You’re three quarters into your AI initiative. The team is engaged. The vendor isn’t the problem. The model is one of the good ones. And yet the deck in front of you is full of “promising signals” instead of shipped wins, and you’re starting to wonder whether the technology is overhyped or your company just isn’t ready.

Neither is usually true. In the AI initiatives we see stall most often, the failure traces back to one of three missing elements — and you can identify which one in under an hour, without bringing in a consultant or rerunning the pilot from scratch. The diagnosis is the easy part. What makes it hard is that most companies don’t know what to look for, so they end up replacing the model when the model was never the issue.

This post is the framework we use, and the three questions we walk through with operating partners and CEOs whose pilots have stalled. You can run them on your own initiatives this afternoon.

The Model Is the Bow. The Big Three Is the Aim.

It helps to think about AI deployment the way an archer thinks about hitting a target. You need two things: a bow with enough draw strength to reach the distance, and the aim to put the arrow on the bullseye. Without draw strength, you can’t reach the target at all. Without aim, the arrow goes wherever physics takes it.

Choosing a model — GPT, Claude, Gemini, whatever — is choosing the bow. It matters, but the providers compete with each other to keep the bow getting more powerful, so most mid-market companies don’t need to overthink it. The aim is what you control. And aim, in AI, comes down to three things.

The Big Three

Every AI deployment that works — from Lumen Technologies’ $50M sales productivity story to JPMorgan’s LLM Suite running across 200,000+ employees — gets all three of these right.

Prompt is the what and how — the instructions you give the AI. Not a search query. A comprehensive brief. The right test: what would you tell a brand-new, highly competent teammate so they could complete this task on the first try without coming back to you for clarification?

Context is the why, where, and who — the background information a seasoned employee has in their head. Brand guidelines, company policies, historical data, customer records, institutional knowledge. The right test: what reports, datasets, or documents would you pull up if you were doing this task yourself?

Tools are what the AI can actually do. Read tools (database queries, file access, API lookups) let it gather information. Write tools (sending emails, updating CRM records, posting to systems of record, generating documents) let it complete the workflow rather than just describe it. The right test: what systems would you touch, and what actions would you take, to finish this job?

These three work as overlapping circles. Each is necessary; none alone is sufficient. The value comes at the intersection. And every stalled pilot we’ve seen has at least one of these circles either missing or thin — usually exactly one, which is what makes diagnosis tractable.

The Three Diagnostic Questions

1. Are the instructions specific enough for consistent results?

The fastest way to test this: ask whether two different employees, given the same prompt and the same task, would get outputs your team would consider equivalent. If the answer is no, you have a prompt problem. The AI isn’t the bottleneck — vagueness is.

Consider a $180M industrial supply distributor whose sales operations team builds an AI to draft RFP responses. The prompt is “draft a response to this RFP based on the customer’s request.” The model is capable. The pilot ships in three weeks. Six months later, win rate hasn’t moved. When you look at the outputs, they’re technically correct, generic, and indistinguishable from what a competitor’s AI would produce. There’s no role definition (sales engineer? account executive? technical writer?), no instruction to mirror the win-themes from recent successful proposals, no length or tone calibration. The model is doing exactly what it was told. That’s the problem.

The fix is rarely a better model. It’s a structured prompt with role, success criteria, format, and reference examples — a brief, not a search query.

2. Does the AI have the information it needs?

If the prompt is good but the outputs are confidently wrong about your specific business — getting policies wrong, citing pricing that doesn’t exist, recommending against vendors you’ve already approved — the failure is context, not capability. The model is filling in the gaps with plausible-sounding general knowledge because you didn’t give it the specific knowledge.

Consider a regional behavioral health network with 14 outpatient locations. They build an AI assistant to help intake coordinators answer the questions that come in from prospective patients — what insurance is accepted, what programs are available, what the wait list looks like. The prompts are tight. The tool can send replies through their portal. But the AI doesn’t have access to the actual current payer mix, in-network status by plan, or program-specific eligibility criteria — which change quarterly. So it generates plausible-sounding answers grounded in general behavioral-health norms. Most of the time it’s close. When it’s wrong, the patient gets routed incorrectly, and the organization picks up either a clinical-fit issue or a state mental health parity exposure.

This is the failure mode that took down Air Canada’s bereavement chatbot. The chatbot wasn’t broken; it had never been given the actual policy. The fix isn’t a new model — it’s a real connection to the source-of-truth documents, kept current.

3. Can the AI complete the workflow, or does a human still have to finish the job?

This is the question most teams skip. Even when the prompt is sharp and the context is rich, the AI can produce a beautiful analysis that nobody acts on — because it has read access to your systems but no write access, no way to actually move the work forward.

Consider a $400M regional 3PL whose finance team builds an AI to flag invoice anomalies. The team wants the AI catching things their accounts payable staff miss when they’re moving fast — duplicate billings, prices that don’t match the negotiated contract, line items coded to the wrong account. The prompts are dialed in. The context includes contract terms, vendor master records, and historical invoice patterns. Detection accuracy on a back-test hits 94%. But the AI has no integration with the AP system to actually hold suspect invoices, no ability to route exceptions to controllers, no way to notify vendors. Every Monday it produces a 40-page PDF the AP team has to manually work through. Six months in, the team stops opening the report. The pilot is technically a success and operationally dead.

The fix is rarely more analysis. It’s giving the AI the write access to complete the workflow — with appropriate guardrails, audit trails, and human approval where the stakes warrant it.

The Failure Patterns

Once you’ve answered the three questions, the diagnosis usually falls into a clean pattern:

  • Great prompt, no context: Generic outputs that sound right but don’t reflect your business. The AI is competent at the task you described and clueless about your company.
  • Great context, no tools: An expensive report generator. Insightful PDFs that nobody has time to act on. Adoption dies quietly around month four.
  • Great tools, vague prompt: Inconsistent outputs at scale. Sometimes it works; sometimes it doesn’t. You can’t tell why, and your team stops trusting it.

Each of these has a different fix, and the fix is almost always cheaper than the rebuild your vendor is going to recommend.

What to Do This Week

Pick one stalled or underperforming AI initiative — the one your team won’t stop apologizing for in the steering committee — and walk it through the three questions. Be honest. Don’t grade on a curve. If a competent new hire couldn’t replicate the prompt, that’s a no. If the AI is filling in gaps with general industry knowledge, that’s a no. If a human still has to close the loop, that’s a no.

By the end of an hour, you’ll know which of the three elements is missing, and you’ll have a clear sense of whether the fix is hours of work or weeks. In our experience, it’s almost always hours.

The Bottom Line

The companies that succeed with AI aren’t the ones with the biggest models or the best vendors. They’re the ones that get the aim right. When a pilot fails, the temptation is to assume the technology isn’t ready. Almost always, the technology is fine — but one of the three things it needs to do its job is missing. Find the missing piece first. Then decide whether you need a better bow.

This is the fourth in a series of five posts on the AI Value Map. Next: even when you’ve nailed the Big Three, there’s still one more way to over-engineer your way into a stalled deployment — and a 700-year-old principle that prevents it.