There are business problems where introducing an artificial intelligence model adds variable costs, latency, uncertainty, and a governance obligation, without contributing anything that isn't already addressed by a written rule. Saying no in those cases isn't conservatism: it's the same technical discipline that leads to saying yes in others.
A company that has never ruled out an AI use case is probably not evaluating: it's adopting.
When the rule is deterministic and written. A commission calculation, a tax validation, a discount threshold. If it can be expressed as a condition, it's expressed as a condition. A model here only introduces variability where there was certainty.
When the cost of error is high and the volume is low. Ten decisions a month with a six-figure impact cannot be automated: they require careful planning and are made by a person. The potential savings are negligible, but the risk is not.
When there is no proprietary data that provides an advantage. If the case can be solved equally well with generic knowledge available to anyone, it doesn't create differentiation. It might be useful as an individual tool, but not as a project.
When the process is not defined. Automating before mapping leads to automation of ambiguity, as we discussed in AI won't fix a poorly designed company.
When the unit cost doesn't add up. If resolving a case costs more than the current process once retries, supervision, and error correction are included, then numbers rule.
When the governance obligation outweighs the benefit. In a use case that falls into the high-risk category under the European AI Regulation, the documentation, monitoring, and traceability apparatus may be disproportionate to the modest savings. It's advisable to calculate this beforehand.
Problem | AI-powered solution | Cheaper alternative |
|---|---|---|
Classify documents by type | Language model | Rules on metadata and structure |
Extract data from invoices of fixed suppliers | Multimodal model | Templates by provider |
Detect duplicate clients | Similarity model | Normalization and deterministic rules |
Answer frequently asked questions | Conversational assistant | Search thoroughly in an organized database |
Prioritize incidents | Predictive model | Explicit business rules |
None of the alternatives on the right are glamorous, and all have three verifiable advantages: fixed cost, reproducible behavior, and trivial traceability.
The rule of thumb we apply: If the problem can be solved with a rule, it is solved with a rule.. The model is reserved for where there is genuine ambiguity — free language, high variability, unpredictable cases.
Rejecting a poorly defined use case creates friction if done as a direct refusal. It's better handled as a redirection.
Name what does solve the problem. «"This doesn't need a model, it needs the customer data to be in one place." The conversation shifts from negative to diagnostic.
Quantify the difference. Cost of the alternative versus cost of the AI project, including operation and governance. The number is more convincing than the argument.
Leave the door open with discretion. The cost of inference fell more than 280-fold between the end of 2022 and the end of 2024, according to Stanford HAI. A case that isn't profitable today might be next year. It's worth recording the calculation and reviewing it, not discarding it forever.
Offer a quick alternative. The energy behind the proposal is usually good. Redirecting it toward something that will be delivered in weeks maintains the momentum.
Organizations that manage this well have something that others don't: a documented list of cases evaluated and discarded, with the reason and date.
This list serves three purposes. It prevents reopening the same discussion every quarter with different stakeholders. It allows for reviewing decisions when economic or technical conditions change. And it demonstrates to a committee, a client, or an auditor that the company evaluates rather than simply following trends.
An inventory of AI governance without any discarded cases is, in practice, an inventory of everything anyone proposed.
There is a second-order effect that offsets the initial friction. When an organization dismisses cases based on sound judgment, those it approves reach the committee with much greater credibility. Internal sponsorship ceases to be an exercise in enthusiasm and becomes a defensible decision.
And there's a business aspect to it. A supplier who recommends against a project when the numbers don't add up is a supplier whose judgment is valuable. We prefer to lose that project and win the next conversation, which is usually bigger.
The question that should be asked in any evaluation is not whether it can be done with AI. Almost anything can be done. It's whether ought "Getting to use AI" is a different and much more useful question.
In six situations: when the rule is deterministic and written, when the cost of error is high and the volume is low, when there is no proprietary data that provides an advantage, when the process is not defined, when the unit cost is not profitable, and when the governance burden exceeds the expected benefit.
Explicit business rules, data normalization, vendor-specific templates, well-ordered search, and deterministic validations offer three advantages: fixed rather than variable cost, reproducible behavior, and trivial traceability.
A rule of thumb is this: if the problem can be expressed as a condition, it is solved with a condition. The model is reserved for cases with genuine ambiguity—free language, high variability, or unpredictable situations.
It shouldn't be. The cost of inference fell more than 280-fold between the end of 2022 and the end of 2024, according to Stanford HAI, so a case that's unfeasible today could be profitable in a year. It's advisable to document the calculation and the date so it can be reviewed.
Because it avoids reopening the same discussion every quarter, allows for reviewing decisions when economic or technical conditions change, and demonstrates to a committee or auditor that the organization evaluates with judgment rather than adopting by default.
Naming what would solve the problem, quantifying the cost difference between alternatives, recording the calculation for later review, and offering an alternative that can be delivered in weeks to maintain momentum.
Is it worth doing with AI? We analyze your specific case and tell you if it's worthwhile, what the cheaper alternative is, and when it would be advisable to review it. If the answer is no, we'll tell you that too. Let's talk →
Español Español 
English
Français
العربية
Português
Deutsch
Italiano
