Customer Discovery and Evidence-Based Validation

Introduction

Entrepreneurial opportunity is not validated because a founder believes in it, because a prototype receives compliments, or because a market appears large in a published report. It is validated progressively when credible evidence shows that a defined group of customers experiences a meaningful problem, values a proposed solution, can access and adopt it, and is willing to exchange money, time, data, or some other scarce resource for it.

Customer discovery and evidence-based validation provide the discipline for reducing uncertainty before substantial resources are committed. They are particularly important in new ventures because entrepreneurs operate with incomplete information about customers, technologies, competitors, regulations, channels, and business models. The central managerial task is therefore not merely to build a product, but to design a sequence of learning activities that tests the assumptions on which the venture depends.

This chapter presents a practical and analytical approach to formulating assumptions, conducting customer interviews, designing experiments, developing minimum viable products (MVPs), and interpreting evidence through lean start-up learning cycles. It also examines the limitations and ethical risks of these methods.

Learning Objectives

By the end of this chapter, students should be able to:

1. Distinguish an entrepreneurial vision from the assumptions that must be tested.
2. Segment customers and formulate testable problem–solution hypotheses.
3. Design and conduct high-quality customer discovery interviews.
4. Select experiments and MVPs appropriate to different types of uncertainty.
5. Interpret qualitative and quantitative evidence without overstating conclusions.
6. Apply build–measure–learn cycles to decisions about perseverance, iteration, or pivoting.
7. Evaluate problem–solution fit and product–market fit using behavioural evidence.
8. Recognize methodological, ethical, and governance limitations in validation practices.

1. The Logic of Evidence-Based Entrepreneurship

A new venture is a system of interconnected assumptions. Typical assumptions include:

• Customer assumption: a particular group experiences the problem.
• Problem assumption: the problem is sufficiently frequent, costly, urgent, or frustrating.
• Solution assumption: the proposed solution addresses the underlying need.
• Behavioural assumption: customers will adopt a new behaviour or workflow.
• Economic assumption: customers will pay enough to support a viable business.
• Channel assumption: the venture can reach customers efficiently.
• Operational assumption: the venture can deliver the promised value reliably.
• Regulatory assumption: the business can operate lawfully and safely.

An assumption becomes a hypothesis when it is stated in a form that could be supported or contradicted by observable evidence. “Small retailers need better inventory software” is too broad. A more useful hypothesis is: “Independent grocery stores with two to five outlets experience weekly stock-outs because replenishment decisions rely on spreadsheets, and at least 30% of interviewed owners will agree to test an automated forecasting service.”

A good hypothesis specifies a population, a situation or problem, an expected behaviour or outcome, and a threshold for action. The threshold prevents founders from changing their interpretation after seeing the results.

Assumption mapping

Assumptions should be ranked by both importance and uncertainty. A high-importance, high-uncertainty assumption deserves early testing because its failure could invalidate the venture. Conversely, a low-importance assumption can often be postponed.

An assumption map is not a prediction of the future. It is a prioritization device for deciding what to learn first.

2. Customer Discovery and Problem Definition

Customer discovery is a structured process of learning about customers’ contexts, behaviours, constraints, and desired outcomes. It is not primarily a sales presentation. Its purpose is to discover whether a problem exists independently of the proposed solution and whether the founder understands the customer’s decision environment.

From demographic segments to problem situations

Traditional segmentation often begins with demographic descriptors such as age, income, or industry. These may be useful, but they are insufficient for early validation. Two customers with identical demographics may have very different needs because they differ in workflow, authority, urgency, or access to alternatives.

A stronger early segment describes:

• the person or organization;
• the situation in which the problem occurs;
• the existing method used to manage it;
• the consequences of failure; and
• the person who experiences, influences, and pays for the solution.

These roles may not be identical. In business markets, the user, economic buyer, procurement officer, and compliance approver can be different people. A product that satisfies users but cannot pass procurement has not achieved practical validation.

Problem severity

Problem severity is multidimensional. A problem may be frequent but inexpensive, rare but catastrophic, or emotionally frustrating but difficult to monetize. Useful dimensions include frequency, financial cost, time cost, risk, emotional burden, strategic importance, and the customer’s current willingness to search for alternatives.

Evidence is stronger when customers describe recent behaviour rather than hypothetical opinions. “Tell me about the last time this happened” is generally more informative than “Would you use an app that solved this?”

