Diop Daily #133 — October 2026

Who Pays for the Second Attempt?

The first attempt is the seductive unit of AI economics. A model receives a prompt, calls a tool, returns an answer, and produces a clean line on a usage dashboard. The institution pays for tokens and declares a workflow. The second attempt is where the institution discovers what it actually bought: a retry after a malformed call, a human review after an uncertain answer, a correction after a polished error, a translation check, or a delay that changes the value of the result.

This is why the price of intelligence cannot be read from the price of a token. OpenAI’s July guidance for AI investment recommends measuring “cost per accepted outcome,” including model and tool use, attempts, completion rate, latency, and human review. Its September usage-analytics guidance connects spend with task categories, model choice, speed, training, and business outcomes. Stanford’s 2026 AI Index reports that productivity gains are largest in structured, measurable work where outputs are easy to monitor, while compute spending and infrastructure costs continue to rise. These sources point toward one accounting problem: an AI system has an economic unit only when an institution can define acceptance and count the work required to reach it.

The cheap answer is the answer that reaches an accepted state without hiding the cost of getting there.

The token is an input, not a result

Token pricing is attractive because it is precise. A provider can state a rate, a dashboard can record consumption, and a buyer can compare models on a familiar axis. That precision becomes misleading when the workflow has consequential failure modes. A low-cost model that requires three retries and twenty minutes of review may be more expensive than a higher-cost model that produces a usable result on the first pass. A fast response that creates a dispute can consume more institutional capacity than a slower response that arrives with the evidence needed for approval.

The full cost of an agentic task has several components:

  • Inference: the model calls, context, reasoning budget, and output tokens used to produce a proposal.
  • Tool work: API calls, browser actions, retrieval, storage, payment attempts, and external services touched by the proposal.
  • Correction: retries, re-planning, data repair, translation review, human checking, and recovery after a failed action.
  • Time: latency, queueing, escalation, and the interval during which a person or process waits for a usable state.
  • Consequence: the cost of a wrong publication, a missed deadline, a duplicated payment, an exposed record, or a decision that must be reversed.

The last two categories are often invisible in model dashboards. They live in calendars, support queues, finance systems, editorial corrections, and the attention of people who were never counted as part of the AI workflow. An institution that measures only inference is measuring the machine’s appetite while leaving the organization’s metabolism unpriced.

Acceptance is a governed state

“Accepted outcome” sounds simple until a real institution has to define it. A support answer may be accepted when the customer’s issue is resolved and the record is correctly updated. A code change may be accepted when tests pass, review is complete, security checks are green, and the change can be maintained by the team. A research brief may be accepted when its sources are accessible, its claims are bounded, its rights are clear, and another reader can follow the reasoning.

Each definition contains a threshold, an authority, and a route for disagreement. Acceptance belongs to the institution that bears the consequence. A provider can report that a model generated an answer; the buyer has to decide whether the answer can enter the work.

This changes the design of observability. A serious outcome ledger should join the model trace to the institutional state:

  1. the task, owner, principal, and intended outcome;
  2. the model, context, tools, and policy version used in the attempt;
  3. the first result and every retry, escalation, or correction;
  4. the reviewer, test, source, or business rule that established acceptance;
  5. the time, financial cost, human capacity, and residual uncertainty carried forward.

This record does more than calculate a margin. It tells the organization which workflows are becoming reliable, which ones are consuming hidden labor, and which forms of intelligence are worth reserving for difficult cases. It also gives the institution a way to compare providers without confusing a cheap proposal with a cheap operation.

The second attempt is where local conditions appear

Imported benchmarks often describe an idealized workflow: stable connectivity, abundant compute, a dominant language, standardized records, and a reviewer who can intervene quickly. The second attempt exposes the conditions that the benchmark omitted. A request may need to be translated into a local language. A mobile connection may drop during a payment call. A document may be stored in a format that the retrieval system cannot parse. A community institution may require a human authority that the imported workflow has no field to represent.

