A business can use an AI assistant to finish a task faster and then start the next task from the same position. It can also use part of the saving to improve the system that performs the task. The second choice can change what happens on the next cycle.
That is the useful idea behind an AI capital flywheel: accepted work creates evidence; evidence improves a retained capability; the capability makes later work easier or more dependable; and some of the resulting capacity is deliberately reinvested.
What makes an AI capability flywheel compound instead of merely generate more output? A genuine loop needs a retained improvement, a credible feedback signal, and enough control to prevent errors from becoming its training material. More activity alone does not establish compounding.
A loop needs something that remains
Imagine a business preparing product records. An assistant helps draft a description. A person checks the physical item, corrects a missing dimension, and accepts the final record. If the correction disappears into a chat history, the next item may produce the same mistake.
If the business instead improves its intake form so dimensions are captured before drafting, the correction changes future work. The retained asset is the better source process. The assistant benefits from the improvement, but the gain does not depend entirely on the model remembering the conversation.
The loop therefore has several stages: perform supported work, observe the result, identify the cause of friction or failure, change the relevant part of the system, and check whether the change helps on new cases. Skipping the final check can turn an attractive theory into an accumulating set of assumptions.
The retained improvement may be a field, a rule, a test, a skill, an integration, or a better customer promise. It should be something that changes future capability. A larger folder of generated documents may have no such effect.
Time Is Becoming Investable Capital explains the reinvestment decision. The flywheel begins when that decision creates a useful change that survives into later work.
Feedback must describe the outcome
A system can collect many signals without learning anything useful. The number of drafts, the number of clicks, or a person’s immediate approval may be easier to observe than whether the customer received an accurate product or the task was actually resolved.
Choose feedback close to the purpose. For product records, inspect corrections, customer questions about missing information, returns caused by inaccurate descriptions, and the effort needed to maintain the record. For a service estimate, inspect whether the promised work was feasible and whether the estimate required later repair.
A proxy can still be useful, but its limitations should be visible. A high acceptance rate might mean good output, hurried review, or an overly permissive standard. A low correction rate might mean fewer errors or fewer people checking.
The loop should therefore retain evidence, not just a score. A correction linked to the source and its cause is more useful than a simple “bad output” label. It allows the owner to decide whether to change the input, the instructions, the tool, or the acceptance rule.
This makes learning a directed activity. The system does not improve merely because it has seen more events. Someone must determine which events support a general change and which are specific to one case.
Complementary assets can accumulate
The Productivity J-Curve framework treats process and skill investments as important complements to general-purpose technologies. It provides a reason to look beyond the software purchase. It does not prove that any particular improvement loop will compound.
In a small organization, the complementary assets can be tangible enough to inspect. A current product record reduces uncertainty. A tested approval rule reduces repeated debate. A documented fallback reduces dependence on one person’s memory. An evaluation case protects a lesson learned from a past mistake.
These assets can support several workflows. A better product record can improve descriptions, customer answers, supplier comparisons, and stock reviews. That reuse is one source of leverage: an improvement made for one task can reduce friction elsewhere.
Reuse still requires checking context. A field appropriate for internal analysis may be unsuitable for public display. A supplier’s claimed specification may need a different label from a measurement made by the business. A shared asset should preserve those distinctions rather than flatten them.
The practical task is to identify which improvement has several supported uses, then validate the transfer. The word “flywheel” should describe that mechanism, not stand in for evidence that it occurred.
Worked example: a product-record loop
Consider a hypothetical store processing fifty product records each month. Each record initially requires twenty minutes of drafting, checking, and correction, or about 16.7 hours. A bounded AI-assisted process reduces that work to fifteen minutes per record, or 12.5 hours.
Suppose the store spends one of the released hours improving the intake form and preserving examples of recurring omissions. The remaining capacity is available for another chosen purpose. The next month, if the improved inputs reduce complete work to fourteen minutes per record, the store has a further gain that can be measured.
These are illustrative assumptions. They are not a forecast of a fixed monthly improvement rate. The useful question is why the second gain happened. Did the form remove a recurring omission? Did the review become more efficient because evidence was easier to find? Or did the mix of items simply become easier?
