Options and numbers
The trade-offs were always there. AI just makes them visible sooner.
This is Article 5 in a nine-part series: The Urban Professional in the Age of AI, based on The Urban Operating System.
1. The gap · 2. Three paradigms · 3. The machine room · 4. Project memory · 5. Options and numbers ← you are here · 6. The digital twin · 7. Participation · 8. Five levels · 9. The question behind the tools
A housing strategist receives a revised programme for a station area on Thursday afternoon.
The numbers don’t work.
Not dramatically. That would be easier to handle. The problem is more ordinary: the residual land value has shifted. Construction costs have moved. The mid-market segment that carried part of the financial logic is weaker than it was eighteen months ago. The housing association can still participate, but not on the same risk allocation. The developer wants more owner-occupied housing. The alderman wants to hold the affordability promise. The neighbourhood group remembers the public courtyard commitment. The finance department wants a clearer view of exposure before the next steering group.
The team asks for options.
What follows is a familiar sequence. The financial advisor adjusts the model. The urban designer checks what extra volume might fit. The policy lead tests the affordability mix. The mobility advisor recalculates parking pressure. The project manager tries to assemble these partial views into a coherent recommendation. Two weeks later, the team has three options.
Option A is politically attractive but financially weak. Option B is financially stronger but socially harder to defend. Option C is called balanced, which usually means nobody loves it and nobody has yet found the hidden problem.
This is how option work often happens. Slowly. Sequentially. With each discipline holding part of the truth, and with the decisive trade-offs only becoming visible after preferences have already hardened.
AI changes this. Not by inventing better answers, but by changing when the trade-offs become visible, and by changing who has to do what to get there.
That is what this article is about.
The option space is larger than the meeting can hold
Urban development is often discussed as if the task is to choose between a small number of alternatives. In practice, the number of possible alternatives is enormous.
Change the share of social rental. Change the phasing. Change the parking ratio. Change the block depth. Move the school. Reduce underground parking. Add a floor on the north side, remove one on the south. Change the tenure mix. Delay the public space investment. Use a different subsidy assumption. Shift risk from the public to the private side, or the other way around. Accept a lower land receipt. Change the energy concept.
Each adjustment affects something else. A programme mix is not only a housing choice. It is simultaneously a financial, political, mobility and legal choice. A parking norm is not only a mobility assumption: it changes construction cost, public space, target groups, marketability and residual land value. A design variant is not only spatial: it changes sunlight, noise mitigation, foundation costs, energy performance and the credibility of what was promised in participation.
This is why teams narrow the option space early. They have to. Human working memory cannot hold hundreds of combinations across ten constraints simultaneously. So teams simplify. They select a few plausible options, test those, and treat the selected set as if it represents the real decision space.
Often it doesn’t.
It represents the part of the decision space the organisation had time to see.
A 2024 systematic review of computational optimisation in urban design describes precisely this gap: Urban Design Optimisation methods allow teams to explore large numbers of design solutions for a district, evaluate multiple objectives simultaneously, and select from Pareto-efficient solutions rather than from a pre-narrowed shortlist.1
A Pareto-efficient solution is one where you cannot improve on one objective without making another worse. In area development terms: you cannot increase the affordable housing share without reducing the land revenue, or improve the public space quality without reducing buildable volume, unless you change the underlying conditions. There is no option that scores best on everything. What the method produces is a frontier: a set of combinations where every step forward on one dimension requires a step back on another. The professional task is then not to find the single best answer, but to choose which trade-off is most defensible given the project’s commitments, its political context, and what was promised to whom.
The tools have been available for years. The constraint has not only been technical. It has also been organisational: most planning teams have lacked the data, workflows and decision routines to use these methods in ordinary project work.
AI changes that equation. Not by replacing the optimisation methods, but by making them usable in the context of a real project: connected to its commitments, its political constraints, its financial assumptions, its partner agreements.
Numbers don’t decide. They discipline the conversation.
There’s a common mistake in discussions about AI-assisted feasibility work. The idea is simple: better models produce better decisions because they calculate more precisely.
That’s not quite right.
The value of numbers in urban development is not that they decide. They discipline the conversation. They force a team to say what it’s assuming. They reveal which ambition depends on which subsidy. They show when a political narrative rests on a financial condition nobody has yet tested.
