness produces a record that looks complete and reads as fatigued. ADM's defense is structural. The gap-detection loop of Section 3.4 asks only what the completeness criteria still require and stops probing a category once it is satisfied or the respondent cannot answer, so depth goes where the record is thin. Two practices reduce the load further. Best-effort estimates are accepted in place of exact figures, which removes the most tiring demand, and inquiry may open with a hypothesis, so a well-briefed practitioner confirms and corrects instead of asking the respondent to assemble the picture from nothing. This also separates delivery from the sales conversation that comes before it. Industry research on buyer fatigue warns against open-ended fishing (Dixon & Adamson, 2011), but that caution concerns how to win an engagement, not how to conduct structured extraction once one is underway. None of this rules out fatigue. The bound holds only as far as the loop's stopping criteria are well specified and honestly applied, which is a demand on the practitioner and the tooling.
5. Discussion
5.1 Where ADM works
ADM is most effective in knowledge work where the problem is ambiguous, contested, or embedded in organizational dynamics an outsider cannot see. That includes management and strategy consulting, technology scoping, organizational design, product discovery, and any setting where the practitioner's first hypothesis about the problem may be wrong.
It also suits organizations where earlier interventions have failed or only partly worked. When several solutions to a persistent problem have already been tried, the Root and Constraint categories are well placed to surface why they did not hold. People rarely volunteer that information, and it is often the most important input for designing an intervention that works.
Complex AI implementations are a growing application. As organizations deploy AI for more functions, the quality of the context those systems work from becomes a strategic variable, and ADM offers a disciplined way to give them accurate context in place of organizational mythology.
5.2 Where ADM does not work
ADM introduces diagnostic latency, because the prescription waits for the cycle to finish. In time-critical operations such as emergency response or acute crisis management, where delay costs more than an imperfect fast solution, that latency is unacceptable. ADM is not a framework for real-time decisions.
Commodity tasks with fully specified requirements do not benefit either. When the problem is known, the constraints are documented, and the solution space is well understood, a diagnostic phase adds process and no value. ADM is for working through ambiguity, and without ambiguity it is unnecessary.
Finally, ADM does not fit engagements whose client relationship cannot support the level of inquiry it requires. Short transactional engagements, RFP-driven procurement, and relationships in which the client is still deciding whether to hire the practitioner are poor conditions for it.
5.3 Epistemic overreach
I use epistemic overreach for the practice of generating solutions, recommendations, or diagnoses from perceived or assumed context. It is an act on the practitioner's part: reaching past what is known toward what feels known, and producing output as if the two were the same.
Epistemic overreach compounds. Each recommendation built on assumed context adds a layer of assumption that later recommendations treat as established. Long-tenured leadership, stable market positions, and past success all speed this up, because success reduces the pressure to question assumptions, and unquestioned assumptions become the ground for the next round of overreach.
ADM is designed to counter it, and the Validation phase in particular brings it into view. Returning a diagnostic summary to stakeholders and asking them to confirm or correct it often yields the most valuable context of the engagement. Their corrections show where the client's self-presentation during Extraction differed from how the client actually recognizes the problem, and that discrepancy is diagnostic data.
5.4 Diagnosis as a strategic position
For a firm that adopts it, ADM is also a strategic position in Porter's sense. Porter distinguishes operational effectiveness, doing similar activities better than rivals, from strategy, doing different activities or similar activities differently, and argues that operational effectiveness alone is not a durable advantage because best practices spread (Porter, 1996). A better discovery process alone would be copied. My claim is that ADM is a different set of activities, and that the trade-offs it demands are what protect it from imitation.
Read strategically, the scope conditions in Sections 4.7 and 5.2 are those trade-offs. A firm practicing ADM does not present a proposal in the first meeting, accepts diagnostic latency, and declines engagements in which the client wants validation instead of diagnosis. Each is a real cost, and Porter's point is that real costs make a position defensible: a competitor whose commercial model depends on solution-led selling cannot adopt diagnosis-led engagement without dismantling its own sales motion (Porter, 1996). This is also why publishing the methodology does not weaken the position. The barrier to imitation is the trade-off, and secrecy plays no part in it. Rumelt makes a complementary point from the strategy side: the kernel of a good strategy begins with diagnosis, and bad strategy often fails by skipping it (Rumelt, 2011). A method that operationalizes diagnosis is therefore the entry condition for strategy work.
