Your Project Will Not Be Killed by Engineering. It Will Be Killed by the Community.
Large capital projects are scrutinised to an extraordinary degree on the dimensions that are easy to model. Grid connection, water availability, planning envelope, cost of capital, construction sequencing: all of it gets a spreadsheet, a sensitivity analysis and a risk register entry with an owner.
Then the project is stopped by a petition, a council vote, a moratorium or a judicial review, and the post-mortem discovers that the thing which actually killed it had no owner, no score and no line in the register. Community acceptance is treated as a communications activity that begins after the decision, rather than a risk that could have been assessed before it.
The Asymmetry
Engineering risk is quantified early, owned explicitly and revisited continuously. Social risk is qualified late, owned informally and revisited only when it becomes a crisis. Both can end a project. Only one of them is managed as though it can.
The asset at stake has a name in the research literature: the social licence to operate — the ongoing acceptance a project holds among the communities affected by it, independent of any permit. Thomson and Boutilier's model describes four levels, and projects move between them in both directions: withheld or withdrawn, acceptance, approval, and finally psychological identification, where the community treats the project as partly its own. The transitions are governed by legitimacy, credibility and trust. A permit grants none of these.
What Conflict Actually Costs
| World-class mining project, capex US$3–5bn | ~US$20m per week of delayed production |
| One project's 9-month construction delay (2010) | US$750m in additional costs |
| Operation halted by a power-line blockade | US$750,000 per day |
| Senior management time diverted to conflict | 35–50% for some executives |
Source: Franks, Davis, Bebbington, Ali, Kemp & Scurrah, "Conflict translates environmental and social risk into business costs", PNAS 111(21): 7576–7581 (2014). The study also found the largest costs are opportunity costs — projects not pursued — and that these are the costs most often overlooked.
Why This Risk Is Getting Worse
Three things have changed for energy, data-centre and heavy infrastructure development in Europe and beyond.
Opposition organises faster than approval does. A local group can reach several thousand signatures in a fortnight using tools that cost nothing, while the corresponding permitting process moves on a timescale of quarters. The asymmetry in tempo is new, and it is structural rather than cyclical.
Resource visibility has increased. Water consumption and grid draw that were once technical annexes are now front-page local issues, and a project's demand figures are routinely compared with the host municipality's entire consumption. That comparison is easy to make, hard to rebut, and extremely effective.
Precedent travels. A community facing a proposal now arrives at the first public meeting already briefed on how a comparable project was fought elsewhere, including which arguments worked. Developers, by contrast, often arrive without having systematically studied the same precedents.
What Gets Measured Instead
Ask most development teams to characterise community risk on a site and the answer comes back as an adjective. Local opposition is moderate. Sentiment is mixed. The council is broadly supportive.
These statements cannot be acted on, compared between sites or falsified. "Moderate" is not a finding; it is a placeholder for the absence of one. Compare it with a statement that carries an actual claim: a petition passing several thousand signatures inside a month, a council with a documented record of refusing comparable proposals, a projected water draw that exceeds what the host municipality itself consumes at peak. Each of those can be checked, argued with and mitigated. The adjective cannot.
| Typical treatment | Assessed treatment | |
| Form of the finding | Adjective | Claim with a source |
| Timing | After site selection | Before commitment |
| Comparability across sites | None | Same scale everywhere |
| Confidence | Implied, uniform | Explicit, graded, variable |
| Owner | Communications, informally | Named, in the risk register |
| Updated | On crisis | On events, continuously |
Two Disciplines That Make an Assessment Usable
Assessment systems that survive contact with a sceptical development committee share two properties. Both sound pedantic and both are the reason the output gets used rather than filed.
Not a citation in a footnote that may have moved, but a retained snapshot of the page as it read on the day it was captured. Local news, council minutes and campaign sites are volatile, and an assessment whose evidence has evaporated cannot be defended in the meeting where it matters most.
Derived from how many independent sources support it, how varied those source types are, and how old they are. A finding supported by one blog post from 2019 and a finding supported by six sources from last month must not appear identical on the page. Uniform confidence is a form of dishonesty, and experienced reviewers detect it immediately.
Where the Model Belongs, and Where Arithmetic Does
The same division of labour that makes executive decision tooling credible applies here, for the same reason: the output is going to be used to argue against a large capital commitment, so it has to withstand a hostile read.
Language models are well suited to reading messy local reality: municipal minutes in three languages, regional press, planning documents, campaign material, transcripts of public meetings. That is genuine unstructured-text work and it is where the leverage is.
Aggregation is not a model task. Once sub-indicators are scored against anchored definitions, combining them into a dimension score and a headline number should be deterministic arithmetic. If a model performs the aggregation, two things follow: the number is not reproducible, and it cannot be calibrated against outcomes you already know. Both are fatal in a procurement conversation.
