AI for ecology consultancies: you already know the rule

Ecology consultancies are not short of work. They are short of hours. So the useful question about AI is not whether it cuts your costs, but whether it hands back the scarcest resource you have: a qualified ecologist's hours.

The demand is real and rising. In a 2025 Home Builders Federation survey, nine in ten respondents had planning applications delayed because local authorities lacked the expertise or resources to handle Biodiversity Net Gain (BNG), and nearly four in ten councils have no in-house ecologist at all.

Two worries hold ecologists back, and both deserve a straight answer. One is whether the tool can be trusted with the work. The other, sharper for this profession, is whether using energy-hungry AI is consistent with the values that brought you into ecology in the first place.

The demand is not the problem. Your capacity is.

It is worth being precise about the market, because it changes what a practice should do.

The Home Builders Federation surveyed the first year of mandatory BNG and published the results in April 2025. They are blunt. 98% of the small and medium-sized builders surveyed find BNG a challenge, and across all respondents, 94% have had planning applications delayed by it, with 90% pointing to a lack of BNG expertise or resources inside the local authority. Nearly 40% of councils have no in-house ecological expertise, a quarter lost ecologists in the past year, and each council now spends an average of £23,000 a year buying that expertise back in from consultants.

Every one of those numbers is unmet demand for ecologists. The whole system is limited by the number of competent-ecologist hours available, and there are not enough of them. CIEEM's 2025 State of the Profession survey reports the same from inside: a shortage of people experienced enough for senior roles, and the workload landing hardest in the April to October survey season.

So the honest framing of AI here is not cost-cutting. A two-person consultancy does not have a cost problem. It has a capacity problem, and every hour spent formatting a report is an hour not spent in the field, or on a judgement only a qualified ecologist can make.

The exception: the smallest sites, where BNG is being lifted

One part of the work is moving the other way, and the change is close. From 6 August 2026, developments of 0.2 hectares or below are generally exempt from mandatory BNG, unless they harm priority habitat, the habitat types singled out for conservation. Separately, other small developments that still fall under BNG can use the simpler small sites metric without an ecologist, run by the project manager, a gardener or a landscape architect. Together these thin the routine, lower-value end of BNG work.

Two caveats matter. "You do not need an ecologist" applies to the metric, not to the site's wider ecology, which can still bring protected species, designations and surveys of its own. And judging whether the priority-habitat exception applies is itself an ecological call, so a qualified ecologist is often still worth involving before anyone relies on the exemption.

So the strategy is clear. As the routine BNG work on small sites thins, what is getting scarcer and more valuable is the licensed, named judgement a planning authority will actually accept. A practice that uses AI to shave time off its cheapest work is solving the wrong problem. A practice that uses it to free hours from admin and drafting, then puts those hours into survey capacity and judgement, is moving toward the part of the market that is growing.

That is the case for AI here. It also invites the fair objection this profession, of all professions, should raise: is it not hypocritical to answer an environmental job with an energy-hungry machine?

Is it not hypocritical to use AI?

It is a fair question, and it deserves an answer rather than a dodge. Whether to use these tools at all is a choice each practice makes for itself, against its own values, and nobody can make it for you. What we can do is separate the part that is overstated from the part that is real.

Start with the overstated part. The figures that circulate about a single prompt are wrong, and the worst of them deserve a factual rebuttal. The claim that one prompt uses a bottle of water comes from a single 2023 estimate of an older model, and has been exaggerated with each retelling. The line that a prompt costs ten times a web search rests on assumptions a 2025 analysis has since called an overestimate. The measured reality is far smaller. Google's August 2025 disclosure puts a typical text prompt to its Gemini model at 0.24 watt hours of energy and about 0.26 millilitres of water, roughly five drops, and that figure includes the cooling and overhead older estimates left out. Independent work on ChatGPT lands around 0.3 watt hours a prompt. The estimates still vary by a factor of ten depending on the model and how the sum is done, so treat any single number as a range, but the order of magnitude is settled, and it is small.

