An explainer for water, transport and energy leaders

AI is getting better at the tasks. Infrastructure changes when it gets better at the teamwork.

A new paper from Google DeepMind, Organizing Intelligence, asks why AI keeps passing expert-level tests while the economy looks much the same. The paper offers an economic model of possible futures; below, I apply its ideas to infrastructure, where delivery depends heavily on coordination.

Dial one

Competence

How good an AI is at a single task: running a hydraulic model, drafting a design report, pricing a scheme. Gains here change costs and pay gradually.

Dial two

Coordination

How well AI agents work together: splitting up a job, handing over without losing context, catching each other's mistakes. Past a tipping point, gains here change industries suddenly.

Darren Russell · 31 October 2026 · Based on Imas, Koh and Legg (Google DeepMind), 29 October 2026

Stop 1

The puzzle: brilliant AI, familiar programmes

Every few months a new AI model clears a bar that looked years away. Yet look at a capital programme in water, rail or power and most of it would be recognisable to someone from ten years ago: the same kinds of frameworks, supply chain tiers and delivery teams.

The authors argue we have been measuring the wrong thing. Almost every AI test checks one model working alone on one problem. Real work is done by organisations that combine people with different strengths and compete for work. An AI can beat every individual on a test and still lose to a well-built team.

£104bnWater company spending allowed in England and Wales for 2025–30 (Ofwat final determinations, December 2024)

1% a yearConstruction labour productivity growth over two decades, against 2.8% for the world economy (McKinsey Global Institute, 2017)

These figures set the scale of the opportunity. Neither comes from the DeepMind paper.

Stop 2

Why a team of specialists still beats a clever AI

People are jagged. A brilliant network modeller may know little about landowner negotiations. A site manager who can run a complex shutdown safely may never have built a cost model. In the authors' quiz of 100 people across six subjects, being good at one subject told you little about the others.

Good organisations put that jaggedness to work. They combine specialists so that the team covers ground no single person could. The paper calls this mosaic intelligence. In the quiz, pairs matched to complement each other beat randomly chosen pairs by 32%.

AI is jagged too. The difference is that copies of the same model share the same strengths and the same blind spots. Adding a tenth copy adds capacity. It does not add a new skill.

Build a scheme team

Illustrative scores, 0–100

What the team covers One AI agent Individual people

–Best individual person, average across skills

–One AI agent, average across skills

–Team coverage, average across skills

Team coverage takes the strongest member on each skill. That assumes the team works together well, which is the subject of the next stop.

Stop 3

The real constraint is coordination

If teams beat individuals, why not one giant team that does everything? The economist Ronald Coase asked that question in 1937. The answer is human limits. Two engineers at one desk share the whole picture. At 20 or 200 people, attention thins, assumptions drop out at handovers, and people's own incentives pull in different directions. Much of what experienced people know is tacit: they know more than they can write down.

Infrastructure feels this more than most sectors. Our schemes cross many disciplines, contracts and organisations, and they run for years. Here is how one scheme can lose context along the way.

A storm overflow scheme, end to end

Composite example

01
Network model flags spills
A modelling consultant shows an overflow spilling too often and tests storage options.
which rainfall data was used, and how well the model was calibrated at this location
02
Option chosen
The client team picks a storage tank over sewer separation on cost and carbon.
the reasons the other options were rejected
03
Outline design and estimate
A designer sizes the tank; an estimator prices it for the business case.
the estimator's assumptions about ground conditions
04
Consents and land
A consents lead negotiates with the landowner and the planning authority.
access promises made to the landowner
05
Detailed design
A different design team, now inside the contractor's alliance, picks up the work.
operations staff's preferences for maintenance access
06
Construction
The site team builds it, adapting to what they find in the ground.
what was changed on site, and why
07
Handover to operations
Operations inherit an asset that works, but is awkward to clean and inspect. The crew who will maintain it were never in the room.

A composite drawn from common patterns in UK water capital delivery. It does not describe any one company or scheme.

Transport

A rail upgrade where track, signalling, power and station works sit with different contractors and all have to fit the same weekend possession.

Energy

A grid connection that needs the developer, the distribution network operator, the transmission operator and the planning authority to line up.

Who holds it together

Programme managers, design managers and interface managers. On any single-skill test they look average. In practice they make everyone else more productive. The paper's human-only example shows how coordinators raise the productivity of specialist teams.

History supports the point. The telegraph and the railway did not make individual workers more skilled. They made handovers faster and more reliable, and that alone made the large modern company possible. Coordination technology changed how industry was organised more than any gain in individual skill.

None of the human limits have to apply to AI. Copies of an agent could pass on their full working state: every file, every assumption, every intermediate calculation. They have no career interests pulling against the team. Today they still coordinate poorly. One researcher who ran thousands of agents on maths problems, OpenAI's Noam Brown, said it is “actually very difficult to get these agents to coordinate in a productive way.” The potential is large. It has not arrived yet.

Stop 4

Competence sets the level. Coordination sets the size.

This is the paper's central idea. Track how much each person produces as a team grows. Better competence lifts the whole curve. Better coordination changes its shape, and with it the team size that works best.

Output per member as the organisation grows

Simplified illustration

Competence1.0×

Coordination45

–

The dashed grey line is a human team today, kept for comparison. The curve shape follows the paper's logic, but the formula is a simplification for illustration, not the paper's model.

