Average Is Not a Customer

The case for decision-specific intelligence

The average small business has no storefront you can visit, no owner you can interview, and no invoice you can inspect.

It has a mean headcount, a median revenue, an industry code, a probability of seeking credit, and a place in a market-size slide. These are indispensable for understanding populations. They become dangerous when mistaken for a person.

A venture does not serve a total addressable market. It serves a sequence of situations. A policy does not reach a category. It asks an individual firm to notice, trust, qualify, apply, implement, and persist.

Between population evidence and individual action lies the work of fit.

From a market to a situation

A market becomes real only when a population-level condition resolves into a situation in which someone can act. Industry, size, and geography narrow the average, but intent, timing, constraints, and the surrounding work determine whether the apparent need is demand.

Founders are trained to move from a large number toward a plausible share: millions of small businesses, billions in spending, a percentage captured. The arithmetic demonstrates scale. It does not demonstrate demand.

Thirty-four million U.S. small businesses do not share one problem, budget, buying process, or reason to change. Even a narrow segment such as “microbusinesses needing better market intelligence” contains firms with different products, data, skills, cash cycles, and ambitions.

The Federal Reserve's 2026 “Firms in Focus” work reinforces the value of disaggregation. Financing experiences, confidence, and operating conditions vary by firm characteristics and owner demographics. The differences are not decorative. They shape what an intervention costs and whether it is trusted.

A market analysis becomes useful when it explains a mechanism. Not “many firms struggle with growth,” but “firms with irregular project revenue spend significant owner time screening opportunities because eligibility and fit data are separated.” The second statement can be tested. It identifies a user, trigger, current workaround, and burden.

Segmentation improves averages by forming smaller averages. Industry, size, age, geography, and owner characteristics can reveal meaningful patterns. Yet two firms in the same cell can still need opposite recommendations.

Consider two illustrative seven-person contractors in the same city. Both have similar revenue and certifications. One has a stable private-client base, limited working capital, and no interest in government work. The other has relevant past performance, a line of credit, and an explicit strategy to diversify into public contracts. A procurement alert has the same formal relevance to both and radically different value.

The missing variables are intent, readiness, relationships, constraints, and timing. These are dynamic. They cannot be fully inferred from administrative data. They must be learned with the organization and corrected over time.

This is why personalization is not a one-time profile. A profile says what the firm is. A decision context says what the firm is trying to do now, what it can risk, and what it learned last time.

The language of a “job to be done” improved product thinking because it shifted attention from demographic identity to the progress a person is trying to make. A drill is hired to create a hole; a financing product is hired to bridge a timing gap. For small firms, however, the focal job is often embedded in several surrounding jobs.

Take an illustrative café buying a new oven. The stated job is equipment acquisition. Around it sit site inspection, electrical capacity, permitting, installation downtime, staff training, menu changes, supplier terms, and enough working capital to survive the transition. A loan that funds the invoice but ignores the surrounding jobs can be perfectly matched to the category and poorly matched to the decision.

This is why needs assessments often produce familiar nouns—capital, customers, skills, technology—without revealing a usable market. The noun is the center of a system. Demand becomes concrete when we see the event that triggered it, the dependencies it carries, the workaround used today, the cost of failure, and the person authorized to change the process.

The venture question is not “How many cafés need equipment?” It is “How many firms face an equipment transition with enough expected value to act, insufficient internal capacity to coordinate it, and a willingness to trust an outside integrator?” That market is smaller. Its economics are visible.

Researchers often prefer behavior to stated preference. A purchase is stronger evidence than an expression of interest. Yet observed behavior occurs inside the current set of constraints. It reveals what people did, not all that they wanted.

A business that repeatedly uses a credit card may appear to prefer revolving credit. It may simply lack access to a line of credit at the right size. A supplier that does not bid on public work may appear uninterested; it may have learned that payment timing would be dangerous. A shop that buys from an expensive nearby wholesaler may value convenience, or it may lack the storage, transport, and cash required to use the cheaper source.

Market analysis must therefore separate preference from feasible choice. Economists call this attention to constraints; product teams might call it understanding the workaround. Either way, the implication is profound. A successful incumbent behavior can be evidence of satisfaction, resignation, or adaptation.

The distinction can be tested. Ask what the owner considered, what ruled alternatives out, and what would have to change. Observe whether behavior changes when a constraint is temporarily removed. Offer an assisted pathway before building a platform. If businesses choose differently when transport, paperwork, or working capital is supplied, the original behavior was not a pure preference.