To distinguish those explanations, retain comparable cases and correction categories. A change that helps ordinary records but fails on unusual goods should have a narrower supported scope. A gain caused by an easier month should not be credited to the revised process.
The loop can then choose its next investment based on actual friction. It might improve condition terminology, clarify which claims require inspection, or stop generating a field that creates more confusion than value. The direction follows the evidence.
Why a constant compounding rate is usually a poor assumption
A spreadsheet can make a small improvement look enormous when it repeats the same percentage for many periods. That arithmetic is valid under its assumptions. The assumptions often lack a basis.
Early improvements may remove obvious waste. Later improvements can be harder because the remaining work involves judgment, physical inspection, or irregular cases. A model update may help one task and disrupt another. The task mix may change as the business grows.
Some improvements also overlap. If better inputs reduce both drafting and review, the benefit should be counted once in complete task time. Adding separate percentages for each claimed mechanism can exaggerate the result.
A more credible approach records discrete improvements with a cause and scope. The business can say that a new intake field reduced a specific correction category, or that a permission rule prevented duplicate sending. It can then estimate the future benefit within the observed conditions.
The flywheel may still be powerful. It does not need a constant exponential curve to matter. A sequence of modest, retained improvements can create a substantially more capable organization without supporting an unlimited growth story.
Better records can be more valuable than more prompts
When an output is wrong, the instinct may be to write a longer instruction. Sometimes the problem is an absent or conflicting source. No prompt can reliably recover a fact that the organization has not established.
A store may need a measured dimension, an actual condition inspection, or a current stock count. A service business may need a defined coverage area or a schedule that reflects available capacity. Improving those records can remove an entire class of ambiguity.
The advantage is that source improvements help both people and software. They also remain useful if the AI component changes. The organization becomes more capable rather than merely more skilled at persuading a particular model to produce a preferred format.
This suggests a useful reinvestment order. Repair the source when the evidence is missing. Repair the rule when the decision is unclear. Repair the interface when the handoff is confusing. Change the model or prompt when the supported interpretation is the actual problem.
The order is a diagnostic aid, not a universal formula. Its purpose is to avoid treating every failure as a language-generation problem. A flywheel improves the system that causes the result.
Learning should not promote every exception into a rule
A single unusual case may justify a specific response without supporting a general change. If the system automatically turns every correction into a standing instruction, it can accumulate contradictory or overly narrow rules.
Suppose one customer needs a different communication format. The business can record that preference for the customer. It should not necessarily change every message. Suppose one supplier quote includes a special term. That term should remain linked to the quote rather than become a universal assumption.
A useful learning record states the proposed scope. Does the lesson apply to one customer, one product category, a particular source, or the entire workflow? What evidence would show that the scope is too broad? When should the lesson be rechecked?
This prevents local evidence from becoming general certainty. It also keeps the procedure manageable. A compact set of well-supported rules is easier to inspect than a long history of corrections that nobody has reconciled.
The Learning Ledger provides a more detailed method for retaining those distinctions. In the flywheel, it is the mechanism that turns experience into a bounded, revisable improvement.
A negative flywheel is also possible
An inaccurate record can produce an inaccurate answer. If the business treats the answer as evidence and feeds it back into the record, the error gains apparent authority. Later outputs may repeat it more confidently because it now appears in several places.
The same pattern can occur with financial claims. A projected benefit becomes a slide, the slide becomes a planning assumption, and the assumption becomes a reported result. Nothing new was measured, but repetition creates the appearance of confirmation.
A healthy loop preserves the distinction between observation and interpretation. A tool’s output should not be treated as an independent source merely because it has been copied into a different document. The original evidence and its limitations need to remain visible.
Independent checks help where consequences justify them. A physical measurement, a transaction receipt, or an authoritative record can interrupt the circularity. Another model repeating the same unsupported claim may not provide meaningful independence.
Provenance and Auditability examines the recordkeeping side of this problem. For the flywheel, the principle is essential: a loop compounds what it retains, including mistakes.