A feasibility model is not reality. It’s an argument in numerical form.
That argument may be useful. It may be honest. It may be carefully built. But it is still an argument. It contains assumptions about sales values, rent levels, cost inflation, phasing, interest rates, land values, subsidy conditions and risk allocation. Some of those assumptions are evidence-based. Some are negotiated. Some are inherited. Some are political choices disguised as technical inputs.
This is where the previous article matters. A project memory room records decisions, assumptions, alternatives, commitments and unresolved tensions. It records why rejected options were rejected, and whether they were permanently closed or only impossible under the conditions of that moment. That is the direct foundation for option work.
Without it, AI can generate variants. It cannot know which variants are responsible.
An AI assistant can run twenty programme mixes in a minute. But if the affordability promise from 2023 is missing from the source base, one of those options may look efficient precisely because it has forgotten the thing the project must still honour.
A model that forgets commitments doesn’t create a better option. It creates a more elegant breach of trust.
The sensitivity analysis becomes the argument
The most useful number in a feasibility model is often not the outcome. It’s the sensitivity.
Which assumption moves the result most? Which variable has become load-bearing? Which cost increase breaks the programme? Which rent assumption makes the affordable share possible? Which phasing delay creates the cashflow problem? Which public investment is treated as fixed, even though it was never formally secured?
AI makes sensitivity work faster. That is useful. But the more important shift is positional: sensitivity analysis can move from the back of a technical appendix to the centre of the decision conversation.
A steering group should not only see that Option B has a stronger financial result than Option A. It should see why. It should see that Option B depends on a sales value assumption based on market data from before the last interest rate shift, and that the same option reduces the public courtyard residents were told would remain. It should see that Option A has a weaker land result but is less exposed to market absorption risk and better aligned with the housing association agreement.
That’s not a technical detail. That is the decision.
The option note of the AI era should not be organised around preferred options alone. It should be organised around load-bearing assumptions.
What must be true for this option to work? What happens if it’s not? Who carries the risk if it fails? Which commitment becomes vulnerable? Which future option does this choice keep open, and which one does it close?
The last two questions are often missing. They are also the ones that matter most in long urban projects.
There is also a false precision problem. A model that reports a residual land value to the nearest euro may look more certain than the situation allows. In early option work, ranges are often more honest than point estimates. A band of uncertainty tells the steering group more than a precise number built on fragile assumptions. This is not a technical argument. It is a psychological one: people trust a number with two decimals more than a range, even when the range is the accurate representation. AI can produce false precision faster and more convincingly than any spreadsheet. That is one more reason to keep the assumption register in the room.
What AI takes over, and what it doesn’t
This is a good moment to be concrete about the division of labour. It’s also where most discussions about AI in planning stay too vague.
AI can now handle some tasks in option work with limited human intervention, provided the inputs, constraints and review rules are clear. Running multiple programme variants against a fixed set of financial parameters. Generating a sensitivity table showing how the residual land value shifts under different cost or revenue assumptions. Flagging whether a proposed option may conflict with a recorded commitment in the project memory. Producing a plain-language summary of what each option costs on each criterion. These are structuring tasks. They used to consume days. They now take minutes.
Some tasks AI supports but cannot own. Translating the results of a sensitivity analysis into a recommendation that a steering group can act on. Deciding which assumptions are load-bearing and which can be loosened. Identifying which combination of conditions is politically negotiable and which is not. Judging whether a formally compliant option is actually deliverable, given the relationships, trust levels and institutional capacity in the room. These tasks require contextual knowledge that lives in the professional, not in the model.
And some tasks AI should not enter at all. Deciding what the city owes to which residents. Setting the weights: whether affordability counts more than land revenue, whether speed outweighs public space quality. Knowing when to surface a number and when to keep it off the table. Carrying the responsibility for a recommendation when the choice is genuinely hard. Those remain human, not because the technology isn’t powerful enough, but because they involve authority, trust and accountability that cannot be delegated to a model.
The practical implication is this: a professional who used to spend two weeks producing three options can now spend two days producing a structured option field. The time freed up is not a bonus. It is the point. The goal is not to produce more options. It is to spend more professional attention on the options that matter.
But only if the organisation makes that trade consciously. If the time saved on variant generation is simply filled with more variant generation, nothing changes.