3. Designing Customer Interviews

Customer interviews are qualitative instruments for discovering language, context, causal explanations, and decision processes. They do not automatically estimate market size or prove demand. Their quality depends on sampling, questioning, listening, recording, and interpretation.

Interview design principles

A high-quality interview usually follows these principles:

1. Recruit for relevance. Select people who have recently experienced the situation, not merely people who match a broad demographic profile.
2. Ask about the past. Recent behaviour is more reliable than predictions about future behaviour.
3. Avoid leading questions. Do not embed the desired answer in the question.
4. Explore workarounds. Existing spending, manual processes, substitutions, and complaints reveal problem intensity.
5. Probe consequences. Ask what happened, who was affected, and what the customer did next.
6. Separate discovery from pitching. Present a concept only after understanding the existing problem.
7. Record evidence systematically. Capture quotations, observations, behaviours, and interpretations separately.

Questions such as “What is the most difficult part of preparing your monthly report?” are preferable to “Would an automated reporting platform help you?” Follow-up prompts include “How do you handle that today?”, “When did this last occur?”, “What did it cost?”, and “Why have you not changed the process?”

Sampling and saturation

Early discovery generally uses purposive sampling rather than a statistically representative sample. The purpose is to identify patterns and meaningful variation. The researcher should nevertheless seek diversity within the target segment and document inclusion criteria.

Interviews may approach thematic saturation when additional conversations produce few new themes. Saturation is not a fixed number. It depends on segment heterogeneity, research quality, and the complexity of the phenomenon. It should not be used to claim that a market has been statistically measured.

Biases and interview validity

Founders often suffer from confirmation bias: they notice statements supporting their preferred idea and discount contradictory evidence. Other risks include selection bias, courtesy bias, recall error, social desirability bias, and interviewer effects. Mitigation includes using a neutral interviewer, recording disconfirming evidence, asking the same core questions across interviews, and having another researcher independently code a sample of transcripts.

A simple coding structure can classify evidence as:

• observed behaviour;
• stated belief;
• current workaround;
• consequence or cost;
• emotional response;
• purchasing authority; and
• willingness to take a next step.

Observed behaviour and costly workarounds generally carry more evidential weight than positive opinions.

4. Experiments for Testing Venture Assumptions

An experiment is a deliberate test that exposes an assumption to observable consequences. A rigorous experiment specifies the hypothesis, target population, intervention, measurement, decision rule, time period, and risks.

The experiment canvas

Before conducting a test, founders should answer:

• What is the riskiest assumption?
• What observable behaviour would support it?
• What result would contradict it?
• What is the smallest ethical test that can produce useful evidence?
• What alternative explanations could account for the result?
• What decision will follow each possible outcome?

A useful distinction is between leading indicators and lagging indicators. A click may be a leading indicator of interest; repeated use, payment, retention, or referral is stronger evidence of realized value. The appropriate metric depends on the stage and the assumption being tested.

Common experiment types

Problem interviews and observation test whether the problem exists and how it is experienced.

Concept tests present a carefully bounded description or prototype to examine comprehension, relevance, and objections. Positive reactions are weak evidence unless followed by action.

Concierge MVPs deliver the proposed value manually to a small number of customers. They are useful when automation is expensive but the service can be performed by people.

Wizard-of-Oz MVPs make the system appear automated while humans perform critical operations behind the interface. This tests customer-facing value without prematurely building complex technology.

Landing-page experiments test whether a defined audience responds to a specific value proposition. They can measure attention and intent, but traffic source, message quality, and novelty can confound results.

Pre-orders and deposits test a stronger form of commitment. They must be handled transparently, with clear refund and delivery terms.

Paid pilots test value, procurement, implementation, and willingness to pay simultaneously. They are often more informative than free trials, although pricing and sales relationships may influence the result.

Basic conversion analysis

For a funnel with n visitors and c desired actions, the observed conversion rate is:

Here, p̂ is the sample conversion rate, c is the number of observed conversions, and n is the number of eligible visitors or prospects. The numerator and denominator must refer to the same population and time period.

For example, if 48 of 600 qualified visitors request a demonstration, p̂ = 48 ÷ 600 = 0.08, or 8%. This does not establish that the true conversion rate in the broader market is exactly 8%. It is a sample estimate subject to sampling error, selection effects, message effects, and implementation conditions.