This is also why survey questions such as “Would you use?” inflate markets. They remove price, implementation, trust, and competing priorities from the decision. The useful experiment restores those conditions early.

Adoption lives in a workflow

Technology makes this distinction especially clear. Counting tools says little about whether information, authority, and practice changed enough for the old burden to disappear.

Technology policy frequently begins with adoption rates. The April evidence on SME technology adoption in the United Kingdom pointed toward differences by firm size, region, sector, management, skills, and perceived relevance. The OECD's 2026 study emphasized advisory support, peer learning, and intermediaries alongside financial incentives.

The adoption gap is real. But “adopt more technology” is not a business outcome.

An owner does not need AI in the abstract. She may need to reduce proposal reconstruction, identify late-payment risk, schedule crews, preserve project knowledge, or answer recurring customer questions. The appropriate tool, data, risk, and training depend on the use case.

U.K. evidence reported difficulty identifying activities or business use cases among the leading barriers to AI adoption. This is a revealing reversal. The market often starts with what the technology can do and asks the firm to find a problem. The firm starts with a crowded week and asks whether change will make it less crowded.

The practical intermediary is not a salesperson for adoption. It is a translator between operating pain and a governed intervention.

Technology adoption statistics count whether a firm uses cloud computing, e-commerce, AI, or enterprise software. The categories are useful for observing diffusion. They do not show whether a business changed the workflow that creates value.

A firm can “use AI” because one employee drafts occasional messages. Another may integrate a model into customer service with approved sources, human review, outcome monitoring, and a process for correction. Both answer yes. Their organizational exposure and productivity effects are incomparable.

The same problem appears with e-commerce. A storefront can accept online orders while staff manually re-enter them into inventory and accounting. Digital demand rises, administrative work rises, and the dashboard celebrates adoption. A scheduling tool can reduce telephone calls but fail when staff availability is not kept current. The interface is digital; the operating system remains fragmented.

The correct unit is the workflow: where information begins, how a decision is made, who acts, what exception occurs, and whether the outcome returns as learning. Adoption is complete when the old burden is retired or deliberately retained, not when the account is created.

This reframes the technology market. The buyer is not purchasing software alone. The buyer is purchasing a transition from one reliable-enough process to another. Vendors who price only the license leave implementation to a customer with no implementation department. Advisers who promote generic adoption can count uptake while firms absorb the failure privately.

The recommendation engine envisaged by CKOS should learn at this level. “You use accounting software” is weak context. “Customer invoices are issued weekly from completed-job sheets, and late updates delay cash” can support a decision.

Consider two illustrative manufacturers with similar employment, revenue, and machinery. Both appear eligible for a grant supporting automation.

The first produces a stable family of parts in predictable volume. Its bottleneck is a repetitive inspection step. Process data is consistent, the operator understands variation, and a customer has indicated additional demand if lead time falls. The grant could accelerate a project already conceptually ready.

The second is a job shop whose advantage is handling irregular, low-volume work. Schedules change daily. Product data is inconsistent because customers send drawings in different forms. The owner is attracted to automation but cannot identify a repeated step whose redesign would pay back. The same equipment may create an island of technical capability surrounded by manual exceptions.

A demographic match treats the firms alike. A decision-specific analysis asks different questions: Is variation signal or waste? Is demand constrained by capacity? Who will integrate the equipment? What process knowledge has been documented? What happens during downtime? How much of the promised productivity depends on orders not yet secured?

The correct recommendation to the second firm may be “not yet.” That answer is economically valuable even if it reduces grant uptake and vendor leads. It may redirect the firm toward process mapping, data standardization, or a smaller pilot. Personalization proves its loyalty when it can advise against the transaction that funds the platform.

A recommendation as a learning relationship

A useful recommendation must therefore do more than match attributes. It should explain itself, price the pursuit in the firm's actual circumstances, learn from refusal, and earn enough trust for the owner to correct what the system thinks it knows.

Imagine an opportunity service that sends a grant to two firms. The first recommendation reads: “92% match.” The second reads:

This grant supports equipment purchases in your county. Your planned packaging upgrade appears eligible, and your prior energy audit may satisfy part of the evidence. The match is uncertain because the guidance excludes routine replacement; confirm whether the new line creates a measurable efficiency gain. The application is estimated at eight owner hours plus a vendor quote. Deadline: 21 days.

The second message is longer and more honest. It connects external rules to internal context, identifies uncertainty, and prices the pursuit in attention. It gives the owner something to disagree with.

