The uncomfortable data on AI strategy
Start with the numbers, because they reframe everything. Adoption is not the problem anymore. Almost everyone has the tools. What almost nobody has is a return.
95%
of organisations deploying generative AI saw zero measurable return
That is not a typo. MIT's 2025 study of more than 300 AI initiatives found 95% produced nothing measurable. McKinsey's State of AI 2025 report tells the same story from the other side: 88% of organisations use AI, but only about 6% are high performers capturing 5% or more of profit from it. Gartner put the success rate of AI use cases at roughly 28%. RAND found around 80% of enterprise AI projects fail to deliver the value they promised.
So the honest starting point for any AI strategy is this: the base rate is failure. If you do what most businesses do, you will most likely get what most businesses get, which is a quieter conscience and no new revenue.
AI adoption is near-universal. AI returns are rare. The gap between the two is the whole game.
Why most AI strategies fail
Dig into the same research and the cause is consistent. It is not the models. It is not a lack of tools. It is that businesses bolt AI onto the way they already work instead of changing the way the work flows.
21%
of organisations have redesigned any workflow for AI. Workflow redesign is the factor most correlated with real financial impact
Read that twice. The one thing most linked to getting paid by AI is redesigning the workflow, and only one in five businesses has done it. The other four out of five are pointing a powerful new tool at an old process and wondering why nothing changed.
For a knowledge business, this matters even more, because your most valuable workflow is not a process diagram. It is your judgement. The way you diagnose a client, sequence a program, and decide what good looks like. If that is not extracted and structured, any AI you deploy is guessing. It produces plausible output that misses your standard, which is exactly the zero-return trap the data describes.
What the 6% do differently
The high performers are not using secret models. They do four things the rest skip. They redesign the workflow rather than decorate it. They aim at growth, not just cost-cutting. They scale the one thing that works instead of piloting twenty. And they treat their own expertise as the asset to be encoded, not a nice-to-have.
Translated into a plan you can actually run, that becomes four questions every AI strategy has to answer before a single tool is chosen.
- Where am I the bottleneck? The one place your business most often waits on you. Not ten places. One. That is where the first build goes
- What decision am I actually making? Name the judgement you apply that the team cannot. That judgement is the thing AI has to learn, and it has to be documented first
- What does good output look like? If you cannot describe your quality standard, no system can hold it. Write down the criteria you check against before you sign anything off
- What happens to the time I free up? A strategy with no plan for the reclaimed capacity just gives you a quieter inbox. Decide what the freed time is for before you start
The framework: Explore, Map, Transform
This is the method we use, written so you can run it yourself. Three phases. Each one has a concrete output. Nothing moves forward until the previous phase is solid.
Explore. Find where you are the bottleneck. Audit your delivery and mark every place the business waits on you. Pick the one with the highest payoff. Output: a single, named constraint to attack first.
Map. Extract the IP behind that constraint. The frameworks, the diagnostic questions, the decision rules. Get it out of your head and onto the page in enough detail that someone else, or a system, could apply it consistently. Output: a structured description of how you actually think.
Transform. Build one AI system on that structured IP, test it on real work, and only extend it once it holds your standard. Output: a working prototype and a roadmap for what comes next. This is the orchestration layer, where the IP becomes a system rather than a document.
Where to start this week
You do not need a consultant to begin. You need a week and the discipline to look honestly at where your time goes. Here is the exact sequence.
- Log every task you personally touch for one week. Mark each one: only you (Y) or could be a system (N)
- Pick the single N task that costs you the most time or most often blocks the team
- Document how you do it. The questions you ask, the order, the decision rules. This is your IP, written down
- Build one small AI system on that documented process. Test it on real work, not a demo
- Only once it holds your standard, extend it. One workflow at a time, never all at once
Do that and you will already be ahead of the 80% who are bolting AI onto a process they never examined. The framework is the same whether you run it alone or with help. What changes is the speed and who does the extraction while you keep running the business.
If you want a shortcut to the diagnosis, the IP Monetisation Assessment maps your delivery against the twelve factors that decide whether your IP is ready to be built into an AI system, and tells you where to start. For the full done-with-you version, see AI consulting services.