When the Right Answer Is to Wait: AI Work Not Worth Doing Yet

Wait on AI when the job is rare, when the data or method it needs is not written down, when mistakes are costly and hard to spot, when nobody will own it, or when a tool you already pay for is about to do it. Waiting is a decision, not a refusal: record why and set a review date.

When is the right answer to wait on AI?

The right answer is to wait when a job fails the basic tests for a good AI project: it is rare, its inputs are missing, its outputs are hard to check, nobody owns it, or a cheaper route is coming. In those cases, AI work costs more than it returns. The money is better spent on a job that passes.

Saying "not yet" is unfashionable. Vendors, boards and staff all push for more AI, sooner. Yet a strategy that never says no is not really a strategy. It is a wish list.

The cost of starting the wrong project is not only the spend. A visible failure makes the team sceptical of the next one, and that can set adoption back months. So a deliberate delay on a weak project often does more for AI adoption than pushing it through.

Which signs say a job is not ready for AI?

Five signs say a job is not ready: it happens rarely, the knowledge it needs is not written down, errors are costly and hard to spot, no one will own the result, or a tool you already have is likely to add it soon. Any one is a reason to pause. Two or more usually means wait.

Each row has a mirror image in what makes a good first project. Frequent, defined, checkable work with a measure attached passes, as the guide to the weekly workflow test explains.

Why does missing data mean you should wait?

Because AI cannot use knowledge the business has never written down. If the method lives in one person's head, or the facts sit in three conflicting documents, the AI will guess. Its guesses will sound right, which makes them dangerous. Writing the knowledge down first is the real project.

The research is consistent on this. According to RAND's 2024 study, two of the five root causes of AI project failure are that the organisation lacks the data, and that the problem is too hard for AI. Gartner gave similar reasons when it predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025. Poor data quality was among the reasons it gave.

Waiting here does not mean doing nothing. It means doing the groundwork: writing down how the job is done, collecting examples of good output and settling which facts are true. That work pays off whether or not AI is added later.

Why should high-stakes, hard-to-check work wait?

Because AI errors are most dangerous where they are costly and look plausible. Legal judgements, pricing decisions, sensitive customer complaints and anything published without review all fall here. Until a person can check every output quickly and reliably, the risk outweighs the time saved. Use AI for research around these jobs instead.

The error rates in specialist work can be high. According to Stanford's RegLab and HAI, state-of-the-art language models hallucinated on 69% to 88% of specific legal queries in their 2023 tests. OpenAI's own researchers explain one cause: training and evaluation "reward guessing over acknowledging uncertainty". The guide to AI hallucinations covers the fixes that work.

Feelings are a poor guide too. In METR's 2025 randomised trial, experienced developers took 19% longer with AI tools, yet still believed AI had made them 20% faster. Where output is hard to measure, people cannot tell whether AI is helping.

When is waiting for your existing tools the smart move?

When the software you already pay for is likely to add the capability you want. Building a custom workflow that your CRM, office suite or AI assistant ships later wastes the build and leaves you maintaining a duplicate. If the job can wait a few months, waiting and then switching it on is often cheapest.

This is the third option in the build, buy or switch on decision, applied over time. Ask your vendors what is on their roadmap before commissioning anything similar. Treat promises with care, and decide on what ships, not on what is announced.

Be equally careful with new products that promise to do everything. According to Gartner, over 40% of agentic AI projects will be cancelled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls. Waiting until a category matures can be sound judgement.

How do you wait well?

Record the decision, the reason and a review date. Do the groundwork that would make the job ready, such as writing down the method or cleaning the data. Name what would change your mind, such as a vendor shipping a feature. Then revisit it at the next quarterly review, alongside the rest of the roadmap.

A simple "not yet" record holds four lines:

1. The job. What was considered.
2. The reason. Which sign it failed.
3. The trigger. What would make it ready.
4. The review date. When someone looks again.

Most of these jobs come back. Some return ready, because the groundwork was done. Others return because the tools improved. Either way, the business decides on evidence rather than enthusiasm.

Skein switches on a tool the client already pays for rather than building, and says so when the answer is to wait; the AI strategy page shows how opportunities are ranked. That suits a growing B2B business with more AI ideas than it can judge. A business that has already decided to build a specific system may want a builder, not an adviser.

Frequently asked questions

When should you not use AI?

Avoid AI for jobs that are rare, depend on knowledge nobody has written down, carry costly errors that are hard to spot, have no owner, or will soon be covered by software you already pay for. In those cases the cost and risk usually outweigh what AI would return.

What tasks should not be done by AI?

Tasks with high stakes and hard-to-check output should stay with people, such as legal judgements, pricing decisions, sensitive complaints and anything published without review. AI can still help with research around them. The test is whether a person can quickly and reliably check every output.

Is it okay to delay AI adoption?

Yes, for specific jobs that are not ready. Delaying a weak project protects time, money and the team's confidence. Delay should be a recorded decision with a reason and a review date, not a general refusal. Start with jobs that pass the basic tests while the others wait.

Why do AI projects fail because of data?

AI cannot use knowledge that is missing, scattered or contradictory, so it fills the gaps with plausible guesses. RAND lists lacking the necessary data among the five root causes of AI project failure, and Gartner named poor data quality as a reason generative AI projects are abandoned.

How do you decide which AI projects to postpone?

Score each idea against five signs: how often it runs, whether its knowledge is written down, how costly and visible errors would be, whether someone will own it, and whether an existing tool will add it soon. Postpone ideas that fail two or more, and record a review date.

Related reading