Chapter 16: From Knowledge to Capability

Demonstrate a bounded task on a fresh input, distinguish your judgment from the assistance you use, and identify the evidence needed before you delegate more responsibility.

The next claim needs a demonstration

Audy has accumulated a substantial working set. She can find the current instruction, explain its history, retrieve a relevant verdict, and recover a small package in a separate working environment. Those records make her work easier to continue. They also create a temptation: to describe the collection as proof that she can now deliver a reliable service.

The next step is smaller and more useful. She needs to show what she can do with those materials on a fresh task. Can she turn unfamiliar notes into a faithful example action list? Can she identify an error before it enters the workshop brief? Can she explain why the correction is appropriate? Can she distinguish a limit in her procedure from a gap in her understanding?

The System arc begins with these questions because a delegated workflow needs something definite to accomplish. A capable operator can name its intended result and recognize at least the important ways it might fail. Without that judgment, adding an agent would mainly increase the number of outputs Audy must trust without understanding.

Himura Inc. remains a fictional prospective business. The immediate demonstration concerns preparing and reviewing workshop-design material. Commercial demand and learner performance will require different evidence.

Open your capability baseline from Chapter 1. Compare the operation you described then with the one you can now attempt. Do not begin by changing your maturity label. Begin by choosing an example that will let another person see the difference.

State the capability as an operation

“Good at AI” is difficult to demonstrate. “Able to prepare a reviewed action list from supplied notes without inventing agreements, owners, or dates” points to a task, an input, a standard, and a result. You can inspect whether the claim is supported.

Audy adds the conditions. She uses her saved instructions and an assistant for drafting. She performs the final source comparison herself. When the notes contain an unresolved conflict, the output must preserve it and request clarification. The capability does not include obtaining that clarification or contacting a participant.

These boundaries make the claim stronger because they make it testable. A reviewer can check a supplied packet rather than debate whether Audy has become a generally capable founder.

Use a verb that corresponds to work someone could inspect: identify, compare, prepare, diagnose, revise, explain, or operate. Then name what the verb acts on. “Understand research” can become “compare a proposed conclusion with the methods and results actually reported in an eligible paper.” “Build systems” can become “restart this bounded workflow from its saved inputs and recover from a missing dependency.”

Include the standard without hiding it inside an adjective. “High quality” may mean accuracy, clarity, completeness, speed, or suitability for a particular person. If you require all of them, say how each will be judged. An attractive output can satisfy presentation requirements while failing the task that made it worth producing.

The capability demonstration starts with this statement. It is a record of evidence, not a numerical examination or an additional maturity framework. Its fields help you make a claim someone else can question.

Distinguish judgment from access to help

A professional can be capable while using references, colleagues, calculators, and software. The relevant question is which parts of the performance depend on those aids and which judgments the person must exercise. Pretending that all work should be unaided would make the demonstration less representative of actual practice.

Audy does not need to memorize every version identifier or research reference. She does need to recognize why a suggestion is not an agreement, why a source citation can point to a sentence that does not support the claim, and why a date conflict cannot be resolved from presentation order alone. These judgments guide the use of her external records.

She therefore records the conditions of each attempt. One short explanation is written before opening the files. A task attempt uses the authorized working packet and the assistant. A review checks the output against the actual input. These observations answer different questions and can all be valuable.

The unaided explanation might reveal a gap that the assisted performance conceals. Suppose Audy says the rule is “leave uncertain fields blank.” Her records are more precise: preserve useful supplied information, identify what remains unresolved, and distinguish missing details from conflicting claims. She can correct her explanation before relying on it to judge a difficult case.

Rozenblit and Keil’s experiments showed how attempting a detailed explanation could expose limits in people’s earlier ratings of their explanatory understanding. The study concerned judgments about explanation, not AI-assisted business performance. It offers a reason to test an explanation through the act of producing it; the demonstration used here is this book’s adaptation. (Rozenblit & Keil, 2002)

Do not make the unaided part a contest in recalling irrelevant details. Choose the distinctions you will need when a tool gives a plausible wrong answer. If you can retrieve a fact reliably and inspect its source, retrieval may be the appropriate skill. If you cannot recognize when the fact matters, additional access alone will not close that gap.

Choose a fresh input for a specific reason

A fresh input should challenge a meaningful relationship, not merely change the names. Audy’s earlier examples separately examined suggestions, missing owners, vague timing, and conflicting dates. The new packet combines those conditions in a short set of notes.

Its first item supplies an agreed action, named owner, and explicit Tuesday deadline. A second item proposes asking a library about a larger room but assigns nobody. Two passages give Thursday and Friday for the same handout, with no established supersession. A final item agrees to shorten the opening discussion without assigning an owner or date.

