Can AI Write a Flood Risk Assessment?

Ask any chatbot to write a flood risk assessment and it will hand you a confident, well-formatted document in seconds. That is the easy part, and it is also the misleading part, because a flood risk assessment is not really a writing task. It is an evidence task. The words are the wrapper around survey data, modelling, current policy and professional judgement, and the wrapper is the smallest part of the job.

Writing is the last 10% of a flood risk assessment

Most of the effort in a flood risk assessment for planning goes in before a single sentence is written. It is gathering current flood data for the site, screening every source of flooding, building or commissioning a model where the maps are not enough, designing mitigation that works, and applying national and local policy to the specific scheme. Only once that evidence exists does anyone write it up. A language model can produce the write-up. It cannot produce the evidence underneath it, and an assessment with a polished write-up over thin or wrong evidence is exactly the kind a consultee sends back.

What a planning authority actually checks in an FRA

The Environment Agency, as statutory consultee, and the lead local flood authority do not grade an assessment on how well it reads. They test it against specific requirements, and these are the points where an AI-written report tends to come apart:

1. Current data and the right climate change allowances. Allowances vary by catchment and epoch, and the maps move: the Flood Map for Planning was updated on 28 May 2026 with surface water climate change extents on the upper-end 2070s allowance.
2. Every source of flooding. Since the September 2025 PPG update, the sequential test must consider surface water, groundwater and sewers, not just rivers and the sea.
3. A genuine sequential and, where needed, exception test, argued correctly for the site. The Yatton judgment (North Somerset DC v Secretary of State, EWHC 1430) shows how finely the planning balance can turn.
4. Safe for the development's lifetime, with no increase in flood risk elsewhere, evidenced by finished floor levels, safe access and a workable drainage strategy.
5. A competent, accountable author. Someone qualified and insured has to stand behind the conclusions, including at appeal.

Where an AI-generated FRA fails those checks

Mapped against that list, the gaps are not stylistic. They are structural:

• On data, a model reaches for figures from its training material, which lag the current maps and allowances. It will quote last year's numbers with full confidence and no flag that they have changed.
• On sources, it defaults to rivers and the sea, the flooding that dominates the text it has read, and under-reads surface water and groundwater, which is where many sites are actually constrained.
• On site-specific evidence, it cannot survey the ground, monitor a water table, or build and defend hydraulic modelling to the EA's river modelling standards. It can describe a model; it cannot produce one a reviewer would adopt.
• On policy, it can recite the sequential and exception tests but cannot weigh them for a real scheme, where the outcome turns on site facts and the planning balance rather than on the wording of the policy.
• On accountability, it cannot be a competent person, hold professional indemnity, or answer for its work if flooding is later disputed.

Where AI does help in flood risk work

None of this makes AI useless to the people preparing assessments. In a consultant's hands it is a reasonable drafting aid: structuring a first draft, condensing long policy into plain English, or explaining a term to a client. The rule that keeps it safe is simple. Treat anything it produces about flood levels, allowances, modelling or policy as a claim to be checked against the primary source, never as a finding. The judgement stays with the engineer; AI just speeds up the typing around it.

Is AI used in flood risk work at all?

Yes, and it is worth separating from the question of writing reports. Tools such as Northumberland County Council's FloodAI project use machine learning for earlier flash-flood warnings, and AI is being trialled alongside planning officers to flag issues faster. We look at this in full in how AI could improve flood risk management. In every credible case the pattern is the same: AI supports the system and a person remains responsible for the decision. Writing an FRA that carries that responsibility is not one of the things it can take over.

What happens if you submit an AI-only FRA?

The practical cost lands later, not at submission. A report that misses a source of flooding or uses superseded allowances draws an objection or a holding direction from the consultee, which means delay while it is reworked, and in the worst case a refusal that has to be fought at appeal with the assessment redone properly anyway. With around 6.3 million properties in England now at flood risk, scrutiny of flood risk assessments is tightening, not easing, so a document that cannot survive that scrutiny is a false economy.

Frequently asked questions

Talk to a flood risk consultant who can stand behind your FRA