Explainability here is not a technical feature. It is the foundation of agency. A recommendation that cannot show why it fits cannot teach the system when it is wrong. The user's corrections are not friction to eliminate; they are the mechanism through which “average” becomes “this business.”

The service should remember the outcome. Did the firm pursue? Why or why not? Was the application eligible? Did the estimated effort resemble reality? A static profile predicts from attributes. A learning relationship improves from decisions.

A relevant opportunity at the wrong moment is irrelevant. This seems obvious, yet many support systems treat deadlines as properties of the program rather than conditions of the customer.

The same restaurant may have capacity to redesign its website in January and none during the holiday season. A contractor may be financially healthy in annual accounts but temporarily extended across three mobilizations. A manufacturer may need finance after receiving a purchase order, not during a general loan campaign. A founder caring for a parent may rationally postpone expansion without changing ambition.

Timing also changes trust. An unsolicited offer of expensive finance during cash stress can feel predatory even if technically suitable. Advice delivered after a failed application can sound like judgment; delivered before it, the same advice is preparation. A network introduction has greatest value when both parties have a reason to act.

Decision-specific intelligence must therefore maintain a living clock: planned events, seasonal load, expiring documents, payment cycles, prior attempts, and windows when change is feasible. This does not require constant surveillance. Much can be declared by the owner and corrected through use. The discipline is to treat time as part of fit.

Static databases decay because firms change. A learning system decays too, unless it knows which facts are durable and which require confirmation. Industry may be stable. Available cash is not. Strategic intent can change after one difficult contract. Confidence scores should fall with age, and the system should say when it is relying on stale context.

Most marketplaces learn from clicks, applications, and purchases because those events are easy to record and monetize. A declined recommendation is often treated as absence. Yet the reason for “no” can be the most valuable fact in the system.

Not eligible, no time, too much cash up front, wrong geography, distrust of provider, prior bad experience, not a priority, already solved: these are different market signals. Repeated across firms, they can reveal that the opportunity is poorly designed rather than poorly marketed.

The interaction must be light enough that explaining does not become another capacity tax. A quick reason can be followed, selectively, by an interview when the pattern matters. The system should also distinguish “not now” from “never” and let the owner erase an inference.

Negative evidence protects against a dangerous feedback loop. If the platform shows what attracts engagement and suppresses what does not, it can gradually confuse its own promotion with demand. Opportunities requiring more explanation may disappear even if their ultimate value is higher. Firms already fluent in the interface generate the signals that shape the interface, while excluded users become statistically silent.

The refusal is therefore not a failed conversion. It is part of the market's reply.

A new intelligence service knows little about a firm. It can ask a long questionnaire, infer from external data, or offer generic recommendations while learning. Each choice imposes a cost.

Questionnaires promise personalization later in exchange for effort now. Owners have learned to be skeptical. They do not know which answers matter, whether the benefit will arrive, or how the data will be used. External inference lowers effort but can be wrong in consequential and invisible ways. Generic recommendations reproduce the noise the service claims to solve.

A better cold start begins with one real decision. The service asks only what changes the answer, explains why each fact matters, and returns a useful analysis immediately. Over time, accepted corrections and observed outcomes build a profile. Trust is earned through a sequence of bounded performances.

This resembles relationship banking at its best. Soft information accumulates because repeated interactions have purpose. The lender does not merely possess more facts; it interprets behavior in context. A digital system can preserve and retrieve context, but it must also create recourse. When the recommendation is wrong, can the owner understand it, correct it, and reach a person if the consequence matters?

Cold start is often framed as a machine-learning problem. For a small-business institution, it is the question of why anyone should teach the system enough to become useful.

The variables that resist a profile

Some of the most consequential variables remain relational and difficult to observe. The attempt to capture them can improve relevance, but it can also convert vulnerability and network inequality into a more efficient form of extraction.

Some of the strongest determinants of action are poorly represented in data: a trusted adviser, a buyer introduction, a peer's warning, confidence built through prior experience, or the knowledge that someone will help if the process goes wrong.

OECD research on social capital and SMEs treats networks and intermediary institutions as part of the enabling environment. Their effect is easy to understate because it appears through other outcomes. A lender application succeeds; a supplier meets a buyer; a technology is adopted. The relational infrastructure disappears from the record.

This matters for algorithmic recommendations. Two firms with identical observable attributes may have different practical access because one has a navigator. If the system learns only from historical success, it may infer that the successful firm's attributes caused the outcome and reproduce the network's exclusion.