The activities also reinforce one another in the way Porter's activity-system analysis describes (Porter, 1985; Porter, 1996). Question-led engagement produces validated context. Validated context feeds AI-assisted execution. Results are compared against the Validation criteria during Adaptation, and Adaptation resets Inquiry from a higher baseline. Copying any one of these activities yields little on its own.
Over repeated cycles, this behaves like the flywheel Collins describes, where no single turn is decisive and each adds to the momentum of the last (Collins, 2001). The completeness levels of Section 3.4 act as its ratchet points. Context validated at Level 3 in one cycle becomes the baseline for the next cycle's Inquiry, so later extraction concentrates on what changed. Over time the practitioner holds an accumulating, validated, machine-readable model of the client's business, which deepens with each cycle and which a competitor arriving later does not have.
The commercial model follows from the medical framing in Section 2.1. A patient pays for the office visit that determines why the symptoms exist, separately from and before any treatment. On this view the diagnostic is the first engagement, with a standalone deliverable (the validated diagnostic record) that keeps its value even if no implementation follows. A firm that gives diagnosis away signals that it has little value, and a firm that prices it signals that it is the core of the work.
5.5 Implications for AI-augmented organizations
AI systems that take part in organizational decisions, communication, and analysis depend on context more explicitly than human practitioners do. A human consultant can draw on embodied experience and on what they notice in the room to partly make up for an incomplete or inaccurate briefing. An AI system has no such fallback; it works from what it receives.
As organizations rely more on AI-assisted work, the accuracy and completeness of the context those systems receive becomes a major factor in their performance, alongside model capability. A less capable model given accurate, well-structured context can outperform a more capable one working from assumed or bloated context.
ADM supplies the discipline for what I call context hygiene: regularly extracting, validating, and structuring organizational context so AI systems work from the organization as it is. Organizational context keeps changing as priorities shift, constraints evolve, and new information redefines the problem, so this cannot be a one-time configuration. The Adaptation phase, feeding new Inquiry cycles, is what keeps the context current.
6. Future directions
6.1 ADM-trained AI assistants
Current large language models are optimized to respond. Their default is to answer, which builds in the prescriptive pattern ADM is meant to counter: given a problem statement, a model generates a solution whether or not the context behind the statement is accurate or complete.
A natural extension is an AI assistant trained to apply the methodology, one that, presented with an organizational problem, first asks a structured question set and only then synthesizes a response. The taxonomy in Section 3.2 could serve as a training scaffold. Fine-tuning on ADM question sequences, with evaluation criteria that reward diagnostic depth over response speed, could produce assistants that work as diagnostic partners.
Such assistants would be most useful in consulting, product discovery, and implementation scoping, where a fast answer that misreads the context is expensive.
6.2 Question quality metrics
ADM identifies question categories and gives examples, but it does not yet offer a scoring rubric for question quality. That limits organizations training practitioners and anyone evaluating an AI system's diagnostic ability.
A rubric would need to assess at least four dimensions: specificity (does the question target a defined information gap?), neutrality (does it avoid leading the respondent?), dependency (is it placed correctly relative to the information it needs from earlier questions?), and yield (does it reliably produce decision-relevant context?). Developing one through structured practitioner evaluation and testing across varied engagements is a worthwhile research program.
6.3 Context hygiene as organizational practice
Combined with a structured context architecture, ADM could serve as an operating layer for AI deployment: a continuous institutional practice in addition to a method for individual engagements.
In that model, ADM provides a quarterly or semi-annual diagnostic cycle that refreshes organizational context, asking what the current state is, what has changed, what the new constraints are, and what the priorities are now. The extracted context is organized in a machine-readable architecture and made available to AI systems across the organization's work.
Current practice usually treats AI context as a configuration step. I expect organizations that treat context as a living resource, extracted regularly and structured formally, to gain a systematic advantage over those that set it once and rarely revisit it. Making this an institutional operating standard instead of a per-project practice is the long-term direction of this work.
7. Conclusion
The pattern this paper describes is old, and naming it precisely is what allows it to be addressed. Practitioners arrive with solutions, and those solutions shape the questions they ask. The questions shape the context they receive, and that context confirms the solutions they already had. The resulting work is technically competent and misaligned with the organization's context, and it recurs because the cycle is structural.
ADM interrupts the cycle where it starts. It requires questions before solutions, extracted context in place of assumed context, and validation before any prescription, so the solution follows from the diagnosis. The process is no harder than the alternative, and it is more honest.