The scoring anchors matter as much as the arithmetic. A "seven on water" has to mean the same thing on every site, which means the scale needs to be defined against real reference cases rather than left to interpretation. This is the same argument we make about deterministic detectors in what AI-native actually changes in your operating model: put the model where judgement about text is needed, and keep it away from anything that produces a number someone will quote in a board paper.
The Score Is Not the Product
A risk score by itself changes nothing. It tells a development team that a site is difficult, which the local team frequently knew already, and provides no route forward. Used badly it becomes a reason to abandon sites rather than a means of derisking them.
The useful version pairs the assessment with the work that moves it: which commitments actually reduce which dimensions, by how much, and under what conditions. A global corporate pledge applied to a specific municipality should not count for the same as a locally audited commitment, because communities do not treat those as equivalent and neither should the model.
That produces a defensible sequence: this is where you stand today, this is where the commitments you have already made take you, this is where a structured engagement programme could take you. The gap between the first and the last is the actual business case, and it is a far stronger internal argument than a number on its own.
Why This Cannot Be Bought Off the Shelf
Every element of this is specific. The regulatory environment is national. The pattern of local opposition is regional. The reference cases that anchor the scoring are sector-specific, and the weighting between dimensions depends on the technology being deployed and where.
More practically, the people who hold the necessary judgement are usually development and permitting staff who have never worked on a data model and have no interest in doing so. Extracting what they know and encoding it into something mechanical is not a requirements-gathering exercise that can be conducted over three workshops. It requires someone sitting with the team, building against real sites, being wrong in the first iterations and correcting.
The practical test of whether the resulting system is any good is whether the domain experts can change it without engineering involvement. If updating a weight or an anchor requires a development ticket, the model has become an implementation detail and it will decay. If the scoring model is a configuration file that the permitting lead can revise directly, it stays alive. That is a design decision made early by someone embedded closely enough to know it matters, which is the point we make more generally in how to buy AI delivery that actually ships.
Where to Start
Take the projects that were delayed or abandoned on social grounds and establish what was knowable twelve months before the block. In most cases a substantial part of it was public at the time and nobody was tasked with looking.
Take your current pipeline assessments and require every statement about community risk to carry a number, a date and a source. The exercise itself surfaces how thin the current basis is, and it costs nothing but discomfort.
Calibrate against known outcomes before trusting the instrument on an unknown site. If the model cannot reproduce your own judgement on cases you know well, it is not ready to inform a decision on ones you do not.
Want Acceptance Risk Scored Before You Commit Capital?
We build the assessment inside your permitting workflow, with scoring anchors your domain leads can revise without opening an engineering ticket.
See how we work →Sources & References
- Daniel M. Franks, Rachel Davis, Anthony J. Bebbington, Saleem H. Ali, Deanna Kemp and Martin Scurrah, "Conflict translates environmental and social risk into business costs", PNAS 111(21): 7576–7581 (2014) — open access; source for every figure in the cost table above, including the ~US$20m per week of delayed production, the US$750m nine-month delay, the US$750,000 per day shutdown, and the 35–50% management-time diversion.
- Rachel Davis and Daniel M. Franks, "Costs of Company-Community Conflict in the Extractive Sector", Corporate Social Responsibility Initiative Report No. 59, Harvard Kennedy School (May 2014) — the companion report (45 confidential interviews, 50 public case analyses, fieldwork in Peru); source for the claim that social risk is systematically under-owned relative to engineering risk. Full PDF.
- Ian Thomson and Robert G. Boutilier, "Social license to operate", in P. Darling (ed.), SME Mining Engineering Handbook (3rd ed., Society for Mining, Metallurgy and Exploration, 2011), ch. 17.2, pp. 1779–1796 — source for the four-level social licence model (withheld/withdrawn, acceptance, approval, psychological identification) and the legitimacy/credibility/trust transition criteria.
- Maarten Wolsink, "Wind power and the NIMBY-myth" (Renewable Energy, 2000) and Rolf Wüstenhagen, Maarten Wolsink and Mary Jean Bürer, "Social acceptance of renewable energy innovation" (Energy Policy 35(5), 2007) — source for the claim that local opposition is poorly explained by simple proximity effects.
- Practitioner observation, SUPALABS engagements 2024–2026. The assessment design principles, the evidence and confidence disciplines, and the calibration approach are drawn from our own client work. This is unpublished, proprietary and not independently verifiable — treat it as informed opinion rather than evidence, and weigh it accordingly against the sourced claims above.
Domande frequenti
Innovation9 min2026-07-29EN
Condividi questo articolo
