A company should decide against a proposed AI use when an adequate alternative serves the task and the additional benefit does not justify the cost, uncertainty or work needed to check it. It should also decline a use whose necessary safeguards it cannot provide. Make that decision about a specified piece of work, then name what will happen instead.
“Do not use AI” leaves employees with a problem if nobody explains how the work will get done. A useful decision identifies the alternative, its owner and the evidence that would justify reconsidering it. That can be a positive operating choice within an enterprise AI adoption strategy.
It also needs an honest boundary. A missing prerequisite might justify a pause. A failed experiment might justify stopping that experiment. Neither establishes that AI is unsuitable for everything the company does.
Define the job before refusing the technology
Suppose a wallcoverings distributor with 180 employees wants to improve its internal sample requests. Sam Taylor, a showroom coordinator, knows the wallpaper pattern code and colourway, meaning the product’s particular colour combination. He needs a sample sent to the right showroom. Claire Bennett, the sample-desk manager, needs a complete request in the queue.
For that job, Claire can consider a form that checks the code against the current catalogue and passes the selected pattern, colourway and showroom into the request queue. A missing or unknown code returns to Sam for clarification. There is no need to interpret an open-ended description merely to transfer known selections.
Now change the input. Sam has a photograph of wallpaper in a customer’s room and does not know the pattern. Identifying a possible match is a different job. A catalogue form cannot do it, and guessing a code would conceal the uncertainty. AI-assisted identification might warrant a separate evaluation, with an appropriate specialist checking the suggested match.
Two sample-desk jobs with different inputs
Claire is choosing how to handle a showroom sample request. The input changes what the arrangement must do.
Select a catalogue entry
Sam supplies pattern code, colourway and showroom. A validated lookup passes the selection to the request queue. Unknown codes return for clarification.
Identify a possible pattern
Sam cannot supply a code. The form does not resolve the image. Any AI-assisted matching proposal needs its own evaluation and specialist check.
Claire can decline AI for known-code requests without pretending she has solved photo identification. That distinction matters more than describing the whole sample desk as either “AI enabled” or “AI free.”
The UK government’s AI Playbook, published in February 2025, advises choosing the right tool and remaining open to traditional technology where it better addresses the problem. It is public-sector guidance, rather than a private company’s operating rule. The useful question carries across: what does this task require the technology to do?
Compare an actual alternative
The alternative deserves the same scrutiny as the AI proposal. “We could use a form” is a design idea until someone has checked the catalogue, request handling and exceptions. A deterministic rule produces the same result for the same specified conditions; it can still be wrong if the rule, data or implementation is wrong.
For Claire, acceptable work means that the requested pattern and colourway are valid, the showroom is correct, and Sam can tell whether his request was accepted or remains unresolved. She also needs a way to correct catalogue entries and deal with failed transfers. Those responsibilities do not disappear because the solution uses familiar software.
Compare the proposed arrangements against the same need:
| Arrangement | What could make it suitable | What still needs checking |
|---|---|---|
Remove a redundant step | The step has no necessary purpose or recipient | Removing it does not leave required work undone |
Use a lookup or fixed rule | The relevant choices and conditions can be stated explicitly | The source information, exceptions and maintenance have owners |
Keep experienced human handling | The task needs judgment that the proposed service cannot support adequately | People have the information, capacity and escalation route |
Evaluate AI assistance | Interpreting variable material could add useful value | Output quality, review work and consequences fit the intended use |
These options can coexist. The point is to compare plausible ways of completing the job, including their continuing work, rather than compare a polished AI demonstration with a deliberately neglected manual process.
Human accounts help expose the difference. In an October 2026 LinkedIn post, consultant Emile Languepin described visiting a company proud of its AI use while he thought shortcomings in its customer relationship management system might explain the underlying problem. The company is unnamed and the alternative was not independently evaluated. His account is useful as a prompt to investigate the diagnosis, rather than proof that fixing the system would have produced a particular result.
A July 2026 Reddit comment offers a narrower alternative: a Reddit user reported automating the cleanup of poorly submitted CSV files with PowerShell and Windows Scheduler. In the same discussion, another Reddit user reported using AI for log parsing and an initial summary before examining a ticket. These are unverified accounts, without a common performance measure. They describe different tasks, and give no reason to force the same method onto both.
Distinguish a non-use decision from a missing prerequisite
An absent catalogue is not evidence that AI cannot help a sample desk. It is evidence that the proposed arrangement lacks an authoritative way to check the product. Depending on the intended use, both an AI service and a fixed-rule system could be affected.
