Five steps, in the order most clients take them
Every one of these ends with something written that a board can hold you to: which decisions the machines may make, who can stop each one, and how you would know if it went wrong. That is what you are buying: control you can show, rather than advice you have heard.
The Executive Read. A short anonymous survey of your executive team before anything else happens. What they can no longer do unaided, who would notice if a system were wrong, how often anyone has overridden one, and what would still be possible if it stopped. Bought on its own, and the cheapest way to find out whether there is anything here worth your time. How it works.
A second opinion on your AI strategy. Send me the strategy you have built and two working days later you have an independent written view: the assumptions I would challenge, the decisions that lack a clear owner, and the three things I would put back in front of the executive team. A fixed fee, agreed before you send anything, and most are four figures. I sign your NDA first if you want one. Send it.
Rules Before Tools. One day with the executive team. You leave with the organisation’s rules written in the room’s own words: the decisions AI may inform, recommend and execute, and the ones it may never own. The session.
The Drift versus Design review. Across a function or the whole organisation. A written view of where you are designing the relationship between your people and these systems, and where you are drifting into one. Four steps: diagnose, map, design, commit. The instrument.
Standing advisor. Monthly. The trusted second opinion as the decisions arrive, and a written brief each month on what has changed and what it means for you. A handful at a time, never more. How it works.
For boards: a briefing, a written view on one decision, or a non-voting seat held over a period. Board advisory
Assignments are scoped one at a time, and most are five-figure. Retainers are monthly, and I take on no more than three or four at once.
The five named instruments behind all of this, in the order they are used, are at the instruments. All five are published in full and none of them needs me to run it.
I have done this job
I am not an AI engineer. I have run, grown, bought and advised businesses with AI in them, which is the vantage point a leader needs.
At Avallain I was brought in to grow the business and put AI to work across it, in an organisation that already had strong tools of its own. That work included the acquisition of TeacherMatic, an early-stage company whose AI product was selling into schools one at a time, and I became its nominated director. Before that I founded EtonX, which prepared students for the world of work in markets already moving fast on AI, and was acquired by Eton College. At Quizlet I led international growth into more than 60 countries for a business whose thesis rested on what AI could do with its data. Since then I have advised early-stage AI-first companies I cannot name, sat on boards in the UK and elsewhere, and spent years in partnership conversations with businesses from China to the UAE, Europe and the US.
All of it came down to the same calls: what to let the technology do, what to keep, and who carries it when it goes wrong.
Two advisory conversations, and most organisations only have one
The technology conversation. What can we automate, which systems should we deploy, how do we implement them, and what efficiencies can we capture? Good questions, answered well by consultancies and internal technology functions, and this work does not compete with them.
The leadership conversation. What should we delegate? Where must judgement remain ours? Which capabilities do we protect? How should roles change rather than shrink? Who is accountable when a machine took part in the decision?
The second set rarely reaches an agenda on its own, because nothing fails while it goes unasked: the work ships, the output often improves, no control is breached. So it tends to surface after something has been lost, which makes it a poor fit for a risk register and a good fit for a leadership session.
Rules before tools, and what is out of scope
No Copilot rollout, AI roadmap, vendor selection, model choice, agent deployment or data architecture. Those are real disciplines, done well by firms built for them.
What comes first is the part they depend on: which decisions a machine may inform, recommend or execute, which stay human, who is accountable for each, and what the people must remain capable of, written down before another tool is bought. The argument was set out in Rules Before Tools in August 2025.
Scope is set out here because an advisory relationship that starts with a misunderstanding about it ends badly, usually about four months in.
Five questions, each ending in a decision
Judgement. Which decisions can AI recommend, which can it execute, and which must a human make? Most leadership teams have never written that list down, and writing it is usually the most useful hour of the engagement.
Capability. Where is AI removing the practice through which people become competent enough to supervise it? An organisation can lose the ability to check the work while its output is still improving.
Work design. When AI removes thirty per cent of a role, what happens to the other seventy, beyond banking the saving? Licences are the cheap part and do nothing on their own. Value, where it appears, comes from reorganising the task, which is a management decision rather than a procurement one.
Leadership. What must the executive team understand personally, rather than delegate to the CIO or an AI steering committee? A board that cannot interrogate its own AI decisions has delegated more than it realises.
Accountability. When an AI system contributed to a bad decision, who can explain why the organisation made it? A human in the loop becomes governance when you can say who, at what point, with what authority to stop it, and how often they actually disagree.
The rule underneath all five is an old one: however you produced this, you are accountable for it. AI has not sat in the meetings or read what you read the way you read it, and the work now arrives faster than the person signing it could have produced it, so the rule has to be said out loud. Its written form, decision class by decision class, is the Decision Quality Protocol: how each decision gets made with the machine in the loop, who owns it by name, and what quality control looks like.
Five ideas that give a leadership team language for this
Each has its own page and a dated first publication, and each is free to use without hiring anybody.
Drift versus design. How deliberate is this organisation’s adoption?
Capability debt. What human competence are today’s efficiencies using up?
Human at the start. Where does human judgement need to frame the problem before the machine begins, rather than review it afterwards?
The missing rungs. What happens to the pipeline when AI takes the junior work people used to become senior on?
Synthetic seniority. Are your people producing senior-looking output without acquiring senior judgement?
Why this person
A client, in her own words. “We brought Rahim in to help our EMEA People Operations leadership team think through AI and human judgement. He surveyed us beforehand, sixteen of eighteen answered, and he built the session from our own responses, so the whole morning was about us. I’d recommend him to any leadership team working out what should stay human.” Karina Govindji, Director, Head of HR UK, Google.
Operator. Twenty years building education technology before writing about it, including starting EtonX in China, with schools in Shanghai and across the country, and, earlier, living in Abu Dhabi and working for its government. A leadership team wants somebody who knows what a decision looks like from inside an organisation.
Researcher. SuperSkills is published by Kogan Page, so the thinking has been through somebody else’s editorial judgement. Behind it sits the research, all of it published before the first invoice.
Independent. The work is not attached to any technology, so there is nothing to sell you afterwards. Six of the seven engagements published at case studies produced further work after the session.
Who this is wrong for
An organisation that wants its AI strategy validated. This is most useful when the answer is allowed to be uncomfortable.
An organisation looking for a technology partner, a systems integrator or an implementation team.
A team wanting predictions about 2030. The argument is about decisions being taken this quarter, and the research is explicit about how weak the forecasting evidence is, including where that cuts against this position.
