A model can draft code, a credit memo, a clinical summary or a board presentation in minutes. Compared with the hours previously required to produce a first version, the productivity improvement can look extraordinary. But what exactly has improved: the speed of producing a draft, or the total effort required to produce something the organisation is prepared to use and defend?
That distinction matters because generation is followed by review, testing, reconciliation, approval and exception handling. I refer to that additional body of work as the verification shadow. It is easily overlooked when the production saving belongs to one team and the checking obligation belongs to another, particularly where the consequences of an error are significant.
The investment case therefore needs to follow the output through acceptance. Faster generation creates commercial value only to the extent that the full workflow becomes more effective or less expensive, including the work needed to establish that the result can be trusted.
What safety engineering tells us about the economics
Every assurance regime rests on a trace: requirements to design, design to implementation, implementation to test and test to evidence. The trace usually assumes that a named person or controlled tool stood at each link and can account for the result.
Generative implementation changes who or what produces the result, which in turn changes the evidence needed to accept it. Functional-safety practice does not answer this with a universal ban. Tool-confidence methods ask two broad questions: how serious an undetected tool error could be, and how likely the surrounding process is to detect it. Independent measures outside the tool—static analysis, unit testing, integration testing and review—can reduce reliance on the generator itself.
Read that in commercial terms: the organisation may not need to qualify every generative tool as though its output were inherently trusted. It can verify the output more aggressively instead. This approach can reduce reliance on qualifying the generator, while increasing the downstream work required to verify its output. The business case needs to capture that transfer of cost.
The least expensive qualification path in mature tool regimes often depends on a record of successful use under stable conditions and a stable version. Fast-changing generative tools struggle to accumulate that history before the version changes. The route that relies on long prior use is therefore structurally difficult for the category.
The practical result is uneven economics. Generative assistance can be highly attractive in low-integrity parts of a system and heavily gated in high-integrity ones. A single productivity assumption across an engineering organisation building both is almost certainly wrong.
Generation cost and acceptance cost are different curves
Suppose AI reduces the production time for an artifact from ten hours to one. That does not imply a 90 per cent productivity gain. The relevant equation is closer to:
Net productivity gain = production time removed − new verification, integration and exception work.
If the old process required ten hours of production and two hours of review, while the new process requires one hour of generation and seven hours of checking, the improvement is real—but it is not the one advertised.
The distinction becomes more important as generation approaches zero. Once the first draft takes minutes, almost all remaining cost sits in whether the organisation can trust and operationalise it. Higher-quality output can reduce correction effort, but average output quality is only one determinant of the verification obligation. The consequence of an error, the applicable regulation, the accountable person's responsibilities and the evidence required after the event also determine what must be checked.
The same pattern appears outside code
The same economic mechanism appears across industries, even though the specific verification obligations differ. In banking, a model-generated decision memo may be fast to produce, but consequential decisions still require data lineage, policy compliance, independent challenge, version identification and evidence an examiner can inspect.
In healthcare, summarisation may save clinician time, but a recommendation that affects diagnosis or treatment creates a different review obligation from a draft discharge note. Contraindications, thresholds and clinical responsibility do not disappear because the average model improved.
In legal work, generation may collapse the time required to assemble a first argument. Citation checking, privilege review, jurisdictional fit and professional accountability remain. In sales, a model can draft a proposal almost instantly. The real bottleneck may be solution feasibility, pricing approval, contractual risk and whether the organisation can deliver what the generated proposal promised.
In finance, an AI-created analysis can be numerically polished and still require reconciliation to source systems, control evidence and an accountable owner. The key variable is not whether the workflow is “creative” or “analytical.” It is the cost of accepting an error and the evidence demanded before and after the decision.
Better models do not eliminate every shadow
Improving model accuracy should reduce some verification work. Fewer defects mean fewer corrections. Better retrieval and structured outputs can make checking faster. Provenance tooling can automate parts of the record. But three obligations may remain even when average quality rises.
Independence. Some decisions require challenge by a person or system that did not produce the original output.
Reproducibility. The organisation may need to show which model, data, policy and version produced the result, particularly if the decision is contested later.
Accountability. Someone must still decide that the output is acceptable for use. A better model can change the evidence available to that person; it does not necessarily eliminate the role.
This creates a floor under the verification cost of certain workflows. Model quality can improve rapidly while the minimum defensible process changes slowly.
The productivity percentage is a portfolio, not a constant
The phrase “AI makes developers 30 per cent more productive” is not a usable operating assumption. Neither is its equivalent in finance, legal, healthcare or sales. A real productivity model separates at least four populations:
- Low-consequence generation. Errors are cheap and easily reversed. Verification can be light.
- High-volume structured work. Automated tests or reconciliation can check output cheaply.
- Consequential judgment. Errors are expensive and independent review remains material.
- High-integrity or regulated implementation. Traceability and evidence are part of the product, not overhead added later.
The same model may be economically transformative in the first two and marginal in the fourth. The organisation-level result depends on the mix. This is why pilots mislead. They often measure the speed of producing a representative artifact under observation. Production economics depend on repeated acceptance, exceptions, integration and evidence across the full distribution of cases.
How to measure the shadow
For each workflow, collect six numbers:
- Time to produce the artifact before AI
- Time to review and accept it before AI
- Time to generate it with AI
- Time to verify, correct and approve it with AI
- Exception rate and exception-handling time
- Rework discovered after acceptance
One illustrative workflow shows why the six numbers matter.
| Workflow: model-generated credit memo | Before AI | With AI |
|---|---|---|
| Time to produce artifact | 6.0 analyst hours | 1.2 analyst hours |
| Time to review and accept | 2.0 reviewer hours | 3.4 reviewer hours |
| Generation time | Not applicable | 0.3 system/operator hours |
| Verification, correction and approval | 2.0 reviewer hours | 3.4 reviewer hours |
| Exception rate and handling time | 8 percent x 1.5 hours | 18 percent x 2.0 hours |
| Rework discovered after acceptance | 3 percent x 2.0 hours | 5 percent x 2.5 hours |
In this illustrative case, expected total effort falls from about 8.24 hours to about 5.26, an improvement of approximately 36 percent. That is a substantial benefit, but considerably smaller than the 80 percent reduction in first-pass production time. Funding and capacity decisions should use the former measure, with the review and exception assumptions made explicit.
Then distinguish time saved from capacity converted. A person recovering three hours a week has not automatically created three hours of economic value. The operating model must redirect, aggregate or remove that capacity. Finally, identify which controls can move outside the model. Deterministic checks, reconciliation, policy rules, tests and permission limits can make verification cheaper and more stable across model changes.
The goal is not to minimise checking at all costs. It is to design the workflow so the cheapest reliable verifier handles each obligation.
What this predicts
AI will automate fastest where output can be checked mechanically, mistakes are reversible and the evidence required for acceptance can be generated automatically. It will move more slowly where judgment must be independent, consequences are asymmetric, provenance matters years later or the organisation cannot define acceptance clearly.
That does not mean the second category will remain untouched. It means its economics will be driven less by raw model capability and more by the architecture of verification. The implication for management is to assess generation and verification together, workflow by workflow. Improvements in model capability will create different returns depending on the consequence of error and the cost of establishing acceptance. Designing a reliable, economical verification process is therefore a core part of capturing AI value.