Now the part that is real. The aggregate is rising fast. The International Energy Agency puts data centres at about 415 terawatt hours in 2024, roughly 1.5% of global electricity, and expects that to more than double to around 945 terawatt hours by 2030, with AI the biggest driver of the rise. A tiny cost per prompt, multiplied by a fast-growing number of prompts, is still a large and rising draw on power and water. Your own use is tiny. The industry's total is not.

The number that should govern your decision is not the global total, which you do not control. It is the footprint of the use you would actually make.

The survey-drive check

Here is a comparison any practice can run for itself. Take a heavy month of ordinary text prompts, 1,000 of them at Google's reported median, which is 30 to 50 every working day. Set that against a single site visit in an average petrol car, using the government's 2025 emissions factors. Uploading long documents or running reasoning-heavy prompts costs more, so this is a short-text illustration, not a figure for every kind of use.

The figures are illustrative, but the difference is large. That one drive emits several hundred times the carbon of a whole month of short-text prompts. The point is not that AI is free. It is that for a practice whose real footprint is fieldwork, travel and rework, the tool is a minor part of the total, and the hours it frees can cut the big parts.

There are two cautions. The rebound effect, also called Jevons paradox, is the tendency for efficiency to raise total use rather than lower it: cheaper drafting can simply mean more drafting. And the profession is reaching a sensible position. A 2026 paper in Frontiers in Ecology and the Environment argues the debate should move past treating AI as simply good or bad, toward what it calls conscientious computation: choosing the lowest-impact method for the job, not the fashionable one. CIWEM, the water and environmental management institution, gives the practical rule for reports: if AI has been used, say so. Measure it, keep it in proportion, and be open about where you used it.

The uses that can cut site visits, not just admin

Everything so far treats AI as a way to save office time. There is a more interesting case, specific to this profession: a few AI tools can cut the biggest part of your footprint, which is the driving to and from sites.

For a field practice, the heavy items are vehicles and repeat site visits, not the electricity a laptop draws. That is a reasonable inference rather than a measured figure, but a sensible one, and CIEEM treats travel as a headline category in its own carbon planning. If AI takes drives out of the year, that saving is far larger than anything the tool costs to run.

Three uses point that way, and none is a chatbot. Each takes a volume of data no person could work through by hand and makes it usable.

1. Automatic call classification. A static detector left recording for many nights produces more audio than anyone can listen to. The BTO Acoustic Pipeline classifies bats, small mammals, bush-crickets and more from those recordings, and tools like BirdNET and Kaleidoscope do the same for birds and bats. In the right survey design, the detector can replace some attended nights on site, and the classifier is what makes it worth collecting that much audio in the first place.
2. Camera-trap triage. In a typical camera deployment most frames are empty, tripped by wind or rain. Open tools like Microsoft's MegaDetector and platforms like Wildlife Insights filter the blanks and flag the animals, turning weeks of review into an afternoon. A European automated camera network reported over 40% cost savings, partly through less regular site visits.
3. Remote habitat screening. Natural England's Living England map and the UK Centre for Ecology and Hydrology's Land Cover Map use machine learning on satellite imagery to map habitats nationally. They are a predictive baseline, not a substitute for the walkover, but they sharpen scoping, add landscape context, and on larger sites can cut how much ground you have to cover blind.

The limits matter. These tools still need ground data to train and check, so they redirect field effort rather than removing it, and a person still confirms every result, exactly as with a bat classifier. Even so, the sharpest environmental case for AI in ecology is not the tiny cost of a prompt. It is that a few of these tools, used deliberately, can take a survey drive out of the diary. That leaves the question that governs everything: can you trust what the tool hands back? Here ecology has an advantage, because the profession already has a rule for exactly this.

You already know the rule

Ecologists have worked with machine classification for years. Point an automated classifier at a night of bat recordings and it tells you what it thinks it heard. Nobody in the profession treats that output as a finding, and they are right not to.

A 2024 study in PLOS ONE tested three commercial bat call classifiers and found accuracy varies enormously by program and by species. The programs were good at avoiding false positives. What they were bad at was the thing that matters most: for the endangered Myotis species, they correctly picked up a real call between 1% and 52% of the time, depending on the species and the software. A classifier that confidently reports nothing while missing half the calls of the species carrying your greatest legal exposure is not a small problem.

