AI Joins the Turf Wars
The Unread Contract, Essay No. 2
Three teams sat down with the same AI-generated flood risk analysis. The water engineers said it confirmed their case for embankment upgrades. The urban planners said it confirmed their case for rezoning the floodplain. The disaster risk managers said it confirmed their case for early warning investment. One integrated analysis. Three different readings. The meeting ended with a request for more analysis.
The AI had given everyone better information. It had settled nothing.
University of Chicago sociologist, Andrew Abbott, predicted this in 1988. In his landmark book, The System of Professions, he showed that when a new technology enters a system of competing expert groups, it doesn’t dissolve the jurisdictional fight. It recruits a new contestant. The engineers, the planners, and the disaster risk managers were already fighting over who owned the flood problem before the AI arrived. After it arrived, they were still fighting, and the vendor whose model architecture had quietly decided which hazard scenarios to run was now in the room too, with the least legible claim of all.
This is the pattern that the “AI as coordinator” argument consistently misses.
The argument is genuinely seductive. If a geospatial foundation model can span hydrology, land use, climate science, and fiscal analysis simultaneously, the institutional cost of coordinating those professional domains should fall toward zero. The technology integrates what humans historically divided. Buy the AI. Skip the committee. It’s a claim you hear from vendors, from reformers, and increasingly from the ministers who commission both. It breaks down in three specific ways that matter enormously for how governments should procure and deploy these tools.
The first: AI inherits the jurisdictions of its training data.
Models are trained on data that reflects existing professional divisions. Hydraulic models embed engineering assumptions about acceptable flood return periods, typically a one-in-hundred-year event for major infrastructure, but that is a value choice, not a technical finding. Land-use models embed planning assumptions about density and setbacks. Climate projections embed assumptions about which hazards to prioritize and which discount rate to apply to future damages.
When you combine these into an integrated AI output, you get a weighted average of pre-existing professional frameworks, with the weights set by whoever controlled the training data and the model architecture. The jurisdictional fight hasn’t been resolved. It’s been frozen into the model.
This is especially acute in Global South cities. Prithvi, the geospatial foundation model developed jointly by IBM and NASA and released in 2023, was trained on Harmonized Landsat Sentinel-2 imagery, a data source that skews heavily toward formally mapped territory with reliable historical coverage. Clay, released by the Clay Foundation in 2024, similarly draws on multi-spectral satellite archives that reflect built environments with functioning land registration systems. Applied to informal settlements in Johannesburg’s southern corridor, or to Juba’s peri-urban fringe where SPARC is currently operating, these models produce outputs that look integrated but embed assumptions about land structure, drainage connectivity, and address systems that were never validated for those contexts. The parcel-level risk score carries a large hidden confidence interval. The clean visualization doesn’t show it.
More important: the profession whose assumptions dominate the training data wins the jurisdictional fight invisibly. Nobody voted on it. Nobody argued about it at the procurement stage. It happened in a parameter file.
The second: AI moves the fight downstream, not away.
Abbott’s three stages are worth returning to: diagnosis, inference, treatment. AI genuinely helps with the first two. An integrated model can produce a multi-domain diagnosis and reason across more variables than any single professional team. Treatment is different. Building an embankment, rezoning a floodplain, collecting a property tax, relocating a household: all of these require political authority, legal mandate, and budget. The AI holds none of them. At treatment, the jurisdictional fight resurfaces, harder than before.
Rotterdam’s Room for the River programme (2007-2015, EUR 2.3 billion) is the clearest documented case. The Dutch faced exactly the water-versus-urban problem. Their solution was structural rather than technical. A Programme Directorate reported simultaneously to the Ministry of Infrastructure and Water Management and to the association of local municipalities. Water safety was the non-negotiable primary objective. Urban quality held co-equal design standing with veto power on implementation. Neither could proceed without the other’s explicit sign-off.
The PBL Netherlands Environmental Assessment Agency evaluated the programme in 2015 and found that the dual-accountability structure was the primary driver of project quality, specifically because it forced integrated design at the front end rather than sequential handoffs at the implementation stage. All 34 projects completed on time, essentially without precedent for a Dutch infrastructure programme at that scale.
The finding that matters here: the joint design process, water engineers and urban planners arguing through trade-offs together, was the mechanism of quality. If an integrated AI analysis had dissolved that argument at the diagnosis stage by producing a seamlessly authoritative output, the political fight would have re-emerged at implementation, where it would have been far harder to resolve.
The third: AI recruits a new contestant.
This is Abbott’s central prediction playing out precisely. The vendor and the data scientists who control the model architecture now have a stake in the diagnosis. They made choices embedded invisibly in the model: which flood return period to use as the design standard, which land-use classification scheme, which climate scenario, which damage function, which threshold triggers an alert. These are professional judgments. Because nobody labeled them as such in the procurement contract, nobody argued about them when it mattered.
