Six layers, eighteen questions, one score. Ninety minutes tells you where your AI programme is genuinely blocked, rather than where a vendor last demonstrated something. The second half is the build guide for whichever layer came out lowest.
Gartner projects $644bn of generative AI spending this year, and expects more than 40% of agentic projects to be cancelled by the end of 2027. Both numbers are real and they are not in tension. The dividing line between them is almost never model quality.
This finds which side of that line you are on, by looking at the six layers underneath the thing you actually bought.
Your capability is set by your weakest layer, not your average.
A total out of 54 is close to meaningless. An organisation scoring 8, 8, 8, 8, 8, 1 is not "mostly ready". It is blocked, and every pound spent on the five strong layers is currently buying nothing.
The most expensive mistake available here is spending on the layer you find interesting rather than the layer that scores lowest. This page computes the minimum for you and then points you at the one guide that matches it.
You can click through this alone in five minutes and get a number that flatters you. The version that works takes three people and ninety minutes.
One person from IT or engineering, one from operations, one from finance or risk. Three is the minimum for an honest answer, because each of them sees a different part of the problem and none of them sees all of it.
Ninety minutes, or two sessions of forty-five. Longer than that and you are debating rather than scoring.
For layer 2, have each person answer separately before you compare. The disagreement between them is the finding. If everybody agrees the data is fine, you have asked people who do not touch it.
Three questions per layer, scored 0 to 3, with every anchor written out so this is not a vibe check. Nine points per layer, fifty-four in total. Open a layer and answer. Your score updates as you go.
Read the minimum, not the total. The button below opens the one guide that matches your weakest layer, and that is deliberately the only one you should be reading first.
One per layer. Concrete enough to start on Monday, and ordered so each step is worth doing even if you never reach the next one. Start with your lowest-scoring layer and genuinely ignore the rest until it moves.
Gartner splits generative AI strategies into three, with rough costs, excluding data transformation.
Ambition and cost are the same decision, and it is usually made once, quietly, before anyone runs the numbers. The common mid-market failure is running an upend budget while calling it an extend project.
File it under IT and it becomes a tool to procure and roll out. File it under workforce and it becomes a colleague to scope, onboard and manage. Same technology, entirely different questions, and only the second set gets you anywhere.
Antonio Neri at HPE put it plainly: an agent is no different from any other employee you have to hire.
Two things, and the second one is the uncomfortable one.
Score the six layers with three people in the room. Find the minimum, and resist the urge to average it away.
Pull the real cost of your most active AI pilot. The full one, including change management, data preparation and tokens, not just the licence line. Then name the single number it is supposed to move.
You already know which side of the cull it is on.
Much smaller job, and most of it is free. The companion guide builds the same architecture for one person across two weekends, with the six layers cut to five and the people layer dropped, because there is only one of you.
Open the personal guide