The market problem: vendors saying yes to everything
There is a structural incentive in the AI consulting and AI tools market to tell founders they should be using AI. Vendors who sell AI tools need buyers. Consulting firms who run AI projects need projects. Boards who pressured the CEO to “have an AI strategy” need an AI strategy to point at.
The aggregate effect is that the assessment most founders take ends in the same conclusion almost every time: yes, you are ready; here is what to buy. The Service Direct SMB survey from 2025 noted that 62% of small business non-adopters said their #1 blocker was “we don’t know where to start.” That cohort is being aggressively pitched right now, by everyone. Most of them shouldn’t be starting yet.
The five signals below are the ones we look at in the AI Readiness Assessment when we are deciding whether to recommend that a business pause. Hitting any one of them is a flag. Hitting two or more is a “not yet.”
Signal 1: Foundations aren’t in place
The first signal is the easiest to spot and the most ignored. The data, governance, and named-ownership underneath AI are not in place.
Concretely:
- Data: Your operational data lives in five systems that don’t talk to each other. An ERP, a WMS, a CRM, a couple of spreadsheets, a Slack channel that has somehow become the source of truth on one workflow. Nobody can produce a single clean dataset of, say, “what happened on each customer order in the last six months” without two engineers spending a week on it.
- Governance: Nobody has decided what AI is or isn’t allowed to do on its own. Who signs off when a model output goes to a customer? What data can a model see, what data is it not allowed to see? What is the escalation path if the model says something it shouldn’t?
- Ownership: No single person on your org chart is on the hook for any AI initiative. The AI work is “led” by a committee, or by a head of innovation who reports to nobody specific, or by a vendor who will not be there in twelve months.
When any of these three is missing, AI work compounds the existing mess. The pilot might demo well, because the data scientist will hand-clean the data and the team will baby-sit the model for the demo. Production runs are different. Production runs need foundations.
Honest signal: if you cannot, in one minute, name the person who owns each AI initiative, the dataset they’re using, and the governance line they cannot cross, the foundation is not in place. Pause. Fix.
Signal 2: No measurable target metric for the next project
The single sharpest predictor of AI pilot failure. You cannot point to a number the next AI initiative will move, with a magnitude and a time horizon.
Statements that fail this test:
- “We’re going to use AI to be more efficient.” Which metric? By how much?
- “AI will help us with customer service.” Cost-per-ticket? Average handle time? CSAT? First contact resolution?
- “We’re piloting AI to see what it can do.” Pilots that exist to “see what it can do” almost universally fail to ship.
Statements that pass:
- “We expect this AI initiative to reduce customer-service cost-per-ticket by 12% over six months.”
- “This will move warehouse pick-pack throughput by 8% in twelve weeks.”
- “This will lift the conversion rate on the checkout flow by 1.5 percentage points in eight weeks.”
The difference is not the size of the number. It is the precision. A pilot with a precise target gets owned, measured, and corrected. A pilot without one floats. Most of the budget gets spent before anyone notices that nothing was supposed to move in particular.
If you cannot complete the sentence “this AI initiative will move ____ by ____ in ____,” you are on a pilot. You are not on a project. Pause and define the metric before you start.
Related reading: Why 95% of AI Pilots Fail to Deliver Financial Returns goes deep on the named-owner + target-metric pattern that separates the 5% from the 95%.
Signal 3: The team can’t sustain it after delivery
The third signal is operational. The team that will run the AI after the vendor or consultant leaves does not have the capacity, the training, or the standing to do so.
A useful test: imagine the AI is in production on Monday. The vendor leaves on Friday. Now ask:
- Who watches the model’s output quality over the next quarter?
- Who is responsible when the model drifts?
- Who decides whether the model retrains, and on what schedule?
- Who handles the conversation with operations when the model recommends something that creates an exception in the existing workflow?
If your answer to any of those four is “we’ll figure that out” or “the vendor will handle it via SLA,” the AI initiative will fail in production. Not because the model is wrong. Because the operating discipline around the model doesn’t exist.
This is the bucket that mid-market businesses underestimate most. The AI work is sold as a project. It is actually a system. Systems need owners and operating disciplines, not just deliverables.
Honest signal: if your operating team is at full capacity right now and you cannot reasonably name who will take on the AI in production, the project is going to drop. Pause until the operational layer is in place.
Signal 4: The last AI initiative didn’t stick (and you didn’t diagnose why)
The fourth signal is pattern-recognition. You ran an AI pilot in the last twelve to twenty-four months. It did not produce a measurable result. You have not run an honest post-mortem on why.
This is the most common failure pattern we see in the businesses that approach us. They want to talk about the next AI thing. The previous AI thing is not discussed, because it is uncomfortable. The previous thing didn’t ship, or shipped and didn’t move a metric, and the team moved on without naming why.
Starting a new AI initiative on top of an undiagnosed failure is an act of optimism. It is also a tax on the next quarter. Whatever broke the last initiative is almost certainly still broken. Same data issues, same ownership gap, same target-metric vagueness, same operating-team capacity problem. The new initiative will hit the same wall.
The post-mortem doesn’t need to be elaborate. An hour with the team, an honest read of what was supposed to happen versus what happened, and a written summary of what would have to be different next time. That hour is worth more than the next pilot’s kickoff meeting.
