Written for the leaders of professional-services firms – the managing partners deciding how their own firms should adopt AI. The “you” in what follows is them. It appears here because it’s also the argument behind how this firm is built.
KPMG’s latest global survey of more than 2,000 senior leaders found that organizations with clear accountability for their AI decisions are three times more likely to report a return on them. Where the CEO is on the hook, the gap is even larger – 14% of those firms report established ROI versus 4% where the chief executive isn’t accountable. And 57% say AI is delivering meaningful business value, versus 21% of the rest.
It’s a tidy statistic, and it will travel through slide decks for the next year. But the number is a symptom. The thing worth understanding is what causes it – because once you see the mechanism, staying personally involved stops being a leadership platitude and becomes an operational necessity.
Those numbers support what we’ve suspected all along: the decisions that make AI pay off are not technology decisions. They’re organizational-design decisions – and those can’t be handed to your IT team. They require active involvement at the most senior levels.
Ethan Mollick stated it clearly: these are “increasingly organizational design and strategy decisions, not IT choices.” So the questions that actually move the needle aren’t “which model?” or “how do we write better prompts?” They’re questions like “which decisions are we willing to let AI influence?”, “what must a human verify before we rely on it?” and “who owns the answer when it’s wrong?” A CIO can stand up a tool. But only the people who own the firm’s risk, its economics, and its client relationships can answer those questions – and when leadership steps back, they don’t get answered by someone else. The firm might buy the tool, but it doesn’t make the decisions that would have made it worth buying.
That’s why a second finding in the same research matters as much as the first: in most organizations, accountability for AI is diffuse – spread across a CEO, a committee, a few business leads, no one in particular. Diffuse accountability isn’t a softer form of ownership; it’s the absence of it. You can’t track impact you don’t own, and you can’t defend a decision no one made. The firms getting this right are doing something very basic, but critical – naming who is accountable for which AI-influenced decisions. One company – Lovable – assigns every internal AI agent a named human “parent,” the person responsible for keeping it honest. That isn’t bureaucracy. It’s the difference between a system you can answer for and one you’re hoping holds.
And timing matters. AI vendors, employers, and the habits of a generation of workers are hardening into expectations that will be hard to reverse, and firms have a window – maybe eighteen to thirty-six months – to define their own defaults before generic ones define them. The firms that use the window choose; the firms that don’t inherit choices made for them. And “inherit” is generous. What actually happens when leadership is absent is that the work goes on anyway – your people are already using AI, just without you, increasingly building unsanctioned tools on top of it that security teams now call “attack vectors” (each new integration is estimated to multiply the attack surface 3–5×). So-called shadow AI isn’t a discipline problem. It’s a governance problem.
This also supports what I preach – governance is not the brake many believe it to be, and adoption is not the gas. Instead, as PwC found in its 2026 AI Performance Study, automation is enabled by trust, not constrained by it. Low-trust firms literally cannot delegate decisions to AI, so they never capture the gains. Clear lines don’t slow a firm down; they’re the only thing that lets it safely speed up. The leader who defines what AI may touch, what it may not, and who verifies the output isn’t policing adoption – they’re the reason ROI-generating adoption is possible.
For professional-services firms, the stakes are especially high, because the work carries your name. The partner who signs the opinion or the deliverable is personally accountable for it – to the client, the insurer, the regulator, the malpractice carrier. You can buy software through procurement, but you can’t procure accountability for a judgment that bears your signature. So in a law or accounting firm, “who owns the AI decision” isn’t something you can delegate down or buy from a vendor. It’s the same question as “who owns the work” – and you already know the answer to that one. At my firm, I own AI governance personally – because as important as every other part of the build is, none of it matters if our work product comes out of a system known to confidently say the wrong things.
None of this asks a managing partner to become technical. It asks for something harder and more familiar: to own the decisions. Map who is accountable for which AI-influenced calls. Decide, explicitly, what AI may touch, what’s restricted, and what’s reserved for human judgment. Define what must be verified, by whom, before anyone relies on a machine – and how you’ll keep an audit trail of those calls. Then ask the question that has always separated firms you can trust from firms you can’t: if a client, the management committee, an insurer, or a regulator asked what happened here – what would you say?
The 3× isn’t magic. It’s what happens when the person accountable for the work is also accountable for the machine doing it. That’s not a technology upgrade. It’s a leadership one – and it’s the one nobody can make for you.