Keeping options open is a strategy, not indecision
Urban development has a cultural bias toward certainty. Decision notes prefer a clear recommendation. Political processes prefer visible progress. Project teams are rewarded for moving forward. Uncertainty is something to reduce before a decision is made.
Sometimes that’s right. Often it’s not.
In long projects, the best decision is not always the one that optimises the current model. It may be the one that keeps the most valuable future options open.
This idea has been formalised in climate adaptation. Haasnoot et al. (2013) developed the Dynamic Adaptive Policy Pathways approach at Deltares and TU Delft, starting from a simple recognition: many long-term decisions must be made under deep uncertainty, so planning should combine near-term action with future flexibility. The method identifies adaptation tipping points: the conditions under which a current policy will no longer meet its objectives and a new one is needed.2
Area development needs a similar discipline.
Not because housing programmes are the same as flood risk strategies, but because both involve long-lived investments, uncertain futures and decisions that can create lock-in before the uncertainty is resolved.
A project may not know whether the mid-market rental segment will recover in three years. It may not know whether a mobility hub will reduce car ownership enough to justify a lower parking ratio. It may not know whether a subsidy scheme will survive the next cabinet. It may not know whether a later climate standard will make today’s public space design insufficient.
The answer is not to wait for certainty. Certainty arrives too late.
The answer is to make decisions that are explicit about thresholds. If sales velocity falls below this level, revisit the tenure mix. If parking pressure exceeds this level after phase one, activate the mobility reserve. If construction costs rise above this band, reopen the phasing sequence. If the subsidy is not secured by this date, shift to the fallback programme.
AI can help maintain this adaptive logic. It can monitor indicators, update scenarios, flag expired assumptions and show which pathway the project is drifting toward. But the adaptive strategy itself remains a governance choice. The organisation must decide which uncertainties matter, which signals count, and what it’s willing to change when conditions shift.
That’s not a model output. That’s institutional discipline.
The metrics trap
There’s a danger here, and it’s serious.
Once options can be scored quickly, the score starts to acquire authority.
This is partly structural and partly psychological. What is measured is easier to defend. What is harder to measure becomes easier to treat as secondary. And when a number comes from a model rather than a colleague, it often gains extra authority. A colleague’s estimate can be questioned directly. A model’s output first requires understanding what the model assumed, which most people in the room do not have time for. So the number stands. Not because it is right, but because nobody knows how to push back against it without looking like they are rejecting evidence.
AI intensifies this tendency because it can produce more scores, more quickly, with more apparent precision.
A 2026 study in npj Urban Sustainability examines this directly. Based on 28 AI implementations selected from 157 deployments across six urban domains, Moghaddam and Cao show how cities have pursued technical accuracy while failing to achieve policy goals. One documented example: an acoustic detection system reached 97 percent accuracy on its own metric while yielding only 9.1 percent effectiveness on the actual policy objective it was deployed to address.3
The example is not imported here because gunshot detection resembles area development. It does not. It matters because the failure mode is transferable: a system can perform well against its own technical metric and still fail against the public purpose for which it was deployed.
For option work in area development, the pattern is close at hand.
A model can optimise for housing numbers and miss household formation. It can optimise for land revenue and miss long-term public value. It can optimise for green square metres and miss whether those spaces are usable by teenage girls, older residents, or children without private gardens. It can optimise for average accessibility and miss who remains disconnected.
The problem is not always biased data. Sometimes the data is clean and the model works exactly as designed. The problem is that the wrong objective was chosen.
This is the uncomfortable part of options and numbers. Better calculation doesn’t remove politics. It moves politics upstream into the choice of variables, weights and thresholds.
Who decides whether affordability counts more than land revenue? Who decides whether a public space commitment is a hard constraint or a soft preference? Who decides when a model is allowed to say that a politically desired option is not financially feasible?
These are not technical questions. They are governance questions with numbers attached.
This is why weighting should not be hidden inside the model. If affordability counts twice as much as land revenue, say so. If delivery speed is allowed to outweigh public space quality, say so. If political feasibility is included as a criterion, name who assessed it and on what basis. A weighting framework is not a technical appendix. It is a political document in numerical form.