If you have not done that hour, do it before you start anything new. We have helped CEOs run that conversation; it is uncomfortable in the first ten minutes and clarifying for the rest.
Signal 5: The pressure is from the board, not from a business problem
The fifth signal is institutional. The driving force for AI work in your business is external pressure (from a board, a PE firm, or industry noise), not an internal business problem that AI is positioned to solve.
BCG’s AI Radar from 2026 reported that about half of CEOs believe their job stability depends on getting AI right. That is real pressure. It is also why a lot of AI work gets started before the underlying business problem is named.
The tell: if you cannot articulate, without hedging, what specific outcome AI is supposed to produce for your business, you are responding to pressure, not to a problem. Outcomes that pass the test sound like “we have customer-service costs growing faster than revenue and we want to flatten that curve.” Outcomes that fail sound like “our board wants us to have an AI strategy.”
The right response to board pressure is not a hastily scoped pilot. The right response is sometimes a board-conversation reset. We have helped CEOs run that conversation, and the result is almost always that the board respects a clear “here’s why we are pacing this carefully” more than another underbaked initiative. Boards want results, not stories. Give them the structured pacing and they give you the runway.
This is not advice to ignore the board. It is advice to bring the board a coherent answer rather than a fast one. The fast answers are how the 95% pilot failure rate gets produced.
Why “not yet” is the highest-trust answer
Saying “not yet” is structurally harder for an assessment vendor to do than saying “ready.” A “ready” answer keeps the engagement alive. A “not yet” answer means the next conversation is about pausing the spend, which most vendors are not commercially incentivized to recommend.
That is why the answer carries weight when it does come. A diagnostic that can say “not yet, here’s the order of operations” is operating in your interest, not theirs. The data backs it up: businesses that paused before clearing the five signals had a meaningfully higher chance of joining the 5% later than businesses that pushed through.
This is also why our AI Readiness Assessment is structured to surface “not yet” cleanly. About one in three businesses that take it score in a band where the right next move is not AI. That output is uncomfortable to deliver and useful to receive. We deliver it anyway, because the alternative is participating in a market dynamic that loses our clients money.
Related reading: What Is an AI Readiness Assessment? and AI Readiness vs AI Maturity cover the full diagnostic frame and the difference between this kind of starting-line read and a scaling assessment.
How to act on a “not yet”
If the diagnostic comes back “not yet,” the answer is not “give up on AI.” The answer is “fix the upstream conditions first, then revisit.” In practice, the moves usually fall in this order:
- Fix the highest-priority foundation gap. Usually one of: data integration, governance baseline, or a named owner on the initiative you were about to launch.
- Run an honest post-mortem on any previous AI work. One hour. Written summary. Real findings.
- Pick the one use case where the metric is sharp and the owner is named. Not three. One. Ship it to production in six to twelve weeks.
- Measure the result over twelve months. Real production, real metric, real number.
- Then revisit the readiness diagnostic. A business that has done steps 1-4 will score very differently than the one that started the cycle.
The order matters. The instinct after a “not yet” is to compress the timeline (“we’ll do all of those things in parallel while we run the next AI pilot”). The compressed version produces another failed pilot. The sequential version produces a paying-back initiative inside a year.
Frequently Asked Questions
How often does the assessment say “not yet”?
About one in three businesses that take ours land in the “not yet” band. The number is roughly stable across industries. It is highest in businesses under $20M (where foundations are most often missing) and in businesses with one or more previous failed AI pilots (where the same conditions remain in place).
Is “not yet” the same as “never”?
No. “Not yet” almost always becomes “ready” inside a year if the upstream work gets done. Data integration projects take three to six months. Ownership reassignment takes weeks. Post-mortems take an hour. None of these are quarter-long initiatives in their own right; they are just rarely sequenced in front of the AI work.
What if my board pushes back on the “not yet” finding?
We have helped CEOs run that conversation. The board pushback usually softens when the alternative is presented honestly: another pilot that will probably join the 95% versus a structured plan that gets the business into the 5%. Boards want results. They will take a coherent six-month plan over an immediate pilot, in most cases.
What does the “not yet” output actually look like?
A one-page scorecard with the six readiness dimensions graded, a position on a cohort curve, and a prioritized list of three to five moves that close the gaps. Designed to be readable in a leadership meeting in five minutes. No upsell required.
Can I just fix the gaps myself without an assessment?
You can. The assessment surfaces gaps you might miss and gives you a structured cohort benchmark, which is harder to build from scratch. The bigger value is usually the sequencing: knowing which gap to fix first when there are three of them.
Does this mean Armstrat won’t take AI work for businesses in “not yet”?
We will take the readiness work. Data integration, governance design, ownership reassignment, the post-mortem on the last initiative. That is the work that turns “not yet” into “ready.” We don’t take the AI implementation work for those businesses until the foundations are in place. That is the AI-integrated, not AI-first stance applied honestly. —
Take the AI Readiness Assessment
If you suspect your business might be in the “not yet” band, ten minutes of honest diagnostic is cheaper than the eighteen months of pilot work you would otherwise spend to find out. The AI Readiness Assessment scores all six foundations, places you on a cohort curve, and is built to say “not yet” clearly when that is the right call. —