
Deepak Mistry · Head of Strategic Advisory · Kaleido Projects
Overview
As AI tools begin to find their way into construction and infrastructure programmes — processing data, surfacing risks, supporting decisions — a question is quietly becoming more urgent: at what point does AI stop assisting and start acting?
This article argues that the real challenge in AI governance isn’t technical. It’s about accountability. When an AI system shapes a procurement decision, flags a subcontractor, or feeds into a gateway recommendation, the question of who owns that outcome rarely has a clear answer — because most programmes have never formally asked it.
Drawing on the governance frameworks already embedded in complex infrastructure delivery, the article makes the case that AI authority needs to be treated with the same rigour as financial delegation or contract management. The article:
- Introduces a spectrum from augmentation to autonomy
- Examines why organisations drift along it without noticing
- Offers eight diagnostic questions any programme should be able to answer before allowing AI to move from supporting decisions to making them
There is urgency here, even if it doesn’t yet feel that way. The frameworks put in place now, while AI is still being adopted and the patterns are still forming, will determine how well the sector is positioned when the pace of deployment accelerates.
Where helping ends and acting begins
The distinction isn’t abstract. It shows up in decisions that occur on major programmes every week — in the gap between what AI surfaces and what it triggers, between what it recommends and what it determines. Where that line falls, and who gets to draw it, is what this article is about:
| Helping | Acting |
|---|---|
| AI that drafts an email | AI that sends it |
| AI that flags a schedule variance on a major infrastructure project | AI that re-sequences the programme, reallocates resource or triggers a contract notice |
| AI that analyses risk across a supply chain | AI that deselects a subcontractor, escalates a claim or initiates a change |
| AI that models cost trajectories ahead of a gateway review | AI that generates the recommendation that goes forward as the decision |
The difference looks obvious when you put it that way. In practice, the line is rarely so clean — and it’s moving in one direction. AI in construction and infrastructure is still, for the most part, in its early stages. Where AI is being adopted, it tends to start with processing, summarising and surfacing information — helping people see more clearly rather than acting on their behalf. That will change. The tools are maturing, the commercial pressure to automate is real, and the industry’s appetite for better programme intelligence is genuine and growing.
Which puts the sector at an interesting point. The patterns aren’t yet locked in. There’s still a genuine opportunity to think carefully about the relationship between AI capability and AI authority before it gets designed by default. In my experience, that kind of window closes faster than people expect.
The instinct, when AI governance comes up, is to treat it as a technical problem. Get the model right, validate the outputs, keep a human somewhere nearby. But the more I work through these questions — in infrastructure, in programme risk, in major capital decisions — the more I think that framing misses what actually matters. The challenge isn’t primarily about AI capability. It’s about accountability architecture. And the governance frameworks most programmes have inherited — built around human decision-makers, delegated authorities and contractual chains of responsibility — weren’t designed with AI in mind.
The accountability gap
When AI acts, who is answerable? It sounds like a theoretical question. In my view, it’s the most practical governance question the sector faces right now.
Consider this scenario
An AI system, embedded in a programme controls platform, identifies a subcontractor’s performance as deteriorating and flags them as high-risk. That flag feeds into a procurement recommendation. The commercial team acts on it. The subcontractor is removed.
Who owns that decision? The person who configured the model? The controls manager who saw the flag and moved on? The commercial director who approved the system’s deployment? The SRO who signed off the platform? The software vendor?
In most programmes, the answer isn’t clearly defined — because the question was never formally asked. The system was procured as a tool. Its outputs were integrated into a workflow. Nobody sat down and decided what it was actually authorised to decide.
This isn’t unique to AI. Programmes have spent decades building decision-rights frameworks — delegated authorities, approval matrices, governance boards, audit trails — precisely because high-stakes decisions need to be owned by identifiable people who carry real accountability. On a complex infrastructure programme, who has authority to approve a scope change, accept a risk or issue a compensation event isn’t left to inference. It’s defined. That clarity exists because experience has shown what happens when it doesn’t.