For African institutions, these conditions are part of the economic object. The Institute for Security Studies has argued that African AI sovereignty depends on investment in power, connectivity, computing, cooling, and data ecosystems. The consequence for agent systems is practical: an outcome ledger must include the scarce resources that make a workflow possible. Compute imported from elsewhere, bandwidth paid at the edge, language review performed by a local expert, and recovery work carried by a small institution all belong in the cost of the result.

The ledger should account for these constraints accurately, without romanticizing them. A workflow that appears cheap because it excludes translation, local validation or intermittent connectivity transfers its cost to the institution that has the least margin for absorbing it. Measurement becomes an instrument of sovereignty when it makes these transfers visible.

From usage analytics to institutional learning

Usage analytics tell a buyer where activity is happening. Outcome accounting asks what the activity changed. The difference is the difference between a meter and a memory.

Suppose an institution sees that one team consumes far more model credits than another. The first reaction might be to reduce access or route the team to a cheaper model. A better investigation asks what work the credits represent. Are the extra attempts caused by poor source data, an ambiguous policy, an inadequate tool contract, a language gap, or a workflow whose acceptance test is too weak? Each answer requires a different intervention.

The ledger should therefore preserve cause, not merely volume. A useful classification might record:

  • productive exploration that has not yet reached a production workflow;
  • repeatable work whose cost and quality are within an agreed envelope;
  • avoidable loops caused by weak instructions, tools, context, or routing;
  • necessary review for high-consequence work;
  • unresolved failure that requires a change to the institution rather than another prompt.

Over time, this classification becomes a map of operating knowledge. It shows where a small model is sufficient, where a frontier model earns its cost, where a human decision is the actual bottleneck, and where the organization should repair its records or policies before buying more intelligence. Capital can then follow proven capacity rather than raw activity.

Where the investable surface is widening

The commercial object is the outcome ledger: infrastructure that attributes the full cost of accepted machine work across models, tools, people, time, language, and consequence. It sits beside usage analytics but reaches into the state that makes work count.

  • Outcome accounting systems: ledgers that join model traces, workflow states, approvals, correction events, and accepted results into a cost-per-outcome view.
  • Correction-aware routing: allocators that choose models and tools based on expected total cost, completion probability, review burden, latency, and consequence rather than token price alone.
  • Acceptance registries: versioned definitions of what counts as good enough for a domain, with owners, tests, evidence requirements, and appeal paths.
  • Capacity and review markets: systems that schedule scarce human expertise, local-language validation, domain review, and recovery work as measurable production resources.
  • Sovereign cost observability: infrastructure that exposes the power, connectivity, localization, data-hosting, and provider-dependence costs carried by an institution or region.

Underwriting can begin with measures that a buyer can inspect: cost per accepted outcome; acceptance rate on the first attempt; correction minutes per task; retry causes; time from proposal to approval; local-language quality variance; percentage of outcomes with complete evidence; and the share of work whose cost remains portable when a provider changes. These measures do not guarantee a good business. They make the business legible enough to govern.

The strongest products in this layer will sell disciplined allocation. They will help a finance leader distinguish productive demand from waste, a technical leader decide where to spend evaluation effort, a public institution protect scarce human review, and an investor assess whether machine capacity is becoming reusable operating leverage. Their defensibility will come from domain-specific acceptance data and from the accumulated record of why a workflow succeeds or needs another attempt.

Build the ledger before scaling the call volume

An institution can begin with one workflow that matters. Define the accepted state in language the owner recognizes. Record the model and tool costs. Add human review, delay, translation, failed calls, and correction. Classify every second attempt. Revisit the ledger after a fixed period and ask which intervention reduced the full cost without lowering the quality bar.

Then test the ledger itself by changing the model, removing a tool, introducing a slower connection, translating the task, requiring a different reviewer and forcing a refusal. The system should preserve enough evidence to show whether the workflow became cheaper because the machine improved, because a human absorbed more work, or because a consequence disappeared from the measurement.

The market will continue to advertise intelligence in tokens, parameters, and benchmark scores. Institutions will buy something more concrete: accepted work under a cost and risk envelope. The second attempt is where that envelope becomes visible. Count it honestly, and AI economics can become a discipline of institutional learning rather than a ceremony around usage.

Sources