Be precise about why you are declining or pausing:
- The alternative is adequate and preferable for this task. State what it does, what was checked and why adding AI would not justify its extra demands.
- The proposal cannot meet a necessary condition. Identify the unsupported action or unacceptable consequence. Keep that action out of use unless the condition changes.
- The evidence is insufficient to choose. Name the missing comparison or unresolved requirement. Decide whether gathering that evidence is worth the effort.
A pause should not turn into an indefinite instruction to “improve readiness.” If Claire needs a current catalogue, name its owner and what must be corrected. If she cannot justify investigating photo identification yet, leave that separate request uncommissioned. There is no obligation to run an AI experiment simply because someone proposed one.
Use the readiness assessment guide when missing information, access or responsibility prevents a meaningful decision. If an experiment has already run, use its actual findings to decide whether to stop, extend or scale the pilot. That is a different decision from choosing not to commission it in the first place.
Record the work you are choosing to support
A short decision note is enough if it makes the operating arrangement clear. Keep the evidence behind the decision available, including any unresolved question. Avoid a numerical suitability score that hides which requirement the proposal cannot meet.
For Claire, the note could read as follows.
That is a proposed operating decision with acceptance conditions. Claire must still obtain the evidence before describing the form as adequate. If the check exposes a problem, the response is to repair or reconsider the arrangement, rather than preserve a “no AI” decision at any cost.
For your own note, include:
- The bounded task and its inputs. Make clear what the decision covers and what remains outside it.
- The chosen arrangement and responsible person. Explain how the work reaches an acceptable result, including unresolved cases.
- The basis for the choice. Keep observed findings separate from assumptions and work still to be checked.
- The condition for reconsideration. Name a change that could alter the comparison, and who would review it.
Do not use an adoption target to reverse a well-supported task decision. A team can make progress by removing unnecessary work or improving a reliable existing route. Equally, choosing a familiar method does not excuse ignoring evidence that it is failing.
Reopen the decision when the comparison changes
NIST’s voluntary AI Risk Management Framework 1.0 supports an initial go/no-go decision based on the intended context. Its MANAGE 2.1 guidance also includes considering viable non-AI alternatives and the resources needed to manage risks. It does not supply a universal score that determines which technology every company should use.
For Claire, a useful reopening trigger could be evidence that photo-only requests are a significant unmet need and that specialists can check proposed matches within the available capacity. That would justify considering a separate task. It would not establish that the known-code form should start interpreting photographs.
Other changes could affect the original decision: new input types, a recurring exception the rules cannot handle, or credible evidence that AI assistance improves acceptable completed work after review. A new product announcement alone does not demonstrate any of those conditions.
Keep the rejected proposal and its reason available so a later sponsor can understand the choice. Set a proportionate review point where needed, and let the process owner raise material changes sooner. Once there is a justified AI use, the buy, build or combine guide can help examine delivery options and their continuing responsibilities.
Questions about deciding not to use AI
Does declining AI mean an adoption program has failed?
No. An adoption program should help the company make useful decisions about its work. Declining a specified AI use can be a sound result when an adequate alternative better meets the need. Record the task, evidence and chosen arrangement so the decision can be understood and revisited. A blanket refusal without examining the work does not provide the same basis. Start with the sample-request distinction.
When are fixed rules a better choice than generative AI?
Fixed rules are worth considering when the relevant conditions and acceptable output can be specified clearly, such as validating a known product code against an approved catalogue. Their suitability still depends on accurate information, exception handling and maintenance. When the task requires interpreting variable material, a rules-only alternative may not meet the need. Compare the actual arrangements using the same task and acceptance conditions.
What if the company has no evidence that either approach works?
Treat the choice as unresolved. Define an acceptable result and identify the smallest useful comparison, or decide that investigating the proposal is not currently worth the effort. Do not report an untested alternative as successful or AI as unsuitable merely because evidence is absent. Give the missing prerequisite an owner when it is worth resolving; the readiness assessment explains how to make those gaps concrete.
Can a company use AI to build a workflow that does not use AI?
Yes. AI assistance during development and AI making interpretations during daily operation are different uses. For example, a developer might use an approved assistant to help write a catalogue lookup, then test the resulting rules and exception handling. That does not establish that the request process needs AI at runtime. Assess each use with its own information permissions, checking and continuing responsibilities.
What evidence should reopen a decision against AI?
Look for a change that could alter the original comparison, such as a material unmet need, different inputs or credible task evidence of better completed work after review. Name who will examine it and which requirement it affects. Reopening authorizes reconsideration, not automatic rollout. The reopening conditions show why a photo-identification request should be considered separately from a known-code sample form.