The same discipline matters more in AI-augmented organizations, because an AI system has no experience to fall back on when its briefing is wrong. The structured outputs of the ADM cycle are designed to give AI systems validated organizational context, offering a practical model for organizations that want AI assistance grounded in what they have extracted about themselves.
The central claim can be stated directly: a solution can be no better than the questions that produced the diagnosis behind it. ADM is a formal method for acting on that claim.
Disclosure
I founded Dovito Business Solutions, which developed ADM, and I use ADM in my companies' commercial work, as described in Section 4.1. ADM™ is a trademark of Dovito Business Solutions.
References
Ahluwalia, R., & Burnkrant, R. E. (2004). Answering questions about questions: A persuasion knowledge perspective for understanding the effects of rhetorical questions. Journal of Consumer Research, 31(1), 26–42. https://doi.org/10.1086/383421
Bauer, T. N. (2010). Onboarding new employees: Maximizing success. SHRM Foundation.
Beyer, H., & Holtzblatt, K. (1998). Contextual design: Defining customer-centered systems. Morgan Kaufmann.
Boehm, B., & Basili, V. R. (2001). Software defect reduction top 10 list. Computer, 34(1), 135–137. https://doi.org/10.1109/2.962984
Burnkrant, R. E., & Unnava, H. R. (1995). Effects of self-referencing on persuasion. Journal of Consumer Research, 22(1), 17–26. https://doi.org/10.1086/209432
Christensen, C. M., Hall, T., Dillon, K., & Duncan, D. S. (2016). Know your customers' "jobs to be done." Harvard Business Review, 94(9), 54–62.
Collins, J. (2001). Good to great: Why some companies make the leap... and others don't. HarperBusiness.
Cooperrider, D. L., & Srivastva, S. (1987). Appreciative inquiry in organizational life. Research in Organizational Change and Development, 1, 129–169.
DellaVigna, S., & Linos, E. (2022). RCTs to scale: Comprehensive evidence from two nudge units. Econometrica, 90(1), 81–116. https://doi.org/10.3982/ECTA18709
Deming, W. E. (1986). Out of the crisis. MIT Press.
Dixon, M., & Adamson, B. (2011). The challenger sale: Taking control of the customer conversation. Portfolio/Penguin.
Dunlosky, J., Rawson, K. A., Marsh, E. J., Nathan, M. J., & Willingham, D. T. (2013). Improving students' learning with effective learning techniques: Promising directions from cognitive and educational psychology. Psychological Science in the Public Interest, 14(1), 4–58. https://doi.org/10.1177/1529100612453266
Fitzpatrick, R. (2013). The mom test: How to talk to customers and learn if your business is a good idea when everyone is lying to you. CreateSpace.
Galesic, M., & Bosnjak, M. (2009). Effects of questionnaire length on participation and indicators of response quality in a web survey. Public Opinion Quarterly, 73(2), 349–360. https://doi.org/10.1093/poq/nfp031
Getzels, J. W., & Csikszentmihalyi, M. (1976). The creative vision: A longitudinal study of problem finding in art. Wiley.
Herzog, A. R., & Bachman, J. G. (1981). Effects of questionnaire length on response quality. Public Opinion Quarterly, 45(4), 549–559. https://doi.org/10.1086/268687
Hofmann, H. F., & Lehner, F. (2001). Requirements engineering as a success factor in software projects. IEEE Software, 18(4), 58–66. https://doi.org/10.1109/MS.2001.936219
Kahneman, D. (2011). Thinking, fast and slow. Farrar, Straus and Giroux.
Kahneman, D., & Klein, G. (2009). Conditions for intuitive expertise: A failure to disagree. American Psychologist, 64(6), 515–526. https://doi.org/10.1037/a0016755
Karpathy, A. [@karpathy]. (2025, June 25). +1 for "context engineering" over "prompt engineering" [Post]. X. https://x.com/karpathy/status/1937902205765607626
Khalsa, M., & Illig, R. (2008). Let's get real or let's not play: Transforming the buyer/seller relationship. Portfolio/Penguin.
