Criteria helps companies get a return on AI. We find where the value is, build it ourselves, and stay until the people who have to use it do.
The problem
Almost every company now has AI somewhere. Far fewer have people who actually use it, and fewer still can say what it earned. Those two are the same problem: a system nobody uses cannot pay for itself. The causes are consistent, and none of them is the model. Automation pointed at a task nobody was complaining about. AI bought because the board asked for AI, rather than because a problem asked for it. Data that was not where the diagram said it was. Cost that escalates the moment it leaves the demo. Risk controls nobody owns. And a workflow nobody redesigned around the thing, so the people it was built for quietly went back to the spreadsheet.
0%
Of companies now use AI in at least one function. Only 39% can point to any measurable effect on earnings, and for most of those it is under 5%. McKinsey, State of AI, 2026.
0%
Generate no material value from their AI investment at all. Five percent create value at scale. BCG, 2026.
0%
Abandoned most of their AI initiatives last year, up from 17% the year before. The average company scrapped 46% of its proofs of concept before production. S&P Global, 2026.
The method
The first happens once. The other three repeat: you measure what landed, and the measurement sets the next roadmap. That loop is the part most programmes skip, which is why their AI never compounds.
Once, before anything else
We interview your executives and, separately, the people who do the work. Those two accounts rarely match, and the gap between them is where AI projects die. You get a company-wide picture of where the value is and a prioritised list of what to do first. If you are not ready, you hear it here.
Executive and stakeholder interviews · current AI usage · key functions and processes · data and technology readiness · skills and organisational readiness · company-wide opportunities · capability gaps · prioritised next steps
Per business area · start of every cycle
We take one area and go deep, mapping the workflow as it is really performed rather than as the process document describes it. You get named use cases, each with a value estimate, its dependencies and a success metric, in the order those dependencies allow. And an explicit list of what not to build.
Business and process discovery · workflow and pain-point analysis · use-case definition · value and impact estimation · feasibility and complexity · prioritisation · business cases · requirements and dependencies · success metrics and KPIs · phased roadmap
Per use case
We build and ship it ourselves, integrations and data connections included. It goes live in shadow mode first, touching nothing, so you can compare its judgement against your team's on real cases. Then challenger, then champion, each move only once the numbers earn it. The code, the prompts and the evals stay yours.
Solution design · automation and workflow development · AI agents · custom AI applications · system and API integrations · data connections · human-in-the-loop workflows · testing and evaluation · production deployment · monitoring and iteration
Every cycle · then back to the roadmap
We train the people who have to live with it, inside their own work rather than in a webinar, and equip the champions who answer the questions once we have gone. A system nobody uses has no value, so adoption is scoped as work. Then we measure against the metric from stage 02, and that sets the next roadmap.
Training on what we built · role-specific training · hands-on exercises using real work · new workflows and ways of working · human-AI collaboration practices · onboarding materials · AI champion identification · train-the-trainer · adoption feedback and metrics
The system watches and records. Your team decides. Nothing changes.
The system proposes. Your team overrules freely. We measure the gap.
The system decides within agreed bounds. Only once the numbers earn it.
Who this is for
Whether you have not started with AI or you already have it running and cannot say what it earned, this is for you.
Where the method comes from
Years of experience in technology came down to one method.
This is not a framework we drew up and then went looking for companies to apply it to. It is what is left after years of building technology inside companies, and of being measured on what it changed rather than on what it shipped.
Every stage is there because we watched something fail without it. Systems built well on data that could not carry them. Roadmaps that opened with the use case that demoed best rather than the one whose dependencies were ready. Automations switched on rather than proved. Good systems going unused while the reporting stayed green.
Four stages is what survived. Anything that did not earn its place in engagement after engagement is not in there. The people who arrived at it are the people who show up.