Prompts are not an AI strategy

Prompt advice has its place. A sharper prompt can improve a draft, tidy a summary, or help someone think through a task they were stuck on. But a folder full of clever prompts is not an AI strategy, and treating it like one is how a lot of businesses end up with nothing to show after six months.

It’s easy to see how businesses land there, because the prompt-library path is the path of least resistance at every step. Someone runs a lunch-and-learn, everyone leaves with ten tips, a shared document gets created, and for a few weeks it feels like the organisation is “doing AI”. There’s a whole cottage industry feeding this: prompt packs, prompt courses, consultants selling laminated prompt cards. None of it is useless. All of it is individual productivity advice wearing a strategy costume. The tell is what happens when someone asks a business question: did quoting get faster? Are errors down? Who’s using it and for what? A prompt library can’t answer any of that, because it was never wired into anything that could.

Business AI needs more than good wording. It needs access to the right data, rules about who can use what, a workflow that says when a human steps in, somewhere to measure whether any of it worked, and a name next to the question of who keeps the thing running.

Prompts do not fix data access

If the model can’t reach your current policy, the customer record, the job history, or the document set, then no prompt is going to save you. Staff end up pasting material in by hand, missing context, or working from a version that went stale three weeks ago.

Watch the hand-pasting pattern for a week and you’ll see the problem compound. The staff member finds the policy document, or a version of it, copies chunks into the chat window, gets an answer shaped by whatever fragment they pasted, and makes a decision on it. Multiply across a team and you’ve got dozens of daily decisions made on different, partial, possibly outdated slices of the same source material, with no record of which slice produced which answer. That’s not augmented judgement; it’s inconsistency with better grammar. And the pasting itself is a governance hole, because nobody is checking what’s being copied into which tool on the way through.

A system worth building connects to approved sources and shows you where each answer came from. Same model, same staff, completely different reliability, because the inputs stopped being freelance.

Prompts do not set permissions

A prompt cannot decide who is allowed to see payroll figures, legal records, customer data, or commercial documents. That call belongs in how the system is designed, not in the text someone types into a box.

Skip that design and a useful assistant quietly turns into a privacy problem. The uncomfortable example: a business loads its shared drive into an AI assistant so staff can “ask questions about our documents”. The shared drive contains, as shared drives do, salary review spreadsheets, a disciplinary letter, and the confidential terms of the biggest client contract. Nothing was hacked. Someone just asked a well-phrased question, and the assistant, doing exactly its job, answered it. Permissions have to be enforced below the conversation, at the level of what the system will retrieve for whom, which is architecture, not prompting. It’s the difference between telling staff not to open the filing cabinet and locking it.

Prompts do not create workflow

A prompt can draft an email. On its own it can’t decide when that draft gets reviewed, where the final message is stored, which customer record gets updated, or what the system does when the answer is shaky. That’s workflow and integration work, and it’s the part that moves AI from a handy tool to something the business actually runs on.

The difference shows up in what happens around the model call. Prompt-level AI: the estimator asks a chatbot to draft a quote email, edits it, sends it from their inbox, and nothing else in the business knows it happened. Workflow-level AI: the quote request arrives, the system pulls the customer’s history and current pricing, drafts the quote, routes it to the estimator for approval, sends it, files it against the customer record and starts the follow-up clock. The model does the same drafting in both stories. Everything that makes the second one valuable, the context arriving automatically, the approval gate, the record-keeping, the follow-up, is plain software design. Which is why the strategy question is never “what should we prompt?” but “which workflow should we rebuild with a model inside it?”

Prompts do not measure value

A prompt library can feel productive while changing very little. Leaders still need to know whether time was actually saved, whether errors dropped, whether turnaround got faster, whether staff used the thing at all. That means a baseline before you start and a way to compare after.

Without it, your AI adoption is a pile of anecdotes about that one time it wrote a good email. Anecdotes have a shelf life: they carry the first budget conversation, but eighteen months in, when someone asks what the AI spend returned, “the team likes it” doesn’t survive contact with a CFO. The measurement doesn’t need to be elaborate. Pick the two or three workflows that matter, write down the current numbers before anything changes, hours per week, error rate, turnaround, and compare quarterly. We’ve covered the mechanics in measuring the ROI of AI and automation. The businesses that do this can also do something the anecdote businesses can’t: kill the uses that aren’t working and double down on the ones that are.

What strategy should cover

A useful AI strategy names the priority workflows, the tools you have approved, where the data boundaries sit, the privacy controls, who governs it, how staff are trained, what integrates with what, and how you will measure any of it. It should also be honest about what you are not going to use AI for yet.

It doesn’t need to be a thick document. It needs to make decisions easier. The test of a real strategy is that it answers the questions that actually come up: a staff member asks “can I use this new tool on client files?”, and the answer takes thirty seconds because the data boundaries are written down. A manager proposes an AI project, and it can be ranked against the named priorities instead of debated from scratch. If your strategy document can’t settle questions like that, it’s a position paper, not a strategy, whatever its page count. A one-page version that settles them, like the minimum viable policy plus a ranked workflow list, beats forty pages that don’t.

There’s also a shelf-life problem nobody mentions at the lunch-and-learn: prompt libraries rot. The clever phrasing that coaxed a good result out of one model version behaves differently on the next, the tool updates quietly, and the laminated card is now advice about software that no longer exists. Systems get maintained and re-evaluated when models change; folders of tips just silently go stale, and nobody notices because nobody was measuring.

Keep prompts in their place

Prompts are useful inside a wider way of working. They help staff talk to the systems you have built. They do not replace those systems, and they never will.

If your AI plan is mostly prompt tips passed around in a Slack channel, move up a layer and design the work properly. The tips can stay; they’re the icing. The layer underneath, data, permissions, workflow, measurement, ownership, is the cake, and it’s the part nobody can paste into a shared doc. The AI readiness assessment is a practical way to start: it looks at your workflows, data and constraints and hands back a ranked list of what’s worth building, which is the thing the prompt library was standing in for all along.