A useful result must preserve all of those distinctions together. The first row cannot become uncertain just because later rows are uncertain. The suggestion cannot become an agreement because the table would look more complete with an owner. The conflicting dates cannot be reduced to one confirmed date. The unassigned agreement must remain visible.

The literal C16 input lets you inspect the combination. These are constructed teaching materials. When performing your own demonstration, keep the expected properties separate from the input you supply to the assistant. Otherwise the answer may already be present in the working packet.

The change in conditions should match the claim you want to make. A new topic may test whether the same operation works on unfamiliar content. A different document format may test whether the procedure depends on a familiar layout. An incomplete source may test whether the workflow knows when to stop. Changing all three at once can expose a failure while making its cause harder to identify.

Research on transfer cautions against treating a skill’s movement across contexts as one simple quantity. Barnett and Ceci distinguish several dimensions of transfer and describe evidence that varies across conditions. Their review does not establish that this specific workshop capability transfers; it supports being explicit about which change a demonstration actually tests. (Barnett & Ceci, 2002)

Begin with a change you can interpret. Preserve the old cases as checks on behavior you intend to retain. If the fresh case reveals a problem, you can then investigate the affected relationship without losing the baseline that made the difference visible.

Write the expected properties before the response

An expected result need not be one exact answer. Audy’s criteria allow faithful paraphrases. They require the same underlying agreements, suggestions, owners, timing, and uncertainty to survive. The wording can differ while the meaning remains inspectable.

She writes those properties before asking for a draft. This gives her review an anchor independent of the candidate’s confidence or organization. If a response proposes a better structure, she can adopt it while keeping the source requirements intact.

For each item, she identifies the relevant lines. This small preparation step makes the later comparison possible. A source number without an inspected supporting passage would only decorate the answer. The reviewer must still ask whether the cited words establish the claimed relationship.

Write at least one unacceptable result that might look reasonable. Audy’s packet anticipates an invented owner for the library inquiry and an unsupported choice between the two handout dates. Another reader might anticipate a summary that omits a consequential exception or a program that returns a result from the wrong input file.

Include a stopping condition. If the required source is missing, the demonstration should show a hold on source-dependent work. If the available instruction does not cover an important new authority question, preserve that gap. Producing a plausible answer at any cost would demonstrate a different behavior from the one you intend to claim.

Keep the expected material in the review area. A drafting assistant can receive the criteria without receiving a prewritten answer to the exact task. For an independent check, record what it actually received. The distinction between an instruction and leaked evaluation material matters even in a small manual exercise.

Inspect the first candidate before repairing it

The constructed first candidate in the artifact is nearly right. It preserves the Tuesday slides deadline, leaves the handout dates unresolved, and keeps the unassigned shorter discussion. Its second row, however, treats the library inquiry as agreed work and assigns the coordinator to it.

The cited line does not support either change. It says the group suggested the inquiry and nobody accepted the task. The coordinator appears in a different source item about uploading slides. The candidate has carried a legitimate owner into the wrong relationship.

Audy rejects that row and retains the supported work. This is a useful judgment because it identifies the exact defect rather than dismissing the whole output or accepting it because the conflict row looks sophisticated. A good result on one difficult-looking part does not settle the remaining parts.

She then inspects the instruction that was supplied. It is the approved instructions-v3. It already says to distinguish agreements from suggestions and use owners only when supported. The diagnosis therefore differs from Chapter 7’s incomplete migration. Here the relevant rule is present and the candidate violates it.

That evidence does not reveal an internal cause in a real model. Audy should not assert that the system “forgot” the rule or “reasoned” in a particular way merely because the output is wrong. She can identify what was supplied, what was produced, and which requirement failed. A causal explanation may require another comparison.

Preserve the first candidate in your actual demonstration. Replacing it with the corrected table removes evidence about the work needed to reach acceptance. It also makes later claims about speed, review burden, and reliability harder to assess.

Choose a repair that matches the evidence

Correcting the row fixes this candidate. Improving the procedure is a separate proposal. Audy considers making source extraction visible before formatting: identify each statement’s status and supported details, then assemble the action list from that inspected material.

The proposed change may help a reviewer notice a borrowed owner. It may also add time or introduce another place where information can be lost. She should try it on a new input and inspect the result before calling it an improvement.

She does not need another version of the core evidence rule, which already states the required behavior. A working procedure can change while the task’s meaning stays the same. Record which level you are changing: the criterion, the supplied context, the order of operations, the review, or the tool.

If you alter several of these at once, be careful about attribution. A better result after changing the model, prompt, examples, and reviewer could justify retaining the combined procedure under its tested conditions. It would not tell you which change caused the difference. Decide whether that distinction matters to the next decision.

For a low-cost personal task, a practical comparison may be sufficient. For a consequential or frequently repeated task, you may need more cases, an independent reviewer, or a specialist’s standard. The amount of evidence should reflect what you intend to rely on, including the effects of a missed error.