When comparing two versions, founders should define the primary metric in advance and avoid repeatedly checking small samples until a preferred result appears. Random assignment, consistent traffic sources, adequate exposure, and a pre-specified stopping rule improve credibility. Statistical significance alone is not a substitute for managerial significance: a small uplift may be unimportant if acquisition costs remain unsustainable.

5. Minimum Viable Products

An MVP is the smallest version of a product or service that enables a venture to learn about a critical assumption from real or realistically engaged customers. “Minimum” refers to the minimum learning-oriented scope, not simply the cheapest or crudest product. “Viable” means that the experience is sufficiently coherent and ethical for the intended learning purpose; it does not mean that the venture is already profitable.

Choosing MVP fidelity

MVP fidelity should match the uncertainty being tested. A paper workflow may be sufficient to test task sequence. A clickable prototype may test comprehension and usability. A concierge service may test outcome value. A paid pilot may test economic exchange and organizational adoption.

Building too much before learning creates sunk-cost bias and delays discovery. Building too little can produce false rejection when the experience is so unreliable that customers cannot evaluate the underlying value proposition. The design question is therefore: what is the smallest credible intervention that allows the customer to experience the relevant benefit?

MVP quality and ethics

An MVP should not expose users to unreasonable safety, privacy, financial, or legal risks. A manually operated service must disclose material limitations where they affect informed consent, safety, or contractual expectations. Sensitive data should not be collected merely for convenience. The venture must define data retention, access controls, deletion procedures, and incident responsibilities before testing with real customers.

6. Lean Start-up Learning Cycles

The lean start-up approach organizes entrepreneurial learning through a build–measure–learn cycle. The cycle is not a license to launch poorly designed products. It is a disciplined method for making uncertainty visible and connecting action to learning.

1. Build: create an experiment or MVP aimed at a specific assumption.
2. Measure: collect behavioural and contextual evidence.
3. Learn: compare results with the pre-specified hypothesis and decide whether to persevere, iterate, or pivot.

A pivot is a substantive change in one or more elements of the venture model while retaining learning already acquired. Examples include a customer-segment pivot, problem pivot, channel pivot, revenue-model pivot, or technology pivot. A pivot should not be an unstructured reaction to disappointing results. It should identify what was learned, which assumption failed, and why the new direction is more plausible.

Evidence hierarchy

Evidence can be arranged roughly from weak to strong, although context matters:

1. Founder intuition and desk research.
2. Positive statements in interviews.
3. Repeated descriptions of a current problem.
4. Use of a prototype or request for access.
5. Time, data, reputation, or workflow committed by customers.
6. Payment, renewal, referral, or repeated use.
7. Economically repeatable acquisition and retention.

This hierarchy is not absolute. A regulated enterprise may require lengthy approval before payment, while a consumer may pay quickly without becoming a retained user. The central principle is to match evidence strength to the claim being made.

7. From Problem–Solution Fit to Product–Market Fit

Problem–solution fit exists when a clearly identified customer segment experiences a meaningful problem and the proposed solution appears to address it in a credible way. Evidence may include repeated problem narratives, strong workarounds, successful prototype use, and customer commitment to a pilot.

Product–market fit is a stronger and more sustained condition in which a product creates sufficient customer value that adoption, retention, referrals, and economic performance become repeatable within a defined market. It is not a single universal number or a one-time event. It is segment-specific and can be lost when markets, competitors, or customer needs change.

Important measures include:

• activation: whether customers reach the first meaningful value event;
• engagement: whether they use the product in the intended way;
• retention: whether they continue using it over time;
• churn: the rate at which customers stop using or paying;
• referral: whether satisfied customers recommend it;
• revenue and gross margin: whether value creation supports the business model; and
• acquisition efficiency: whether customers can be reached at sustainable cost.

For a cohort of customers, a simple retention rate at time t can be expressed as:

N₀ is the number of customers in the cohort at the starting point, and N_t is the number from that same cohort still active at time t. The measure is meaningful only when “active” is defined consistently and the cohort is tracked without replacing departing customers with new ones.

A venture may have high activation but low retention, indicating initial curiosity without durable value. It may have strong retention among a narrow segment but weak acquisition, indicating product value with an unresolved go-to-market problem. Managers should therefore interpret metrics as a system rather than celebrate a single headline measure.

