Expert Insight: Drawing the Line Between AI and Human Judgement in Infrastructure Delivery
)
In this interview, Oliver Cronk, Founder and Chief Architect at Cronk Advisory, cuts through the hype around AI in infrastructure delivery, pointing to where the evidence for real-world impact is strongest, such as flood forecasting and automated consenting review, and where it is thinnest, particularly the generative design claims that are often just parametric optimisation rebadged.
Oliver sets out a practical test for governance, arguing that organisations should stop searching for a single AI "source of truth" and instead decide in advance which data source wins in which circumstance. And explains why, as AI takes on more analysis and reporting work, the value consultants offer is shifting from producing the work to owning the judgement behind it.
Ahead of chairing the panel Harnessing AI & Digital Tools to Transform Infrastructure Delivery at the Sustainability Delivery Summit London, Oliver shares where he draws the line between AI and human oversight.
EA: You've built a career on the practical, and often critical, adoption of AI in business. Where are you seeing meaningful real-world impact in infrastructure delivery, including planning, consenting and investment decisions?
From my work and research: impact comes from where three things coincide: plenty of historical data, an expensive manual process, and a clear decision that gets made.
Flood and water is the most mature example. JBA Consulting's machine-learning tidal model for the Thames Barrier cut forecast error to within 10cm for 92% of high tides, against 68% for the existing hydraulic model, and it is in operational use rather than a pilot. On the consenting side, government is doing some of the most interesting work itself: the i.AI Consult tool analysed the Independent Water Commission call for evidence, agreeing with human reviewers about as often as they agreed with each other. Asset inspection using computer vision is moving from pilot to production too.
What I see far less of is verified evidence for the big "generative design" and agentic workflow claims. Much of what is badged generative is parametric optimisation, which is valuable but not new. The critical adopter's job is to tell those apart.
EA: What are the biggest multi-stakeholder challenges to AI adoption you've encountered, and what's worked to overcome them?
Data sharing and re-use across stakeholders is often a major issue. AI is only as good as the data it works from. This data problem sits behind an incentive problem. Infrastructure projects generate huge amounts of environmental and performance data that often never leave the project, either because there is no standard for it, no one is paid to share it, or because sharing it might show something uncomfortable. Retrofitting that onto live projects is harder still.
EA: How should organisations think about governance and data integrity to have real confidence in AI's "source of truth"?
Start by accepting there is no single source of truth, only sources of varying trustworthiness, and design for that. The model, the reporting system and the site survey will disagree. Governance is deciding in advance which one wins in which circumstance, and who signs it off.
Three practical tests:
- Provenance: can you trace an AI-informed output back to the data it drew on and the version of the model that produced it?
- Transparency and visibility of reasoning: can a competent human see why the tool reached its conclusion, not just what it concluded? If not, it should not be near a design or a funding decision.
- Evidence hygiene: separate vendor-reported figures from independently verified ones. A striking number of quoted AI savings in this sector are suppliers marking their own homework.
Look for evidence behind AI use cases you are looking at. Is it a demo, pilot, production? Are the outcomes claimed from vendors / client / independent reviews?
EA: As AI takes on more of the analysis and reporting work, how is that changing what clients expect from the consultants and architects advising them?
Clients are asking a reasonable question: if the tool can do the analysis, what am I paying you for? The honest answer is that the value is shifting from producing the work to owning the judgement about it: framing the problem, knowing which data to trust, spotting the error the model made confidently, and carrying the professional accountability when it matters.
There is also a competitive angle. Clients are increasingly building their own tools. The firms that will thrive are those that own domain IP or specialist data rather than simply reselling access to someone else's model.
EA: The session will look at achieving the correct balance between AI and human oversight. In your experience, where should that line be drawn?
One of the panelists frames this really nicely: "Human first, human last". AI is usually best in the middle as a supporting tool: generating options, sifting volume, flagging anomalies, drafting. Humans belong at the framing / directing and at the sign-off, and those two points should have a named person / role.
The line moves depending on consequence and reversibility. Automating the triage of ten thousand consultation responses, with a human checking the themes, is a good trade. Letting an agent pass unverified data into a structural design is not, and the liability question there is unresolved.
The test I use: if this output were wrong, would a human catch it before anything physical, financial or legal happened? If the answer is no, you have drawn the line in the wrong place.