The aim is not to make every experiment formal. It is to avoid hiding the gap between a repaired answer and a more dependable way of producing answers. That gap becomes important when you begin delegating a workflow and cannot inspect every intermediate step at the same depth.

Look for a boundary, not a universal verdict on AI

A tool can help substantially on one task and mislead on another that appears nearby. That makes task selection and verification part of the capability you are developing. Asking whether AI is generally “ready” does not settle whether this procedure works on your inputs under your standard.

In a field experiment with consultants, Dell’Acqua and colleagues found benefits on tasks within the studied AI capability frontier and worse correctness on a selected task outside it. The published study concerns specific participants, tasks, and technology; its boundary is not a permanent map of today’s tools. Its practical relevance here is the need to examine a plausible failure case alongside the task you expect to benefit. (Dell’Acqua et al., 2026)

Audy’s synthetic task has its own provisional boundary. The instructions cover supplied commitments, suggestions, missing details, vague timing, and unresolved deadline conflicts. They have not established a complete procedure for interpreting an explicitly authorized amendment or deciding whether a real person had authority to make it. Those conditions require further work.

A useful boundary statement identifies what to do when the input reaches the edge. Preserve the evidence, identify the unresolved question, and route it to the responsible person. A workflow can be useful while holding cases outside its supported scope. Its value depends partly on how accurately it recognizes those cases.

Do not turn a failure into a universal conclusion either. The invented library assignment does not show that every assisted action list is unusable. It shows that this constructed candidate requires correction and that the proposed procedure needs evaluation.

Keep examples of both acceptable and unacceptable behavior in the review set. A collection containing only obvious failures can make a strict filter appear effective while it rejects useful work. A collection containing only easy successes can conceal the conditions in which review is most needed.

Count the work needed to reach an acceptable result

A demonstration that records only generation time leaves out much of the operation. Audy had to assemble the input, retrieve the instruction, write expected properties, inspect the candidate, correct a row, and decide what to test next. Those activities are part of the effort required to produce a reviewed example.

Record preparation, production, review, and repair separately when effort matters. A reusable instruction may reduce preparation after its initial creation. A new type of input may increase review. Keeping those parts visible helps you identify where the next improvement would matter.

Compare tasks under the same acceptance standard. If the manual attempt requires a source-faithful result while the assisted attempt is timed before review, the comparison answers an unhelpful question. A faster unreviewed draft and a slower accepted output are different products of work.

Also record who performed the judgment. An experienced reviewer may catch a defect a beginner misses. A workflow that depends on that expertise should name it rather than credit every successful correction to the assistant. The practical question is whether you can supply that review consistently at the scale you intend to operate.

Do not infer a commercial margin from this exercise. The time needed to prepare synthetic materials excludes acquiring a customer, understanding a real brief, obtaining appropriate inputs, agreeing scope, and supporting a delivery. Later chapters will connect capability with those business operations. For now, the record can reveal an internal bottleneck without pretending to price a complete service.

A slower attempt can still be useful during learning. Perhaps Audy now understands why the owner was unsupported and can diagnose the same defect elsewhere. Record that learning outcome separately from an efficiency claim. You need not force every valuable exercise to prove immediate time savings.

Make the demonstration legible to another person

Imagine handing the evidence to a colleague who has not followed the book. They should be able to locate the task statement, input, supplied instruction, first output, review, and corrected candidate. They should also know which parts are constructed examples and which parts are your actual recorded work.

The record need not include every conversation. Preserve enough to reconstruct the consequential choices. If a missing attachment determined the result, include an authorized copy or explain that the reviewer cannot independently inspect it. A long transcript does not compensate for an absent source.

Ask the reviewer to challenge one conclusion. They might notice that a source line permits more than one interpretation, that the expected answer assumes an unstated owner, or that an apparent improvement came from providing the answer during drafting. Their challenge can improve the demonstration even if the final candidate remains unchanged.

If they disagree with your verdict, return to the criterion and source. Do not treat agreement with you as the acceptance standard. The earlier verdict practice provides a way to record and resolve such disagreements, including correcting your own judgment when the evidence requires it.

A demonstration becomes useful beyond the immediate task when its limits are as readable as its successes. Another person can decide whether the operation fits their situation or which additional test they would need. That is a stronger basis for collaboration than an unexplained claim of expertise.

Choose the next responsibility you can justify

The current TFIS homepage describes a progression from operating tasks with AI assistance to orchestrating workflows and supervising agents. Those labels direct attention to changes in responsibility. This book asks you to support such changes with work another person can inspect; its exercises are not official certification. (The Future Is Solo, n.d.-b)

Audy’s demonstration does not justify unattended delivery. It can justify a more structured rehearsal: let an assistant carry a bounded sequence from supplied input to a review packet, while Audy retains the final judgment. The next chapter will make that delegation concrete through a task contract and defined stopping conditions.