Krosnick, J. A. (1991). Response strategies for coping with the cognitive demands of attitude measures in surveys. Applied Cognitive Psychology, 5(3), 213–236. https://doi.org/10.1002/acp.2350050305
Loewenstein, G. (1994). The psychology of curiosity: A review and reinterpretation. Psychological Bulletin, 116(1), 75–98. https://doi.org/10.1037/0033-2909.116.1.75
Lütke, T. [@tobi]. (2025, June 19). I really like the term "context engineering" over prompt engineering [Post]. X. https://x.com/tobi/status/1935533422589399127
Miller, V. D., & Jablin, F. M. (1991). Information seeking during organizational entry: Influences, tactics, and a model of the process. Academy of Management Review, 16(1), 92–120. https://doi.org/10.2307/258608
Morrison, E. W. (1993). Newcomer information seeking: Exploring types, modes, sources, and outcomes. Academy of Management Journal, 36(3), 557–589. https://doi.org/10.2307/256592
Mumford, M. D., Reiter-Palmon, R., & Redmond, M. R. (1994). Problem construction and cognition: Applying problem representations in ill-defined domains. In M. A. Runco (Ed.), Problem finding, problem solving, and creativity (pp. 3–39). Ablex.
Norman, G. R., Monteiro, S. D., Sherbino, J., Ilgen, J. S., Schmidt, H. G., & Mamede, S. (2017). The causes of errors in clinical reasoning: Cognitive biases, knowledge deficits, and dual process thinking. Academic Medicine, 92(1), 23–30. https://doi.org/10.1097/ACM.0000000000001421
Petty, R. E., Cacioppo, J. T., & Heesacker, M. (1981). Effects of rhetorical questions on persuasion: A cognitive response analysis. Journal of Personality and Social Psychology, 40(3), 432–440. https://doi.org/10.1037/0022-3514.40.3.432
Porter, M. E. (1985). Competitive advantage: Creating and sustaining superior performance. Free Press.
Porter, M. E. (1996). What is strategy? Harvard Business Review, 74(6), 61–78.
Pressley, M., McDaniel, M. A., Turnure, J. E., Wood, E., & Ahmad, M. (1987). Generation and precision of elaboration: Effects on intentional and incidental learning. Journal of Experimental Psychology: Learning, Memory, and Cognition, 13(2), 291–300. https://doi.org/10.1037/0278-7393.13.2.291
Rackham, N. (1988). SPIN selling. McGraw-Hill.
Reiter-Palmon, R., & Robinson, E. J. (2009). Problem identification and construction: What do we know, what is the future? Psychology of Aesthetics, Creativity, and the Arts, 3(1), 43–47. https://doi.org/10.1037/a0014629
Richardson, W. S., Wilson, M. C., Nishikawa, J., & Hayward, R. S. A. (1995). The well-built clinical question: A key to evidence-based decisions. ACP Journal Club, 123(3), A12–A13. https://doi.org/10.7326/ACPJC-1995-123-3-A12
Rittel, H. W. J., & Webber, M. M. (1973). Dilemmas in a general theory of planning. Policy Sciences, 4(2), 155–169. https://doi.org/10.1007/BF01405730
Rothstein, D., & Santana, L. (2011). Make just one change: Teach students to ask their own questions. Harvard Education Press.
Rumelt, R. P. (2011). Good strategy bad strategy: The difference and why it matters. Crown Business.
Schein, E. H. (1969). Process consultation: Its role in organization development. Addison-Wesley.
Schön, D. A. (1983). The reflective practitioner: How professionals think in action. Basic Books.
Senge, P. M. (1990). The fifth discipline: The art and practice of the learning organization. Doubleday.
Sherbino, J., Kulasegaram, K., Howey, E., & Norman, G. (2014). Ineffectiveness of cognitive forcing strategies to reduce biases in diagnostic reasoning: A controlled trial. CJEM, 16(1), 34–40. https://doi.org/10.2310/8000.2013.130860
Spangenberg, E. R., Kareklas, I., Devezer, B., & Sprott, D. E. (2016). A meta-analytic synthesis of the question–behavior effect. Journal of Consumer Psychology, 26(3), 441–458. https://doi.org/10.1016/j.jcps.2015.12.004
Van Clief, J., & McDermott, D. (2026). Interpretable context methodology: Folder structure as agentic architecture (arXiv:2603.16021). arXiv. https://arxiv.org/abs/2603.16021
Womack, J. P., & Jones, D. T. (1996). Lean thinking: Banish waste and create wealth in your corporation. Simon & Schuster.
Submitted to arXiv, cs.AI / cs.HC David Coleman, Dovito Business Solutions, Fort Collins, Colorado Contact: david@dovito.com