Allocate investment across the system
A business can overinvest in one workflow because its benefits are easy to see. Faster content production may receive constant attention while fulfillment, records, and customer service remain neglected. The resulting imbalance can reduce total value.
Review the bottleneck after each meaningful change. If drafting is no longer the constraint, the next investment may belong elsewhere. The goal is a more capable business, not an endlessly optimized local task.
A simple portfolio can separate maintenance, improvement, and exploration. Maintenance keeps the current promise dependable. Improvement addresses known friction. Exploration tests a new possibility within a defined budget. Each has a different expectation and stopping rule.
The distinction protects existing customers while allowing invention. An organization should not finance speculative experiments by neglecting the records or controls that its current service depends on. Nor should maintenance consume all capacity without examining whether an unnecessary process can be retired.
A flywheel remains useful when it directs attention toward the next supported improvement. It becomes a habit when it always funds the same category regardless of the actual constraint.
Revalidation keeps the loop honest
An improvement that worked last year may fail under a new model, input source, or customer promise. The organization should retain enough evidence to recheck the capability when those conditions change.
A useful revalidation trigger is specific: a model update, a permission expansion, a changed source format, a new category of work, or a material increase in exceptions. These events can alter the mechanism that produced the earlier gain.
Keep a small set of protected evaluation cases. They should include ordinary work, a consequential exception, and a case that should produce no action. Rechecking them can reveal whether a change preserved the supported capability.
The evaluation also needs fresh cases. A system can become tuned to familiar examples while performing poorly on new work. The protected set tests continuity; new cases test whether the claimed scope still fits the actual environment.
This is a maintenance obligation, but it need not become an elaborate research program. The evidence should be proportionate to the task’s consequences and the size of the claimed improvement.
Completion matters more than the improvement queue
A loop can become overloaded with proposed changes. Each error generates a task, each task generates a plan, and the owner eventually spends more time managing improvements than performing useful work. A visible queue helps only if the organization can choose what to finish.
Limit active improvements to the number the owner can inspect. Give each a predicted effect, a small scope, and a decision date. Leave lower-value ideas in a backlog instead of treating them as current commitments. If a change does not produce its expected result, close or revise it rather than allowing it to remain indefinitely “in progress.”
This creates a cadence of completed learning. The organization can explain which changes were tried, what happened, and what remains in use. That record is a stronger sign of a functioning flywheel than the size of the improvement backlog.
AI Leverage in Practice
Choose one recurring workflow with a measured baseline and an accepted result. Identify the most frequent source of avoidable friction. Use part of the released capacity to address that specific cause, preserving the change and its rationale.
Compare new cases with the earlier process. Look for the predicted effect, such as fewer missing dimensions or less duplicate review. Also look for a counterexample: a category where the change makes work worse or does not apply.
Record the result as supported, unsupported, or unresolved. Keep its scope and revalidation trigger. Do not turn a promising association into a causal claim if the task mix or several other conditions changed at the same time.
Today’s tools can help organize feedback, draft improvements, and implement bounded changes. Future systems may perform more of that loop automatically. The authority to change operating rules should still follow a defined evidence and approval process, especially when customer commitments or financial actions are involved.
Stop improving a task when the next change has little expected value or the bottleneck has moved. Retain what works, retire what does not, and direct the next investment toward the organization’s actual purpose.
A flywheel with a visible mechanism
The AI capital flywheel is a proposed way to organize improvement, not a guarantee of exponential returns. It becomes credible when each cycle leaves a useful asset, uses evidence about the outcome, and demonstrates a benefit on later work.
The durable gain may come from better records, clearer rules, improved judgment, or a simpler process. The model helps, but the organization learns. That distinction is what keeps the loop from becoming an expensive cycle of generated output.
The next part of The Age of AI Leverage turns to Optionality Capital: how cheaper exploration can create several feasible choices before the business commits to one.
Sources
- Erik Brynjolfsson, Daniel Rock, and Chad Syverson, The Productivity J-Curve: How Intangibles Complement General Purpose Technologies, published 2021; complementary assets inform the framework, while local compounding claims require separate evidence.
Loading comments…