AI systems are now shaping decisions of comparable weight without being placed into that same architecture. They arrive as analytical tools, get embedded in reporting workflows and gradually become part of how the programme thinks — often before anyone in a governance role has had reason to ask what authority they have been quietly granted.
That accountability gap is, in my view, the central problem in AI governance. Everything else follows from it.
Augmenting isn’t the same as autonomy
A lot of the confusion in this space comes from treating augmentation and autonomy as points on a single scale of AI capability, when they’re actually different in kind. I find the distinction worth being precise about, because the governance implications sit on completely different ground depending on which side a system falls.
Augmenting means AI helps humans perceive, analyse, compare, summarise, simulate, challenge or decide. The human remains the agent. Their understanding may be transformed by what the AI provides, but the decision — and the ownership of it — stays with them.
Acting means AI executes something: sending, approving, escalating, allocating, triggering, prioritising, rejecting. The AI becomes the agent. Something changes that wouldn’t have changed without it.
Crucially, the distinction isn’t about how advanced the technology is. A simple algorithm that automatically rejects a planning application is acting. A sophisticated model that produces a nuanced risk assessment is still augmenting. The difference is about agency — what authority the system has been given, and by whom.
In a programme context, the boundary runs through familiar territory. A tool that identifies potential conflicts and surfaces them for a planner is helping. The same tool, given authority to re-sequence activities or trigger float consumption alerts that feed into contractual notices, is acting. The underlying technology may be identical. The governance question is entirely different.
What tends to get lost is that increasing AI capability and increasing AI authority aren’t the same thing. You can have a highly capable AI that stays firmly in an augmenting role. You can have a relatively simple one with significant autonomous authority. On complex programmes, where contractual, commercial and reputational consequences are serious, keeping those two questions separate isn’t just sensible. It’s a governance requirement.
The drift up the ladder
Between pure augmentation and full autonomy there’s a spectrum — and organisations tend to move along it gradually rather than by design.
◎
Where does your programme sit?
Click the row in the table below that best reflects how your AI systems are currently being used. Your governance focus areas will appear beneath the table.
AI role Human role
Observes You are here Interprets
Summarises You are here Checks
Analyses You are here Evaluates
Recommends You are here Decides
Challenges assumptions You are here Justifies
Drafts action You are here Approves
Executes within limits You are here Oversees
Decides & acts You are here Reviews after the fact
← Augmenting Autonomous →
Select your level
Click the row on the left that best reflects where your programme currently sits with AI. Your governance focus areas will appear here.
Programmes don’t make a deliberate choice to move from recommendation to execution. They drift there. I think understanding why that drift happens is more practically useful than most abstract governance principles — because the drift is rarely visible to the people inside it.
- Efficiency pressure On a programme where the controls team is already stretched, once AI recommendations are consistently accepted, the approval step starts to feel procedural. The time saved by shortening the loop is real and visible. The governance cost isn’t, until something goes wrong.
- Capability normalisation When AI outputs are consistently useful — or at least better than nothing — confidence in them grows faster than scrutiny of them. What once felt like a significant recommendation starts to feel like routine output.
- Invisible authority transfer AI capability often enters programmes through technology or data teams, embedded in platforms and dashboards before anyone with governance responsibility has thought carefully about what authority is being quietly transferred.
I want to be clear that this isn’t a criticism of the people involved. It’s a description of how new capability tends to move through complex organisations. But on programmes where a misallocated risk, a misjudged compensation event or a disputed change can run to millions of pounds and years of delay, the drift from assistance to authority deserves active attention.
The governance theatre problem
At some point in most AI governance conversations, someone says: “But a human is still in the loop.”
It’s worth sitting with that claim for a moment rather than accepting it, because a human can be technically in the loop while being practically irrelevant to the decision.