Goodhart’s Law applies here with particular force: when a measure becomes a target, it ceases to be a good measure. In option work, this means that the moment a score becomes the goal, the underlying purpose it was meant to serve starts to recede. In area development, this is why a housing number can start to crowd out housing need, why a green space norm can crowd out actual use, and why a financial score can crowd out public value. AI does not create this risk. But it accelerates it.
What the decision package should contain
The most valuable output from AI-assisted option work is not a beautiful image or a ranked list.
It’s a better decision package. Six things, at minimum.
A constraint map: what is legally fixed, financially constrained, contractually committed, spatially bounded, politically preferred or self-imposed? Many planning processes collapse this distinction, treating soft preferences as hard constraints and ignoring actual hard constraints until they cause a problem.
An assumption register: which numbers drive the result, where do they come from, who owns them, and when do they expire? No number without an assumption. No assumption without an owner.
A commitment check: which promises, participation outcomes or partner agreements are touched by each option? This is where the project memory room and the option field connect. An option that performs well against financial criteria but contradicts a formal commitment to a neighbourhood group is not a free choice. It has a cost.
A sensitivity summary, written in plain language. Not appended as a technical annex. The purpose is to show the steering group where the project is fragile, and why.
An option history: what was tested, what was rejected, and under what conditions could it return?
A recommendation that separates calculation from judgement. The model may show that an option performs better on selected criteria. The professional still has to explain whether that option is defensible.
This is where AI becomes genuinely useful in a planning organisation. Not as a machine that chooses, but as a machine that makes the choice harder to obscure.
The new professional skill: specifying the question
The urban professional in this workflow is not replaced by the model. But the nature of the work changes.
The tasks that used to define option work move partly to AI: running the model, generating variants, building the comparison, writing the first draft of the trade-off summary. What remains is harder to delegate and harder to learn from a textbook.
The most important skill becomes the ability to specify the question precisely enough that the output is actually useful.
A weak prompt asks: make three options for this site.
A stronger prompt asks: generate programme variants for 450 to 650 dwellings, with at least 35 percent affordable housing, no loss of the public courtyard commitment, no net increase in public parking pressure, and a maximum public investment exposure of X. Use current cost assumptions but test sensitivity for construction cost increases of 5, 10 and 15 percent. Flag options that depend on assumptions not yet formally approved. Identify which earlier commitments each option affects. Do not rank the options until the constraints and assumptions have been shown.
That’s not just better prompting. It’s better planning.
The professional who can write that specification understands the project. She knows which constraints are real, which assumptions are fragile, which promises matter and which numbers are politically loaded. AI makes her more effective because she already knows what the work requires.
The opposite is equally true. A professional who can’t specify the question will receive outputs that look useful and may be fundamentally misdirected.
This is also where the character of professional development shifts. Junior professionals used to learn option work by doing option work: building models, running variants, seeing what changed when you moved a parameter. That learning path is partly disrupted when AI handles the mechanical steps. The risk is that something important is lost before something better replaces it. The skill of reading a model critically, of knowing which number to distrust, of recognising when a result is too clean, develops through practice with the model itself. If that practice disappears, the professional may be able to specify questions but unable to evaluate answers.
That is not an argument against using AI. It is an argument for being deliberate about what junior professionals still do by hand.
Where the discomfort must stay
There’s a temptation to make option work smoother. AI can help produce cleaner dashboards, clearer scorecards and more confident recommendations. That’s useful. But some discomfort must remain.
The purpose of option analysis is not to remove discomfort. It’s to locate it accurately.
If all options look good, the model is probably hiding the conflict. If one option clearly wins, the evaluation framework may be too narrow. If the numbers settle the matter, the wrong questions may have been excluded.
Good option work should reveal where the project is genuinely constrained. It should show the alderman that keeping the affordability promise requires either more subsidy, lower land revenue, higher density, reduced parking cost or a different phasing strategy. It should show the developer that viability depends on public choices that cannot be treated as technical inputs. It should show residents when a promise is being changed, and why. It should show the project team which decision can be made now and which one should remain conditional.
And it should show where the institutional capacity to make those decisions is thin.
The question is not only which option is best. It is whether the organisation has the governance capacity, the political will and the partner trust to actually implement it. An option that depends on a decision the organisation cannot make is not a real option.
AI can make the conversation more precise. It cannot make it painless.