8. An Integrated Validation Process

A practical validation process can be organized into seven stages:

Stage 1: State the opportunity

Describe the customer, problem situation, current alternative, desired outcome, and proposed value proposition without exaggeration.

Stage 2: Map assumptions

List assumptions across customer, problem, solution, demand, channel, economics, operations, and regulation. Rank them by importance and uncertainty.

Stage 3: Conduct discovery

Interview and observe relevant customers. Seek disconfirming evidence and document actual behaviour, not only opinions.

Stage 4: Select the smallest credible experiment

Choose a test whose design matches the assumption. Do not use a survey to answer a behavioural question that requires observation or payment.

Stage 5: Define metrics and thresholds

Specify the primary outcome, measurement period, inclusion criteria, and decision rule before collecting data.

Stage 6: Run and interpret

Separate raw observations from interpretations. Examine alternative explanations, segment differences, and implementation failures.

Stage 7: Decide and update

Persevere when critical assumptions are supported, iterate when the problem is real but the solution is weak, and pivot when a central assumption is contradicted or a more attractive opportunity emerges.

Applied Case: RouteWise Fleet Services

The following is an illustrative case designed for educational purposes.

Background

A founding team proposed RouteWise, a subscription service intended to help small urban delivery companies reduce failed deliveries. The founders assumed that dispatchers lacked adequate route visibility and would pay for predictive delivery-time alerts.

Problem

The initial assumption combined several issues: the problem might be failed deliveries rather than route planning; the user might be a dispatcher while the buyer might be the owner; and a software subscription might not be attractive if companies already used messaging applications and driver incentives.

Analysis and discovery

The team interviewed 22 people across eight delivery companies. Instead of beginning with the proposed software, they asked about the last failed delivery, its consequences, and the actions taken afterward. They found that failures were concentrated in apartment buildings with difficult access procedures. Dispatchers spent substantial time calling drivers and customers, but owners measured the problem primarily through refunds and repeat-customer complaints.

The team coded interviews by event frequency, cost, existing workaround, and authority. Four companies had already created manual building-access notes. This was stronger evidence than the general statement that “better routing would be useful.”

Experiment and MVP

The team built a concierge MVP. For six weeks, customers sent delivery schedules and building information through a secure form. A RouteWise employee prepared access instructions and flagged high-risk stops. The team measured completed deliveries, dispatcher time, and customer complaints. The service was offered as a paid pilot with a clearly defined scope and cancellation terms.

The pilot showed that the intervention reduced dispatcher follow-up time, but only when building information was current. Two companies renewed; one did not because its delivery volume was too low to justify the fee. The team learned that the highest-value segment was not all urban delivery companies but firms serving dense apartment districts with recurring routes.

Outcome

The founders did not immediately automate the full platform. They narrowed the segment, redesigned the value proposition around failed-delivery prevention, and tested a recurring data-maintenance service before investing in predictive routing technology. The next experiment focused on renewal and referral rather than clicks on a marketing page.

Learning points

1. The original solution framing concealed the more important operational problem.
2. Observation of workarounds revealed customer value more clearly than general enthusiasm.
3. A concierge MVP tested outcome value before expensive software development.
4. Paid pilots exposed volume and renewal constraints that free trials would have hidden.
5. Segment refinement was a productive pivot, not evidence that the entire opportunity had failed.

9. Challenges, Limitations, and Governance

Evidence-based validation reduces uncertainty but cannot eliminate it. Early customers may be atypical innovators. A founder may interpret a temporary trend as durable demand. Competitors may respond after launch. Regulatory approval may alter the product. Market size estimates may be unreliable because categories change as new solutions emerge.

Experiments also create risks of metric fixation. Teams may optimize clicks, sign-ups, or engagement while neglecting customer welfare, retention, profitability, or safety. A/B testing can identify which message performs better without explaining why or whether the message is honest.

Ethical validation requires:

• informed and voluntary participation;
• truthful claims about prototype capability;
• protection of personal and commercially sensitive data;
• special care with vulnerable populations;
• fair compensation where participants provide substantial time;
• transparent handling of deposits, cancellations, and failures; and
• avoidance of dark patterns that manipulate consent or payment.

