Humans + AI: The Operating Model

By Ross Dawson, building on "Augmenting Human Capabilities When All Work Becomes Humans + AI."

Almost every organisation now uses AI. Almost none has an operating model for it. The difference between those two statements is where the next decade of competitive advantage will be won and lost.

What most organisations have done is adoption: they have given people powerful tools and told them to be more productive. The tools are extraordinary, the individual gains are real, and yet at the level of the organisation the promised transformation keeps failing to arrive. Output gets faster but not better. Everyone is busier and the work is strangely more uniform. The advantage that was supposed to compound somehow does not.

The reason is that an operating model is not a set of tools. It is the system by which an organisation makes decisions, allocates work, builds capability and learns. When you introduce a genuinely new kind of participant into that system — one that can reason, draft, analyse and recommend at a scale no human can match — and you leave the system itself unchanged, you do not get a transformed organisation. You get the old organisation, slightly faster, with new risks. Designing the Humans + AI operating model is the actual work. This is what it involves.

Start from the decision, not the tool

The unit of an operating model is the decision. Organisations are, at bottom, machines for making decisions — about customers, products, prices, people, risk and direction — and the quality and speed of those decisions is what separates the great from the mediocre.

So the right question is not "where can we use AI?" It is "how is this decision made, and how should humans and AI each contribute to making it better?" Those are profoundly different questions. The first leads to a scattering of tools bolted onto existing tasks. The second leads to a redesign of the decision itself: what AI should surface, what humans should weigh, where the handoff happens, and who is accountable for the outcome.

When you start from the decision, you quickly discover that not all work is the same, and AI's role should not be the same across it. This is the core of the operating model: matching the mode of human-AI collaboration to the nature of the work.

The core patterns of Humans + AI work

When you start from the decision rather than the tool, a small set of patterns covers almost everything. I set these out in a framework on the core patterns of agentic AI for organisation design, because leaders kept asking for a vocabulary to plan with rather than another catalogue of use cases.

There are three mechanisms by which AI creates value, and they are not interchangeable. The first is automation: the direct substitution of a specific human function with a digital worker. The second is productivity: optimising who — human or agent — performs each task based on their comparative advantages, rather than defaulting either to the person or the machine. The third, and the one most organisations underuse, is augmentation: cognitive support that expands what a human can do rather than replacing them. The evidence is now consistent that output per worker rises on the order of 10 to 40 percent where AI augments human decision-making, and barely moves where AI is bolted on as a standalone tool. Augmentation is not the soft option. It is the yield argument. Most of the value being left on the table today comes from organisations reaching for automation when augmentation would compound, and treating productivity as a headcount question rather than an allocation one.

On top of those mechanisms sit two collaborative workflow patterns, and the difference between them is where accountability lives. In the Humans + AI sandwich, a human frames the task with intent and context, AI executes, and a human reviews and refines the output — the person owns both ends. In AI with humans-in-the-loop, AI performs the primary task and humans provide approval at the critical junctures — the machine owns the middle, the human owns the gate. The operating model's job is to choose the pattern deliberately for each kind of work and make the choice explicit, rather than letting it default by accident to whatever the tool vendor shipped.

The mistake organisations make is to apply one pattern everywhere — usually crude automation, treating AI as a faster typewriter — or to drop AI into a consequential workflow with no clear pattern at all, so that no one can say where the decision is actually made.

Recursive, self-improving decisions

The deepest source of advantage in a Humans + AI operating model is not any single decision. It is the loop. Organisations that build systems to learn from their decisions — at every level, continuously — will compound their advantage at a rate that organisations making one-off decisions cannot match.

This is the recursive heart of the model. Every consequential decision is an opportunity to learn: what was decided, on what basis, and what actually happened. Most organisations throw that information away. The decision is made, the outcome arrives months later disconnected from the decision that caused it, and no learning is captured. A Humans + AI operating model closes that loop deliberately. The reasoning behind a decision is recorded. The outcome is connected back to it. AI surfaces the pattern across thousands of such loops, and humans interrogate the pattern to improve the next decision.

Done well, this turns the organisation into a system that gets better at deciding the more decisions it makes — a recursive, self-improving engine. It is also where humans and AI are genuinely complementary rather than merely coexisting: the machine sees the pattern across the volume, the human brings the judgment, the context and the accountability that the pattern alone cannot supply. The loop is the asset. The decisions are just the raw material it learns from.