How to start
A planning organisation doesn’t need a full digital twin to begin with options and numbers.
It needs one live feasibility model, one structured assumption register and one repeatable option template.
Start with a project where a programme decision is coming. Take the existing model. Identify the ten assumptions that matter most. Record their source, owner, date and review condition. Then ask AI to help generate a structured option field around those assumptions: not a final recommendation, but a map of feasible combinations and exposed trade-offs.
Use the first round internally. Don’t present it as truth. Present it as a challenge to the team’s existing understanding.
Which option did the team expect to work, but doesn’t? Which rejected option becomes viable under a changed assumption? Which preferred option depends on an assumption nobody owns? Which political promise has a financial consequence that hasn’t yet been named?
That’s enough to begin.
The next step is to connect the option field to the project memory room. Every option study should leave a trace: what was tested, what was rejected, why it was rejected, which assumptions were used, which commitments were affected, and what should be reviewed if conditions change.
That creates cumulative intelligence. The next time the market shifts, the team doesn’t start again. It asks which assumptions have expired and which alternatives should return.
What this changes
Article 3 argued that the machine room is where AI first changes the operational conditions of urban work. Article 4 argued that long-running projects need a project memory room, because decisions without memory become brittle, and organisations that cannot remember cannot adapt responsibly.
This article adds the next layer.
Once the machine room works and the project memory exists, option generation becomes different. The team is no longer limited to the three variants it had time to prepare. It can explore a wider field, test assumptions earlier, expose trade-offs more honestly and keep future pathways visible.
The hardest constraints remain: land ownership, finance gaps, political courage, legal procedure, public trust, implementation capacity. AI doesn’t dissolve any of them.
But the nature of the professional’s job shifts. Less time producing options, more time understanding them. Less time building models, more time knowing which assumptions to challenge. Less time writing the trade-off summary, more time explaining it to people who will resist parts of it. The work becomes more interpretive and more political, not less.
Back to the housing strategist on Thursday afternoon. The numbers don’t work. That hasn’t changed. But what changes is what she can do with that problem before the next steering group. She can run the option field against the actual constraints. She can test which assumptions are load-bearing and which can be moved. She can check which commitments are affected by each variant. She can show the team not three options but the contours of the decision space, and the conditions under which each path holds.
She still has to decide which path to recommend.
AI has not removed judgement from the work. It has changed the material judgement works with.
The promise of options and numbers is not certainty. It is a wider decision space, a clearer view of what each path costs, and a better record of why one path was chosen over another.
Next in this series: The digital twin. Once options can be generated and tested, the question becomes where that testing happens. The digital twin is moving from expert visualisation tool to shared governance workspace.
Yes, I used AI. Here’s how.
Yes, I used AI. It helped with source finding, structure, critical review and drafting alternatives. I used it as a sparring partner, not as an authority. The argument, source selection and final wording are my own.
Notes
Tay, J. Z., Ortner, F. P., Wortmann, T. & Aydin, E. E. (2024). Computational Optimisation of Urban Design Models: A Systematic Literature Review. Urban Science, 8(3), 93. The review covers research from 2012 to 2022 and defines Urban Design Optimisation as a way to explore many design solutions for a district, evaluate multiple objectives simultaneously, and select from Pareto-efficient options. Pareto-efficient here means: combinations where improving one objective requires accepting a cost on another. mdpi.com/2413-8851/8/3/93
Haasnoot, M., Kwakkel, J. H., Walker, W. E. & ter Maat, J. (2013). Dynamic adaptive policy pathways: A method for crafting robust decisions for a deeply uncertain world. Global Environmental Change, 23(2), 485–498. The DAPP method, developed at Deltares and TU Delft, introduces the concept of adaptation tipping points: conditions under which a current policy no longer meets its objectives and new action is required. The Deltares website provides a current overview of the method and its applications. deltares.nl/en/expertise/areas-of-expertise/sea-level-rise/dynamic-adaptive-policy-pathways
Mashhadi Moghaddam, S. N. & Cao, H. (2026). The metrics trap: how technical sophistication masks social harm in urban AI systems. npj Urban Sustainability. The authors analyse 28 AI implementations selected from 157 deployments across six domains and show how technical accuracy can mask policy failure. nature.com/articles/s42949-026-00394-1