Meaningful oversight requires more than a signature. The person reviewing an AI output needs:
- To understand the basis of the recommendation — not just the output, but the reasoning and the assumptions behind it
- To see alternatives, not just the chosen path
- Uncertainty made visible — not just what the AI concluded, but how confident it is and where that confidence is thin
- Time to think
- The organisational permission to push back, to reject the recommendation without being required to over-justify themselves against a system the programme has already invested in accepting
On busy programmes, those conditions aren’t always present. A controls manager reviews a risk dashboard in a Monday morning meeting, sees a red flag the system has generated, and passes it up the chain because it looks authoritative and there are fourteen other things on the agenda. The flag may be well-founded. It may also rest on assumptions the model inherited from incomplete data, a skewed baseline or a configuration nobody has revisited since go-live. The manager didn’t have the time, the visibility or the supporting information to know either way.
This is governance theatre — the performance of oversight without its substance. The accountability gap stays intact, even though a human was technically present. If something goes wrong, the approval step shows up in the audit trail. But whether the person who approved had any genuine opportunity to exercise judgement is a different question — and usually an unanswered one.
A human in the loop isn’t the same as human judgement in control. Involvement is only a safeguard when it’s real — meaning informed, empowered and carrying genuine ownership of the outcome.
Decision rights, not decision assistance
So if the question isn’t “what can AI do?” and not “is a human reviewing the output?”, what is it?
In my view, it’s this: what decisions has this system been formally authorised to make or shape, by whom, with what limits, and who carries accountability for the outcomes?
Programmes already manage this question for people. Delegated authorities, RACI structures, approval matrices — these exist because the right to make a consequential decision isn’t the same as the capability to make it. On a major infrastructure programme, a junior commercial analyst doesn’t approve a compensation event simply because they have done the calculation. The authority to approve sits with someone specific, at a defined level, with a defined scope. That boundary exists because when it’s unclear, disputes follow.
AI systems need to sit inside that same architecture — not float alongside it as tools that nobody formally authorised. Before any AI system moves from augmenting to acting on a programme, the programme should be able to answer these eight questions:
Governance Readiness Diagnostic
For each question below, select the response that honestly reflects your programme’s current position.
Once all eight questions are answered, click Generate governance summary.
Your personalised readiness summary — including specific recommendations — will appear directly below.
1Consequence
What is the worst credible outcome if the system is wrong?The severity of potential harm sets the threshold for how much human oversight is warranted. A low-consequence error is recoverable; a high-consequence one may not be.
2Reversibility
Can the action be undone, and at what cost?Irreversible actions demand a higher standard of confidence before acting. If the system gets it wrong and the action can’t be undone, the cost of that error is permanent.
3Uncertainty
Does the system surface its own uncertainty, or does it present all outputs with equal confidence?A system that presents uncertain outputs with the same confidence as reliable ones creates a false picture of the decision landscape. Visible uncertainty is a precondition for meaningful oversight.
4Transparency
Can a qualified person reconstruct the basis for a decision after the fact?Without the ability to reconstruct how a decision was reached, there’s no basis for challenge, learning or accountability. A black box that acts is a liability, not an asset.
5Authority
Who formally granted this system permission to act in this domain?Capability isn’t the same as permission. If nobody has formally granted the system the right to act in this domain, the authority gap is real — regardless of how well the system performs.
6Accountability
Who carries the outcome if something goes wrong?If accountability isn’t assigned before something goes wrong, it will be contested after. Ambiguity here doesn’t protect anyone; it just delays the argument.
7Escalation
Under what conditions does the system stop and ask?Every autonomous system needs to know when it’s out of its depth. Without defined escalation conditions, the system will act in situations it was never designed to handle.
8Auditability
Can we reconstruct what happened, why, and what was visible to whom?Governance without a record is unverifiable. The ability to reconstruct what happened, and why, is what distinguishes a governed process from an opaque one.