In sectors such as health, finance, education, employment, and mobility, validation may involve regulatory, professional, and safety obligations before a conventional MVP can be tested. Governance should be designed into the experiment rather than treated as a later obstacle.

Technology and AI

AI can accelerate transcription, coding, segmentation, personalization, and prototype generation. However, automated analysis can reproduce sampling bias, misclassify meanings, expose confidential interview content, or create fabricated patterns. Entrepreneurs should obtain permission before uploading recordings, minimize personally identifiable information, verify machine-generated codes against source data, and preserve an audit trail of analytical decisions. AI-generated prototypes should not be represented as reliable production systems until their performance, security, explainability, and human oversight have been evaluated.

Best-Practice Checklist

Before declaring an assumption validated, ask:

• Is the target customer precisely defined?
• Is the evidence based on behaviour or merely stated preference?
• Was the experiment appropriate to the assumption?
• Were success thresholds defined in advance?
• Were negative cases actively sought?
• Could another explanation account for the result?
• Does the evidence persist across customers and time?
• Has willingness to pay or another meaningful commitment been demonstrated?
• Are retention, delivery, economics, and regulatory requirements addressed?
• Were customer rights, privacy, safety, and data governance protected?

Chapter Summary

Customer discovery and evidence-based validation are methods for converting entrepreneurial assumptions into testable hypotheses and learning before committing major resources. Strong discovery focuses on defined customer situations, recent behaviour, current workarounds, and consequences rather than hypothetical enthusiasm. Experiments should identify a critical assumption, expose it to observable customer behaviour, use a clear metric and threshold, and lead to a predefined decision.

MVPs are learning instruments whose fidelity should match the uncertainty being tested. Concierge services, prototypes, landing pages, paid pilots, and pre-orders each provide different kinds of evidence. The build–measure–learn cycle supports disciplined decisions to persevere, iterate, or pivot. Problem–solution fit concerns meaningful need and credible response; product–market fit requires sustained, repeatable value reflected in adoption, retention, referral, and viable economics.

Validation remains uncertain and ethically constrained. Founders must account for sampling bias, confirmation bias, weak proxies, regulatory duties, privacy, safety, and the limitations of AI-assisted research. The aim is not certainty, but progressively stronger evidence and better-informed resource allocation.

Key Terms

Assumption: A belief about customers, problems, solutions, or business conditions that has not yet been sufficiently tested.

Hypothesis: A specific, observable, and testable statement derived from an assumption.

Customer discovery: Systematic learning about customer contexts, behaviours, needs, and decision processes.

Problem situation: The particular context in which a customer experiences an undesirable outcome.

Workaround: An existing method customers use to manage a problem without the proposed solution.

Evidence: Information that supports, contradicts, or qualifies a venture hypothesis.

Experiment: A deliberately designed test connecting an intervention to measurable customer behaviour.

Minimum viable product (MVP): The smallest credible product or service used to generate learning about a critical assumption.

Concierge MVP: A manually delivered service that tests customer value before automation is built.

Wizard-of-Oz MVP: A customer-facing system that appears automated while key operations are performed manually.

Conversion rate: The proportion of eligible prospects who complete a defined action.

Problem–solution fit: Evidence that a meaningful customer problem exists and the proposed solution addresses it credibly.

Product–market fit: Sustained and repeatable customer value within a defined market, reflected in adoption, retention, and viable economics.

Activation: The point at which a customer first experiences the product’s intended value.

Retention: Continued use or subscription by customers over a specified period.

Pivot: A substantive change in venture direction based on learning while retaining relevant insights.

Confirmation bias: The tendency to seek or interpret evidence in ways that support existing beliefs.

Thematic saturation: The point at which additional qualitative inquiry produces few new relevant themes.

Leading indicator: An early measure that may signal future behaviour or outcomes.

Decision rule: A pre-specified criterion linking experimental results to an action.

Review Questions

5 Short Answer Questions

1. What is the difference between an assumption and a testable hypothesis?
2. Why are questions about recent behaviour generally more useful than questions about hypothetical future use?
3. Define a concierge MVP and identify one situation in which it is appropriate.
4. What is the difference between activation and retention?
5. Name two ethical risks associated with customer validation experiments.