1 · Solo work wins

Below a basic level of coordination, working together costs more than it adds. Today's AI agent swarms often sit here.

2 · Bounded size

The world we know. Teams help up to a point, then get too big to manage. Infrastructure's fragmented supply chain is a symptom of this ceiling.

3 · No ceiling

Past the super-cooperative threshold, each extra member makes the whole more productive. Bigger is better at every scale. Only the size of the market limits growth.

Stop 5

Four futures, depending on which dial moves

The paper treats the two dials separately, which gives four regimes. To see what each means for our sector, sort infrastructure work into the paper's three kinds of market:

Market	What it means	Infrastructure examples
Deep, standardised	Large volumes of similar work, where demand can absorb a lot of output	Sewer CCTV survey coding; standard mains replacement design; routine asset condition reports; grid connection application checks; lineside inspection analysis; smart meter and EV charger rollout planning
Niche, one-off	Specialised work where limited demand caps the payoff from scale	A new reservoir or major tunnel; a bridge replacement over a live railway; a complex treatment works rebuild on a constrained site; a first-of-a-kind energy project
Relational	Work where human involvement is part of the value	Regulator and government negotiation; community consent for a pipeline or pylon route; landowner agreements; incident command in a flood or outage; support for vulnerable customers; board accountability

Pick a regime

Select a quadrant

Coordination →

Roughly today

Competence →

“Roughly today” follows the authors' view that AI is improving fast on individual competence while lagging on coordination. The infrastructure examples are 38thMove's application of the model, not the authors'. Physical work on site also depends on robots and plant, which is a separate question.

Stop 6

What to watch, and in what order

The paper's most practical finding is about timing. Changes in pay and roles come before changes in how the industry is organised. So stable employment figures and an unchanged supply chain do not show that AI is having no effect.

Early signals

Day rates and pricing for routine desk work such as survey coding, drawing production and standard reports
Fewer graduate and entry-level roles in the tasks closest to what AI can do alone
Steady or rising demand for experienced programme, design and interface managers
Small, AI-heavy firms winning small, price-sensitive packages
Firms' spending on computing rising against spending on people

Later signals

A supplier that suddenly scales in standardised work without hiring to match
Consolidation of the supply chain in high-volume work
Frameworks and contracts rewritten around AI-run delivery
Sector productivity statistics finally moving

The tipping point may give little warning. A small improvement in how agents coordinate, too small to show up on any AI leaderboard, could be enough to reorganise a market sitting close to the threshold. That is why the authors call for tests of AI teams as well as individual models: how well they divide a job without duplicating work, hand over cleanly, and catch rather than compound each other's errors.

Stop 7

What infrastructure organisations can do now

These moves make sense in all four futures, which is why they are worth doing before anyone knows which one arrives.

Judge AI pilots by team resultsMeasure accepted output, rework, review effort and handover losses across the whole workflow. A task done ten times faster can still slow the scheme if someone has to check and fix it downstream.
Fix your handoversAgree information requirements, record decisions and the reasons behind them, and keep one trusted source for asset data. The things that help people coordinate are the same things AI agents will need.
Develop your coordination capabilityCoordination is valuable now. Develop the capability, and watch how the human role changes as agents improve.
Protect how people learnRoutine tasks are where graduates pick up tacit knowledge. If AI takes those tasks, design another route to experience.
Sort your portfolio into the three marketsKnow how much of your spend is standardised, high-volume work. That is where change could come suddenly.
Watch the supply chain for the tipping pointAsk suppliers how output scales as they add AI agents. Check whether your frameworks assume human team sizes and rates.

Stop 8

The argument in 60 seconds

AI is now as good as a capable professional at many individual tasks, yet infrastructure delivery looks much the same. A new DeepMind paper models how this could change.

Our sector runs on teams of specialists: modellers, engineers, estimators, consents leads, site managers. Each is strong in different places. Put together well, that team beats any single expert. Today it also beats a crowd of AI copies, because copies of the same AI share the same blind spots.

What limits our teams is coordination: handovers, lost context, the interfaces between contracts. Better AI skills will make work cheaper and teams leaner, gradually. The big change comes if AI agents learn to work together well. Then, in standardised, high-volume work, a few large AI-heavy organisations could quickly outcompete everyone else.

So watch pay and roles first, and improve coordination now. Better handovers help today's teams and prepare the information that future AI systems will need.

What would this mean for your programme?

If you want to think through where coordination limits your own delivery, or how to judge AI pilots by team results rather than task speed, I would be glad to talk.

Tell me what you need →

Reference

Plain-English glossary

Competence: How good an agent, human or AI, is at a specific task.
Coordination: How well agents work together: sharing context, dividing work, handing over, keeping aligned.
Jagged: Strong in some areas and weak in others. True of people and AI, but copies of one AI are jagged in the same way.
Mosaic intelligence: What a well-built team can do by combining specialists whose strengths cover each other's gaps.
Coasean ceiling: The point beyond which a firm gets too big to coordinate well, named after the economist Ronald Coase.
Super-cooperative threshold: The level of coordination at which adding members always increases output per member, so the ceiling disappears.
Technology effect: How AI changes the value of people inside the firms that exist today.
Reorganisation effect: How pay changes when firms change size and make-up in response. The paper shows this can reverse the technology effect.