The right to act should be earned, bounded and reviewable. If a programme can’t answer these questions for a system that’s currently acting in consequential areas, it doesn’t yet have governance. It has a process.
Where autonomy belongs — and where it doesn’t
The argument here isn’t against AI autonomy. There are plenty of decisions on major programmes where AI acting within defined parameters isn’t just acceptable but genuinely preferable.
Automated clash detection in design coordination
Document routing and correspondence management
Progress data aggregation from site systems
Alerts when earned value thresholds are breached
Safety observation categorisation
Low-stakes operational adjustments within agreed parameters
In all of these, the action is bounded, the rules are clear, the consequences of error are limited and the audit trail is solid. Requiring human approval at each step would add friction without adding real oversight. That’s not governance. It’s bureaucracy.
The question isn’t whether AI should ever act. It’s whether the boundary between augmenting and acting has been consciously drawn — and whether the accountability architecture around it’s real rather than assumed.
Where AI should stay in an augmenting role is when decisions are high-consequence, irreversible, contractually significant, dependent on professional judgement, or where they affect safety, rights or public confidence. This includes:
- Gateway decisions
- Risk acceptance
- Scope change approvals
- Contractor performance assessments
- Claims adjudication
- Safety case sign-off
AI may have a genuinely valuable role in preparing analysis, stress-testing assumptions or modelling consequences for all of these. But its role should be to inform the decision, not own it.
I also think it’s important to be honest here — humans aren’t automatically better at these decisions. People are biased, overloaded and often poor at reasoning under uncertainty. On a major programme, AI may actually improve decision quality in meaningful ways: surfacing assumptions that go unchallenged in governance meetings, identifying patterns across a risk register that no individual would catch, or modelling schedule consequences more rigorously than time allows. The goal isn’t maximum human control. It’s better decisions with clear, functioning accountability. Those aren’t the same thing.
The boundary should be designed, not discovered by accident
Construction and infrastructure aren’t starting from zero here. The sector already understands governance architecture. Project controls, change management, delegated authority, audit trails. The question is whether those frameworks get extended deliberately to cover AI, or whether AI is allowed to accumulate authority quietly in the gaps between them.
The programmes that navigate this well won’t necessarily be the ones with the most sophisticated AI. They will be the ones where the question of AI authority gets treated as a governance design problem from the start — with the same rigour applied to financial delegation, contract management and operational risk.
That means doing four things deliberately:
- Being explicit about what AI systems are authorised to do and what stays a human decision.
- Building oversight that’s substantive rather than ceremonial.
- Making sure the accountability architecture around AI is at least as robust as the architecture around the people it works alongside.
- Revisiting those boundaries as the capability evolves, rather than letting drift do the work of policy.
We haven’t yet reached the point in this sector where these questions are urgent. AI in construction and infrastructure is still, largely, helping people work better rather than acting independently. That’s the window. The frameworks put in place now, while the technology is still being adopted and the patterns are still forming, will shape how authority is allocated when things mature.
AI should be allowed to illuminate choices, test assumptions, surface consequences and — within agreed, bounded, auditable limits — act. But where decisions carry contractual, commercial, safety or strategic weight, capability isn’t the same as legitimacy. The question is never simply whether AI can do something. It’s who gave it permission — and who answers when it does.
Working with Kaleido
The gap described in this article is the gap Kaleido was built to close.
We work with leaders and project teams across infrastructure, energy, and defence to bring Decision Intelligence into the way programmes are run. That means helping organisations frame the right problems before committing to answers, build investment cases that are honest about uncertainty, and put the analytical and governance structures in place that allow consequential choices to be made with clarity rather than momentum.
Our work spans strategic advisory, integrated risk and cost modelling, and digital programme controls — not as separate offerings, but as connected capabilities that serve a single purpose: ensuring that the decisions at the heart of a programme are as well-founded as the assets they produce.
If any of what this article describes feels familiar, we would be glad to talk.