5 Long Answer Questions

1. Explain how a venture team can construct and prioritize an assumption map. Include examples of high-importance and high-uncertainty assumptions.
2. Discuss the principles of effective customer discovery interviews and evaluate the limitations of interview evidence.
3. Compare landing-page experiments, concierge MVPs, and paid pilots in terms of the assumptions each can test and the evidence each produces.
4. Explain the build–measure–learn cycle and distinguish a disciplined pivot from an unstructured reaction to failure.
5. Critically evaluate the claim that product–market fit can be established by a single metric.

3 Analytical / Application Questions

1. A start-up receives 1,000 landing-page visitors and 70 demo requests, but only two of 15 pilot customers renew. Diagnose what the evidence suggests about attention, activation, value, and retention. Recommend the next experiment.
2. Design an interview and experiment plan for a proposed digital service for independent restaurants. State the target segment, three assumptions, interview questions, MVP type, primary metric, and decision rule.
3. A health-technology venture wants to test an AI symptom-assessment tool using real patient data. Develop a validation approach that addresses learning objectives alongside privacy, safety, informed consent, regulatory, and AI governance requirements.

Multiple Choice Questions with Answers

Question 1

Which statement is the strongest example of a testable customer hypothesis?

A. Small retailers need better technology.

B. Everyone will appreciate automated inventory management.

C. Independent grocery stores with two to five outlets experience weekly stock-outs and at least 30% of qualified owners will agree to a paid pilot.

D. Inventory software is a large and attractive market.

Correct Answer: C. Independent grocery stores with two to five outlets experience weekly stock-outs and at least 30% of qualified owners will agree to a paid pilot.

Question 2

Which interview question is most appropriate for discovering actual customer behaviour?

A. Would you use our proposed application?

B. What features should our product include?

C. How much would you pay for this idea?

D. Tell me about the last time this problem occurred and what you did next.

Correct Answer: D. Tell me about the last time this problem occurred and what you did next.

Question 3

What is the principal purpose of a concierge MVP?

A. To maximize automation before launch

B. To test customer value through manual delivery before investing in automation

C. To eliminate the need for customer interviews

D. To demonstrate that the venture is already profitable

Correct Answer: B. To test customer value through manual delivery before investing in automation

Question 4

A landing page receives 48 demo requests from 600 qualified visitors. What is the observed conversion rate?

A. 0.8%

B. 8%

C. 12.5%

D. 80%

Correct Answer: B. 8%

Question 5

Which evidence generally provides the strongest indication of willingness to pay?

A. A customer says the concept is interesting.

B. A customer clicks an online advertisement.

C. A customer pays for and renews a pilot.

D. A customer completes an unpaid survey.

Correct Answer: C. A customer pays for and renews a pilot.

Question 6

A team discovers that customers experience the stated problem, but the proposed solution does not improve outcomes. Which response is most appropriate?

A. Ignore the result because the market is large.

B. Continue scaling the solution unchanged.

C. Consider iterating or pivoting the solution while retaining the problem insight.

D. Declare product–market fit based on interview enthusiasm.

Correct Answer: C. Consider iterating or pivoting the solution while retaining the problem insight.

Question 7

If a cohort begins with 200 customers and 150 remain active after three months, what is the cohort retention rate?

A. 25%

B. 50%

C. 75%

D. 150%

Correct Answer: C. 75%

Question 8

Which practice most directly reduces confirmation bias during discovery?

A. Interviewing only people recommended by enthusiastic supporters

B. Recording and examining disconfirming evidence

C. Changing the hypothesis after every interview

D. Presenting the full product pitch at the beginning of every interview

Correct Answer: B. Recording and examining disconfirming evidence

Question 9

Why can high activation coexist with weak product–market fit?

A. Activation measures initial value, while retention may reveal that value is not durable.

B. Activation and retention always measure exactly the same behaviour.

C. High activation proves that customers will renew.

D. Retention is relevant only to physical products.

Correct Answer: A. Activation measures initial value, while retention may reveal that value is not durable.

Question 10

Which action is most consistent with ethical validation of an AI prototype using sensitive data?

A. Upload all recordings to any free AI tool for convenience.

B. Collect more personal data than the experiment requires.

C. Obtain appropriate permission, minimize data, verify outputs, and maintain human oversight.

D. Describe prototype outputs as clinically reliable before testing them.

Correct Answer: C. Obtain appropriate permission, minimize data, verify outputs, and maintain human oversight.