Good intelligence should therefore distinguish capability from access. It should be able to say, “You appear qualified, but this pathway usually depends on a reference, certification, or relationship you do not yet have.” That is less flattering than a high match score and more actionable.

The deeper a system understands a business, the more valuable—and vulnerable—the data becomes. Goals, constraints, pricing logic, relationships, unsuccessful bids, and cash concerns can improve recommendations. They can also be used to sell, rank, or exploit.

A platform funded by lead sellers may be tempted to define relevance as what a provider wants to promote. A lender with access to operating context may price risk more accurately or identify desperation. A marketplace can use demand data to compete with its own participants.

The design principle must be loyalty: whose outcome is the system optimizing when interests diverge?

User control requires more than a privacy notice. The firm should see what the system believes, why it matters, where it came from, who can access it, and how long it persists. Sensitive context should be purpose-bound. Correction must change future behavior, not merely edit a decorative profile.

The better the personalization, the stronger the governance must become.

Market research often removes outliers to find the pattern. Design research should sometimes begin with them.

The owner who declines a seemingly perfect grant may reveal a hidden cash requirement. The firm that refuses a free tool may reveal implementation burden. The business that thrives in a weak local market may reveal a relationship or route absent from the model. Outliers expose variables the average omitted.

This does not mean building for every exception. It means treating surprising behavior as evidence before labeling it irrational. The question “What would have to be true for this decision to make sense?” is one of the most productive in venture research.

The result may narrow the market. That is a success. A smaller group with a repeated, consequential problem is a better basis for an experiment than a vast category united by a noun.

Synthesis without market theatre

Market synthesis earns credibility by joining these levels without collapsing them: the broad pattern, the decision mechanism, the actual friction of adoption, the economics of service, and the evidence that would prove the premise wrong.

A credible market analysis should make it possible to be disappointed. It should specify the group, triggering situation, current alternative, cost, willingness and ability to pay, pathway to adoption, and observable outcome. It should also say which evidence is inferred and what would falsify the premise.

Large top-down numbers still have a role. They show the outer boundary and help estimate whether a narrow mechanism can support an organization. But the analysis must build upward from cases. How frequently does the trigger occur? How expensive is the workaround? Who captures the benefit? Can the buyer authorize a purchase? Does serving the case require so much custom work that the venture cannot sustain itself?

This is where qualitative depth and quantitative discipline meet. Interviews reveal mechanisms and language. Administrative data estimates prevalence. Concierge experiments reveal behavior under real friction. Cohorts show retention and heterogeneity. Unit economics prevent a socially resonant problem from being confused with a viable service.

The strongest synthesis does not announce that demand exists. It shows the chain by which a condition becomes a decision, the decision becomes a purchase or participation, and the intervention becomes an outcome worth paying for.

The age of abundant intelligence creates a new scarcity: the confidence that a recommendation is about this situation.

More data can improve relevance, but only if the system understands the limits of each fact. Industry is not strategy. Revenue is not liquidity. certification is not readiness. A prior decision is not a permanent preference.

The April argument is simple: average is a lens, not a customer. Market analysis should move from population pattern to decision mechanism; product design should move from profile to learning relationship; governance should move from consent once to control over time.

Would an owner rather receive one opportunity that explains itself or fifty that technically match? The answer is a venture premise. Whether people will trust such a system enough to teach it is the experiment.

Research lineage

Observed evidence. Technology and financing experiences vary across firm size, sector, geography, management, and owner characteristics. Use-case clarity, skills, and trusted intermediaries shape adoption.

Interpretation. Segments remain incomplete until intent, readiness, constraints, relationships, and timing enter the decision context.

Hypothesis. Transparent, correctable recommendations grounded in a living account of the firm can outperform volume-based alerts and generic support.

Questions carried forward. What is the minimum context needed for relevance? How quickly can a system learn fit? What governance makes deep personalization worthy of trust?

Sources and further reading

1. Federal Reserve Banks, Small Business Credit Survey, 2026 firm-level reports and methodology.
2. OECD, SME Technology Adoption in the United Kingdom, April 22, 2026.
3. U.K. Office for National Statistics, “Management practices and the adoption of technology and AI in UK firms”, March 24, 2025.
4. OECD, SMEs and entrepreneurship, research through April 2026.
5. U.S. Census Bureau, Business Trends and Outlook Survey, April 2026.

Evidence cutoff: April 30, 2026. The contractor and grant examples are illustrative composites.