The result: water, urban, and the vendor’s technical team are all in the fight. The vendor’s claim is the most dangerous because it’s the least legible. You can argue with a water engineer about their assumptions. The model’s assumptions live in a configuration file that nobody in the Ministry of Urban Development has read, or has the standing to interrogate.
FEMA’s Hazus platform in the United States offers the long-run evidence. Hazus integrates engineering, economics, and urban planning into a multi-hazard loss estimation tool and has been in operational use since the late 1990s. It did not eliminate jurisdictional disputes between FEMA, state emergency managers, city planners, and local elected officials. A November 2022 GAO report on the federal approach to disaster recovery found programmes fragmented across more than 30 agencies, with state and local officials reporting they could not navigate the multiple programmes and their differing requirements. The GAO recommended Congress establish an independent commission to address the fragmentation. (GAO-23-104956, Disaster Recovery: Actions Needed to Improve the Federal Approach, November 2022.) Three decades of Hazus have not resolved that. The tool produces better loss estimates. The institutional fight over who acts on them remains.
Where AI genuinely helps
Three places specifically, and they’re worth naming. This is not an argument that AI makes coordination harder across the board.
Cheaper iteration reduces the cost of the coordination conversation itself. Rotterdam worked partly because the Dutch could model trade-offs quickly enough to make negotiation productive. When water safety and urban quality requirements conflict, an integrated model can show both teams the precise consequences of each other’s constraints in near real-time. That narrows the fight to where it belongs: value choices, not factual disputes about consequences. Reducing that friction is significant even when it doesn’t end the fight.
Shared data reduces bad-faith arguing. One reason water and urban teams fight is that they run different models with different input assumptions, and neither can fully interrogate the other’s numbers. A shared AI-generated data layer forces both teams onto the same factual substrate. Rwanda’s National Land Authority used AI-assisted parcel identification and drone mapping for exactly this purpose: not to automate decisions about land allocation or tax assessment, but to create a single verified factual layer that forced planning, revenue, and land administration into the same institutional conversation. The coordination didn’t happen automatically. The shared data made it significantly harder to avoid.
AI surfaces hidden trade-offs that professional silos conceal. When water engineers design a flood embankment, their models typically don’t compute displacement costs or urban connectivity effects on the communities behind it. When urban planners approve density in a risk zone, their models rarely compute downstream hydraulic consequences. An integrated analysis that makes both effects simultaneously visible is uncomfortable for both professions. That’s precisely what makes it useful.
The dangerous scenario
Governments use AI as a reason to defer the institutional design work.
The logic runs: commission the integrated analysis, skip the messy inter-ministerial coordination. The technology delivers an impressive output. Nobody has settled who acts on it or who is accountable when it’s wrong. The output joins the collection of accurate-but-unimplemented analyses that line the walls of every urban ministry in the Global South.
This is the pilot cemetery failure mode in a new form, and it has a specific political economy that makes it likely. An AI system feels neutral in a way that a water engineer or an urban planner does not. Commissioning an AI analysis is easier for a minister than forcing a coordination settlement between directorates that have been fighting for a decade. The technology becomes a procedural substitute for the decisions nobody wanted to make.
Bangladesh is the live case. The Bangladesh Water Development Board controls national flood infrastructure. Urban local bodies own land use within cities. These jurisdictions have never been formally settled. Dhaka’s chronic flood problem is documented in World Bank urban resilience assessments with a consistent finding: the technical solutions are available, the financing gap is manageable, and the coordination failure is binding. BWDB engineers design for river flood protection. Municipal engineers design for urban drainage. Nobody designs for the interaction between the two systems, which is where Dhaka actually floods. Deploying AI will produce a better map of where the coordination failure causes harm. It will not cause the coordination to happen.
The coordinator the AI can’t replace
The institutions that use AI successfully for multi-disciplinary problems share one structural feature. They built the coordination architecture first, then gave AI a defined role inside it.
The Netherlands Delta Programme (2010 onwards, approximately EUR 1 billion per year through 2050) created a Delta Commissioner: a politically appointed cross-sectoral authority with convening power but not command authority. The Commissioner cannot order ministries to act. The Commissioner can convene, publish integrated risk assessments, and hold politicians publicly accountable for inaction. AI-generated climate and flood risk analyses now feed directly into the annual Delta Programme report, giving the coordination function better information and a harder-to-ignore evidentiary base. The AI made the coordination more productive. It didn’t replace the coordinator.
Singapore took the statutory route. PUB’s mandate was expanded to require that drainage infrastructure function simultaneously as public space, legally embedding the urban-water integration into a single agency’s accountability. AI-assisted design tools now help engineers model the dual requirement. They work because the legal settlement came first.
Before any government buys an AI tool for a problem that spans professional domains, four questions need written answers. Who has convening authority when the disciplines disagree? Who makes the final call and is accountable for it? What does each technical team contribute, and on what does it not have standing? Who answers politically if the integrated output is acted on and fails?
The AI will improve every step of the process once those answers exist. It cannot produce the answers. Abbott saw this coming. The procurement contract should too.