The homogenisation trap

There is a danger at the centre of all this, and the research now names it precisely. A Johns Hopkins and MIT study of more than two thousand three hundred participants found that human-AI teams outperform human-only teams on productivity and, at the same time, homogenise their output. Both things are true at once. When everyone draws on the same models, prompted in similar ways, the work converges, and the variance that is the raw material of genuine insight, of differentiation, of breakthrough, gets quietly compressed toward a competent average. The study's real lesson is about design: where AI sits in the creative process determines whether diversity is preserved or destroyed.

For an individual task, homogenisation is invisible and even welcome — the output is good, fast and clean. For an organisation, sustained homogenisation is an existential risk, because competitive advantage lives in the tails, not the mean. An organisation whose every analysis, pitch, design and strategy regresses toward the same model-shaped centre is an organisation that has traded its distinctiveness for speed without noticing the trade.

The operating model must therefore protect variance on purpose. That means using AI to expand the range of options considered rather than to converge on the first plausible one. It means valuing and rewarding the human contributions the model cannot make — the contrarian read, the contextual knowledge, the taste, the willingness to reject the competent answer in favour of the better one. And it means designing collaboration so that the machine does the work of breadth and the human does the work of depth and difference, rather than both collapsing toward sameness. The goal is to keep people sharper, not more artificial. An operating model that makes everyone more average is working against the very thing it was supposed to create.

Designing the work: roles, decision rights and workflows

Concretely, three things have to be redesigned, not bolted on.

Roles change because the value of human work moves. The work that AI does well — recall, drafting, first-pass analysis, breadth — stops being where humans add value, and the work that remains human — judgment, relationship, accountability, the consequential choice, the creative leap — becomes the whole of the role rather than a part of it. Job design that still rewards people for the volume of the work AI now does is job design pointed at the past. Roles should be rewritten around the contributions that remain distinctively human, with AI fluency assumed rather than celebrated.

Decision rights change because a new participant has entered the decision. For every consequential decision the operating model must be explicit about what AI contributes, what the human decides, and crucially where the authority and the accountability sit. The failure mode is ambiguity: the human assumes the machine has it handled, the machine has no accountability to assume, and the decision falls into the gap. Accountability cannot be delegated to a model. The human owns the consequential decision and its consequences, full stop, with AI as the most powerful adviser they have ever had. As the systems become more agentic, transparency has to be built into the structure itself — so that you can always see where a decision was made, what authorisations were given, and follow the audit trail. Delegating to an agent without that structure is not delegation. It is abdication.

Workflows change because the handoffs between human and machine are where value is created or destroyed. A workflow that routes every decision through a human review that no one has time to do well creates oversight theatre. A workflow that automates a consequential decision with no point at which a human can intervene creates unaccountable risk. The design work is to place the human exactly where their judgment matters most, give them the information and authority to exercise it, and let the machine carry everything else.

Capability: keeping people sharper, not more dependent

An operating model is only as good as the people running it, and there is a real risk that as AI does more of the cognitive work, the humans atrophy — losing the very judgment the model relies on them to provide. An organisation of people who can no longer evaluate what the machine produces is not augmented. It is hollowed out, and dangerously so, because it has lost the ability to catch the model when it is confidently wrong.

So capability building is part of the operating model, not an afterthought. People need to become fluent enough with AI to direct it well, and critical enough to challenge it — to know when the competent answer is the wrong one. That is a higher-order skill than either using the tool or doing the work by hand, and it does not develop on its own. It develops when the operating model deliberately keeps humans engaged in the reasoning rather than reducing them to approving the output, when it rewards the override that turned out to be right, and when it treats the cultivation of human judgment as a strategic investment rather than a training-budget line item.

Where to start

The temptation is to redesign everything at once. Do not. Choose one consequential, repeated decision that matters to the organisation — how you price, how you qualify a lead, how you assess a risk, how you prioritise a backlog. Map how that decision is made today. Assign the right pattern — automation, productivity or augmentation — to each part of it. Make the decision rights explicit. Close the loop so the organisation learns from each instance. Protect the variance that keeps the decision sharp. Then do the next one.

The organisations that win the next decade will not be the ones that adopted AI earliest or bought the most tools. They will be the ones that redesigned how they decide — deliberately, decision by decision, into an operating model where humans and AI each do what they do best and the whole system learns from every choice it makes. That is not a technology project. It is the defining design challenge of the era, and it belongs to leadership, not to the lab.