Answers
Plain answers to the questions I'm asked most about AI in business. Each one starts with the short answer, then the detail and where to read more.
the short version first
Getting AI to pay
Why do AI pilots fail to reach production?
Most AI pilots stall because they were built against the documented process, not the real one. When they meet real data, exceptions and handoffs, costs rise, value is hard to show and nobody can say who owns the risk. Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027.
How do you measure the ROI of AI automation?
Measure it in EBIT and cash, not hours saved. Put every benefit into one of four boxes: cost out, cost avoided, leakage stopped or cash accelerated. Agree the baseline before anything is built, give each number a business owner, and check the case still works if only half the benefit arrives.
What is process mining, and why should it come before AI?
Process mining reads the event logs your systems already keep (case, activity, timestamp) and reconstructs how work really flows: every route, how long each step takes and where work waits. It should come before AI because it shows the real process, sets the baseline and often finds cheap fixes that need no technology.
Governing AI and agents
What is governed AI?
Governed AI is AI that runs inside clear limits, with a named owner, human approval at the points where risk calls for it, and a record of every decision it takes. The aim is AI you can explain to a board, an auditor or a regulator after the fact, not just AI that works in a demo.
What should a board ask before approving an AI project?
Five questions: who owns it, what can it do on its own, what's the number, where's the record, and what happens when it's wrong. If any answer is 'we'll work that out later', the project isn't ready to fund. I call this the Working Test.
What should an AI audit trail record?
For every action, an AI audit trail should record what the system saw, what it decided and the alternatives, which rule or policy permitted it, who owns that rule, and whether a human checkpoint applied and who cleared it. It should be written at the moment of action and readable by a compliance officer, not just an engineer.
What does 'human in the loop' mean for AI agents?
Human in the loop means a named person must review or approve certain AI decisions before they take effect. It works when the trigger points are defined in advance (by value, risk or uncertainty), the person sees the full history, and their decision is recorded. It fails when a person is asked to rubber-stamp everything.
When do the EU AI Act's high-risk rules apply?
After the Digital Omnibus changes, obligations for stand-alone high-risk AI systems (Annex III) apply from 2 December 2027, and for AI built into regulated products (Annex I) from 2 August 2028. The deadlines moved; the duties, including human oversight and record-keeping, did not.
Buying AI
What is agent washing?
Agent washing is relabelling an existing chatbot, assistant or workflow tool as 'agentic AI' without it having real agent capabilities. Gartner coined the term and estimated that only around 130 of the thousands of vendors claiming agentic AI actually offer it.
What is the difference between agentic AI and RPA?
RPA robots follow a fixed script exactly the same way every time, which makes them predictable and easy to audit. AI agents reason about a goal, handle unstructured input and exceptions, but cost more and won't always decide the same way twice. In regulated businesses the best designs use agents for judgement and robots for execution.
How do you choose an AI automation partner?
Choose the partner who will agree a measured baseline before building, show you the evidence their systems produce before the demo, put you in touch with customers in production rather than pilot, and tell you when a process isn't worth automating. Ask how they're paid and what leaving looks like.