Most of the failed AI projects we get called into did not fail because someone chose the wrong model. They failed months before the first line of code was written, because the organisation could not export its own data, could not describe the process it wanted automated, or had nobody with real hours to own the work. Industry surveys put AI project failure rates somewhere between 70 and 85 percent. In our experience, most of those failures were predictable on day one, and the prediction had nothing to do with the technology.
Predicting them does not require a consulting engagement. It requires one honest afternoon. What follows is the assessment we effectively run on the first call with every prospective client, written up so you can run it yourself: five dimensions, each scored 0 to 2, for a total out of 10. Score it with the people who actually do the work, not just the leadership team, and resist the urge to round up. Rounding up on this test is how companies end up funding a pilot the organisation cannot digest.
Dimension 1. Data: where does it actually live?
Every useful AI system is downstream of the data you can put in front of it. So the first question is not what AI can do; it is what you have. Three sub-questions:
- Where does it live? One system? Nine? A shared drive of Excel files named
final_v3_ACTUAL.xlsx? Every extra system adds friction, and every spreadsheet maintained by exactly one person is a single point of failure. - Can you export it? Do not answer from memory; try it this week. Ask for a full export of customers, orders, or tickets from your main system. If the answer involves a vendor support ticket, a fee, or the phrase "we would have to check", write that down.
- Is there one source of truth? If a customer's phone number exists in three systems with three different values, which one is correct? If nobody can answer that, a model cannot either.
Score 0 if data is scattered across spreadsheets and inboxes and exports are painful. Score 1 if core data lives in one or two real systems but has known duplication and quality issues. Score 2 if there is a clear source of truth, exports are routine, and a named person is responsible when the data is wrong.
Dimension 2. Process clarity: can you write the workflow down?
Pick the workflow you most want to automate: invoice matching, lead qualification, support triage, quote generation. Then run a cheap test. Ask two people who do it daily to write down the steps independently, including what they do with the exceptions. Compare the documents.
If two employees describe the same workflow differently, AI will not resolve the disagreement. It will automate it.
This is the dimension leaders overestimate most. From the corner office a process looks like five clean steps; at the desk where it happens there are eleven steps, four undocumented exceptions, and a rule that exists only in one veteran's head. None of that blocks an AI project, but all of it has to be surfaced and written down first, and the writing down is work the organisation, not the vendor, has to do.
Score 0 if the workflow exists only in people's heads and the two write-ups disagree on the basic sequence. Score 1 if the happy path is documented but the exceptions are folklore. Score 2 if the process is written down, the write-ups broadly match, and exceptions have named handling rules.
Dimension 3. Integration surface: can software talk to your systems?
An AI system that cannot read from and write to your existing systems is a demo. Before any build, we ask which systems the target workflow touches and whether each one has an API. The honest news for most mid-sized businesses: modern SaaS tools almost always do; older on-premise systems sometimes offer direct database access instead, which is workable; and a desktop-only ERP with no API and no export beyond PDF is a genuine constraint that shapes the whole project.
You do not need perfect APIs everywhere. You need at least one workable path into each system the workflow touches. We covered the full menu of options, including what to do when there is no API at all, in our guide to adding AI to existing software without a rewrite.
Score 0 if the core systems are closed: no API, no database access, no scheduled exports. Score 1 if some systems have APIs but the most important one does not. Score 2 if every system in the target workflow has an API or an equivalent programmatic path.
Dimension 4. People: an owner and a team that is not hiding
Two separate questions here. First: is there an internal owner who will spend real hours on this? Not a sponsor who approves budget. An owner who answers questions, chases access credentials, reviews outputs weekly, and makes the call when edge cases show up. In the pilots we run, this person spends four to six hours a week for the first two or three months. If nobody can commit that, the project stalls no matter who builds it.
Second: how does the team feel? Staff who see the project as a threat go quiet in exactly the way that kills it: they stop volunteering the exceptions, the workarounds, the ugly truths about how the process really runs. Staff who are curious hand you the edge cases for free. Which one you get is set by how leadership frames the project, and it is set early. "This will handle the repetitive 60 percent so you can do the judgment work" lands very differently from silence plus rumours.
Score 0 if there is no named owner and the team learned about the AI plan through the grapevine. Score 1 if there is an owner with limited hours and the team is wary but talking. Score 2 if there is an owner with committed weekly hours and the people doing the work helped choose the target workflow.
Dimension 5. Expectations and budget
A pilot-sized budget attached to a measurable target beats a moonshot mandate every time. The best readiness signal we see is a leader who can say: cut first response on support tickets from four hours to thirty minutes, and here is a pilot budget to prove it. The worst is a transformation mandate with a big number and no metric attached. Big budgets without metrics produce demos, steering committees, and slideware.
Budget realism matters too. The meaningful custom pilots we scope usually land in the low tens of thousands of dollars, not the hundreds; if the number in your head is wildly above or below that range, recalibrate before you start. We broke down where the money actually goes in our guide to AI development costs.
Score 0 if the goal is "become an AI company" with no workflow or metric attached. Score 1 if there is a target workflow but no agreed success metric or budget. Score 2 if there is one workflow, one measurable target, and a pilot-sized budget someone has already approved.
Reading your score
Add up the five numbers. The total matters, but so does the shape: a zero on data or process clarity caps what any pilot can achieve, whatever the other dimensions say.
- 0-3: Not ready, and knowing that just saved you money. Do not start an AI project; the same budget spent now buys frustration and a stalled pilot. Spend a quarter on foundations instead, starting with data, which is almost always the binding constraint at this band.
- 4-6: Pilot-ready. Pick exactly one workflow, define one metric, and run an 8-12 week pilot at the smallest useful scope. Do not sign an enterprise AI platform deal, and do not run three pilots in parallel. The organisation's attention is the scarce resource, not the technology.
- 7-10: Scale-ready. Rarer than you would think; perhaps one in ten of the businesses we assess. Your constraint is prioritisation, not readiness. Rank candidate workflows by hours saved times frequency, divided by integration difficulty, and work down the list.
Red flags that predict failure
Separate from the score, a few signals correlate with failure so strongly that we treat any one of them as a stop sign:
- An AI strategy with no workflow named. If the strategy deck says transformation and never names a single process, there is no project yet. There is a mood.
- Nobody owns data quality. When bad records are everyone's problem, they are no one's, and every AI output inherits them. A system grounded on data nobody vouches for produces answers nobody trusts.
- Headcount replacement expected in month one. Realistic AI systems assist first and earn autonomy over months. A business case that only works if people leave the payroll in week four is a business case that fails, and it poisons Dimension 4 on its way down.
- The project exists because the board asked about AI. A defensible answer to a board question is a strategy memo, not a build.
Where to start if you scored low
Low totals almost always cluster on the same dimension: data. That is good news, because data work is the one investment that pays off whether or not you ever build AI on top of it. Consolidating the spreadsheets, deduplicating the customer records, and standing up one queryable source of truth improves reporting and daily operations immediately. It is also the single highest-ROI move on this list, precisely because it is unglamorous and everyone has been putting it off. We wrote about the mechanics in turning legacy data into business intelligence.
The decision rules, in short:
- Scored 0-3: spend the next quarter on data consolidation and on writing down your top workflow. Re-run this assessment afterwards; most businesses jump three or four points.
- Scored 4-6: pick one workflow, run the two-employee write-up test, fix the disagreements, then pilot against a measurable target.
- Scored 7-10: stop assessing and start ranking use cases. You are losing more to delay than to risk.
And if you want a second pair of eyes on your scoring, send us your five numbers. Telling a business it is not ready yet is the cheapest advice we give, and we give it often.