Readiness for that step concerns more than one correct answer. Can you specify what the workflow may read and produce? Can you tell when the required input is missing? Can you review its output without reconstructing the whole task? Can you stop or repair an attempt that violates a requirement?

If one of those abilities is missing, practice it directly. You may need a clearer source packet, a smaller task, or a better review example. Adding more agents would not necessarily help. A useful next step should address the observed gap rather than resemble the next rung in a diagram.

You may also choose to keep a task human-led. If it happens rarely, depends on difficult judgment, or takes less effort than maintaining a delegated procedure, that can be a sensible operating decision. Moving up the ladder should expand your choices about how to work, not require automation for its own sake.

Choose between three possible next tasks

Suppose Audy must decide what to attempt tomorrow. She could prepare another example herself with assistance, rehearse a delegated sequence, or promise an unattended service. Each choice places a different demand on the evidence she has just produced. The useful comparison names that demand rather than assigning her a broad confidence score.

The second row is the productive stretch for the next chapter. Audy need not wait until she can support the third row before learning how delegation behaves. A rehearsal creates room to discover a missing input or a bad handoff while the result is still an internal candidate. That narrower next step has a clear purpose: produce the evidence that the present demonstration lacks.

Notice that the first row remains useful work. Repeating a human-reviewed task with different inputs may reveal whether the rule violation recurs and how much correction costs. If review consumes most of the time, Audy should investigate that cost before adding coordination. If the task cannot yet be specified clearly, drafting the contract may expose the ambiguity before any agent runs. In either case, the next responsibility follows the observed gap.

Demonstrate capability in your own field

A service professional could demonstrate preparing a client-ready scope from authorized intake notes. The capability would include distinguishing a client’s request from a confirmed commitment, preserving unanswered questions, and identifying dependencies before promising a date. The reviewed output would remain a draft until the responsible person authorized its external use.

An educator could demonstrate designing an activity that produces a learner attempt and giving feedback tied to the actual response. The initial evidence might concern design and feedback quality. A separate learning claim would require appropriate learner evidence. Keeping these claims distinct helps the educator improve one operation without overstating another.

A software builder could demonstrate reproducing an error, making a bounded repair, and showing that the relevant behavior changed under a meaningful check. The evidence should include the actual starting condition and resulting behavior. A generated explanation of a fix is not the same as a program that now performs the task.

In each setting, choose a fresh case that challenges the claimed operation. Name the assistance used, preserve the first result, and inspect a plausible defect. The details will differ, but a reviewer should be able to connect the claimed capability with the evidence offered for it.

Field assignment: earn one precise claim

Complete the capability demonstration for one recurring operation. State what you will produce, for whom, from which input, and under what review conditions. Identify the distinctions you should be able to explain without the assistant, then make that explanation before opening the records.

Choose a fresh input for a reason you can state. Prepare the expected properties separately, including one legitimate alternative and one plausible failure. Supply the permitted working materials, retain the first response, and compare each consequential part with the source.

Record the defect if one appears. Inspect the actual instruction and input before deciding whether the problem concerns missing knowledge, an incomplete handoff, or violation of a present requirement. Correct the candidate, then make a separate proposal for improving the process if the evidence suggests one.

Ask another reader to inspect the chain of evidence when possible. Revise your capability statement to match what the demonstration actually supports. Include the help used and the next condition you have not tested. If the attempt fails, the resulting boundary and repair can still be valuable progress.

Audy can now state a specific next responsibility: rehearse a bounded assisted workflow that produces a source-linked draft and review packet, with her final judgment retained. Chapter 17 turns that responsibility into a delegation contract someone—or an AI agent—can follow.

References

Barnett, S. M., & Ceci, S. J. (2002). When and where do we apply what we learn? A taxonomy for far transfer. Psychological Bulletin, 128(4), 612–637. https://doi.org/10.1037/0033-2909.128.4.612

Dell’Acqua, F., McFowland, E., III, Mollick, E., Lifshitz, H., Kellogg, K. C., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2026). Navigating the jagged technological frontier: Field experimental evidence of the effects of artificial intelligence on knowledge worker productivity and quality. Organization Science, 37(2), 403–423. https://doi.org/10.1287/orsc.2025.21838

Rozenblit, L., & Keil, F. (2002). The misunderstood limits of folk science: An illusion of explanatory depth. Cognitive Science, 26(5), 521–562. https://doi.org/10.1207/s15516709cog2605_1

The Future Is Solo. (n.d.-b). The future is solo. Retrieved September 10, 2026, from https://thefutureissolo.com/