KNOWLEDGE 4 min READ

PART OF TECHNOLOGY STRATEGY

AI is very good at tasks that are already well defined. It is spectacularly bad at inventing definitions nobody wrote down. Many AI projects are not blocked by the technology. They are blocked by a spreadsheet called FINAL_v7, an undocumented process, or a leadership team whose success metric is having an AI initiative.

Why it matters

The cost of an AI project that fails is not only the budget. It is the confidence spent, the data exposed, and the internal conclusion that 'AI does not work for us', which then blocks the use cases that would have. A readiness audit is cheaper than discovering three months into a build that the data was never clean enough.

The four conditions

1. 01 A defined use case

   One task, one owner, one measurable outcome. 'Improve customer service with AI' is an aspiration. 'Draft first replies to the forty repeated questions our support team answers weekly, reviewed before sending' is a use case.

2. 02 Data you trust

   Accessible, consistent, and owned. If two systems disagree about the customer, the model will confidently pick one. See CRM strategy.

3. 03 A process worth keeping

   AI applied to a process nobody has examined reproduces every redundant step. Run the automation sequence first: map, question, then encode.

4. 04 A way to evaluate

   How you will know it is working versus merely confident. Sample review, error rates, time saved, and a human who owns the decision to keep it running.

The one-sentence test

For any proposed use case, write: 'This saves [whom] [how much time or money] on [which decision or task], and we will know because [which number moves].' If the sentence cannot be completed, the use case is theatre. If it can, writing it often reveals something better: the fix may not need AI at all. It may need a form with fewer fields, or two systems that finally talk. That is less exciting to announce and considerably more exciting to bank. We wrote the argument at length in AI Is Not the Strategy.

Where AI does fit

• Assistants that answer from your own material: policies, product details, past decisions. Bounded, checkable, useful.
• Agents that do the repetitive middle of a process: classification, extraction, first drafts, routing. With a human at the point where a mistake would cost money.
• Analysis over data you already trust: patterns in support tickets, reviews, sales notes, once those are consistently captured.
• Content assistance under a defined tone of voice, edited by someone who owns the brand.

Where it does not

• As a strategy. AI is a capability; the business outcome is the strategy, and the capability inherits the quality of the system it is dropped into.
• On undefined processes. It will automate the confusion.
• Where the failure cost is high and evaluation is absent: legal, financial, medical and safety decisions without human review.
• As a substitute for the definition work the team has been avoiding.

Common failure modes

• Starting from the tool. 'We have access to a model, what can we do with it?' is a search for a problem.
• Skipping the data audit. Discovering the data is inconsistent is cheaper before the build than during it.
• No evaluation. Confident output is mistaken for correct output because nobody is measuring.
• Announcing before delivering. The initiative becomes a reputational commitment rather than an experiment.

The problem is rarely that you need AI. It is usually a spreadsheet called FINAL_v7.

What does an AI readiness audit involve?

A review of candidate use cases against the four conditions, a look at the data those use cases depend on, and a clear yes, no or not-yet for each, with the reasoning written down and the not-yets turned into a fix list.

Should we build or buy?

Buy where a product fits a well-defined use case; build where the process is specific enough that no product does. Either way, the readiness questions come first, because they decide whether the thing works regardless of who made it.

How do we avoid exposing customer data?

Decide what data the system may see before it sees anything, read the provider's terms on whether your inputs are stored or used to train its models, prefer systems that keep your material within your control, and treat access like any other security decision rather than a feature toggle.

GO DEEPER

Reading is free. The diagnosis is a conversation.

Say Hi

NEXT TOPICS TO LEARN