Should We Still Think in KPIs and Functions in the Age of AI?

In my last post I suggested KPIs per function: customer service, sales, finance, IT and so on. While writing it I kept having an uncomfortable thought. What if the frame itself is the problem? What if we are measuring a new kind of work with an org chart designed for the old one?

I don’t have a finished answer. This is a set of exploratory thoughts, and I expect to be wrong about some of them.

Faster, not different

Deloitte’s State of AI in the Enterprise (29 October 2026) says 81% of companies see faster decisions and execution, but only 25% are deeply changing how the business works. Deloitte names organisational change, the capacity to redesign work and governance, as a key blocker.

That matches what I see. Most AI is bolted onto existing functions. Customer service gets a chatbot, finance gets invoice matching, IT gets a coding assistant. Each function optimises its own KPI. Each gets a bit faster. And the flow the customer actually experiences, from question to order to delivery to invoice, often looks the same.

Functions vs flows

A function owns a step. A customer experiences a flow. When work was done by people, splitting it into functions made sense: specialisation, clear managers, manageable teams.

Agents don’t care about department lines. An agent can read the support case, check the order, see the stock situation and draft the credit note. If that work is split across four functions with four KPIs, the agent will be configured four times, owned by nobody end-to-end, and every handover becomes a place where quality leaks.

So one idea is to measure and own outcomes across the flow: time from customer problem to solved problem, cost per correct delivery, share of orders that go through without a human touching them. Functional KPIs remain, but as diagnostics, not as the goal.

Humans and agents in the same team

G2’s 2026 AI Agents Report found that 45% of buyers lose more than a quarter of the expected gains to oversight and rework, and that many organisations manage their agents informally, without clear ownership. That is an organisational problem more than a technical one.

If agents are part of the work, they should be part of the team: with a named owner, a clear scope, and the same expectation of quality as anyone else. A team of five people with a set of well-run agents may handle what used to take a larger department. I think smaller teams with agent capacity will become normal, not because people are removed overnight, but because new work won’t be staffed the old way.

Roles move towards orchestration, judgement and accountability

When I led a team of 18 consultants at EVRY, and later as co-owner at XLENT, much of the leadership work was about matching people to problems and making sure someone owned each result. Coaching hockey teams taught me the same thing in a different setting: a team works when roles are clear, everyone knows who makes which call, and the structure supports the play rather than the other way around.

With agents, I think more human roles shift towards:

• Orchestration: deciding which work goes to which agent or person, and designing the flow.
• Judgement: handling the exceptions, the ambiguous cases, the things that need context or empathy.
• Accountability: owning the outcome, including when an agent gets it wrong.

Deloitte notes that 40% of enterprises let AI execute decisions without human approval at least to a moderate degree, while nearly one in four say accountability is unclear when AI produces a negative outcome. That gap is the design problem.

Decision rights become explicit

In a traditional organisation, decision rights are often implicit: they follow the hierarchy and the habits. With agents they need to be written down. What may the agent decide on its own? Up to which amount, which risk, which customer category? When must it ask, and whom?

This is also where my dashboards-to-alerting experience feels relevant. At a global manufacturer, moving from dashboards to alerts with drill-down forced us to decide who should act on what, at which threshold. That is decision rights in miniature. Organisations with agents need the same thing at scale.

Governance as design, not as an afterthought

GDPR, the EU AI Act and security requirements are often treated as a review at the end. With agents acting across systems, that doesn’t work. Even with parts of the AI Act timeline adjusted through the Digital Omnibus, requirements such as AI literacy and transparency already apply.

I’d rather see governance built into the flow: what data each agent may access, logging of what it did, clear human escalation, and someone who can explain a decision afterwards. Done well, that is not a brake. It is what allows you to give agents more responsibility, because you can trust and verify what they do.

What to remove

Perhaps the most underrated question: what should disappear?

• Handovers that only exist because work was split between departments.
• Reports and dashboards that nobody acts on.
• Approval steps that never stop anything.
• KPIs that reward a function for pushing work onto the next one.
• Systems kept alive only because a manual process depends on them.

If the organisation of tomorrow just adds agents to the organisation of today, we get the double cost I wrote about earlier. If it removes things, we get room to work differently.

So, KPIs or not?

I think we still need measurement, probably more than before. But I suspect the KPIs that matter will increasingly describe flows and outcomes, owned by small cross-functional teams of people and agents, rather than functions defending their own numbers.

I don’t know exactly what that organisation looks like. I’m fairly sure it isn’t the org chart we have today with a chatbot in every box. A good way to start might be modest: pick one flow, give it one owner, measure it end-to-end, and see what you can remove.