The authors' conclusion is the sentence to keep: a qualified analyst should verify automated classifications before species-specific conservation, regulatory and permitting decisions. That study is on North American species, so the numbers do not transfer to the bats you survey, but the lesson transfers exactly.

So the profession already runs a good rule for machine output:

Let the tool do the first pass. A qualified person checks every result before it goes on the record. The tool's raw output is never the finding.

That is the correct rule for ChatGPT too. Practices are treating generative AI as a strange new thing that needs a strange new policy, when they have followed the right policy for a decade. You do not need to invent an AI position. You need to notice you already have one, and apply it to the tool that writes sentences instead of the one that reads calls.

There is one difference, and it is the only one that matters. A bat classifier tells you when it is unsure: it gives you a confidence score, and a file you can open and listen to. A language model gives you neither. It may hedge in words, but not with a calibrated score you can check against a source, so its assured prose reads the same whether it is right or wrong. The check has to be yours.

Where it genuinely helps

Start with what AI does not touch, because it is most of the job. The actual practice of ecology runs on fieldwork and judgement: protected-species surveys and the licences that follow them, Ecological Impact Assessment, Habitats Regulations Assessment for internationally designated sites, Ecological Clerk of Works supervision on site, and years of post-consent monitoring. The tool does none of it. What it helps with is the writing and admin wrapped around the work, which is where the recoverable hours are.

Turning field notes into a first draft. This is the strongest use in the sector by a distance. A structured set of survey notes, target notes, conditions and times becomes a competent first draft of the survey write-up in minutes instead of an afternoon. That is the walkover half of a Preliminary Ecological Appraisal, not the whole report. The records-centre search and the designated-site data behind it come from real searches the model cannot make up, and the draft is checked against them. Nothing is being decided; the typing is being done, and the typing is a real cost.

Standard, reusable text. Report structure, standing legal and policy background, and established method descriptions are house wording you can keep in a controlled library and reuse. The parts that only look standard, the site-specific methodology, the survey limitations and the mitigation, still have to match what you actually did and found. Reuse the settled text, and do not ask a model to invent the rest.

Desk study and literature review, with one condition. Pulling together policy context, designations and species ecology is quick work for AI. The condition is that every citation and legal reference gets opened and checked, for reasons the next section makes plain.

Client and planning correspondence. Explaining to an impatient developer why the required emergence survey cannot be completed in February is drafting, not judgement: high volume, low stakes, easy to check.

The unbillable back office. Proposals, scoping responses, fee quotes, tender answers, and the CVs and project lists that go into every framework submission. None of it is billable, all of it is necessary, and it is what keeps a small consultancy working late.

Where it bites

Fluency is not evidence. The risk to watch is subtler than an AI writing your conclusion. It is an AI-drafted report that reads beautifully, describes the survey effort with total confidence, and quietly hides an ambiguity you noticed in the field. The draft sounds more certain than the survey was, and in a backlog, in October, that reads like a finished report rather than a warning.

The metric is arithmetic with rules, not prose. The statutory biodiversity metric is a defined calculation in a defined tool. Ask a language model for biodiversity units and it will give you a plausible number, because plausible text is all it makes. It does not understand the trading rules or the spatial multipliers, and a wrong figure goes straight into the planning submission, the management plan and the section 106 obligations. Use the tool the calculation is done in. Do not ask a chatbot to do it, or to check it.

Habitat condition is a field judgement. Condition is the input that most affects the numbers, and it comes from a competent person standing in the habitat applying published criteria. Nothing does that from a photograph and a description. Let AI infer a condition score and you have let it invent the basis of the whole assessment.

Fabricated references are not hypothetical. In June 2025 the High Court dealt with two cases together, Ayinde v London Borough of Haringey and Al-Haroun v Qatar National Bank. In one, five of the cases cited did not exist; in the other, eighteen authorities did not exist, and several of the real ones did not say what they were claimed to say. The lawyers involved were referred to their regulators. A planning officer is entitled to assume your references are real, so open every one.

A wrong protected-species call is not a typo. If a report says there are no bats and there are bats, the consequence is not a corrected paragraph. It can be a permission granted on a false basis and later challenged, works that amount to an offence under the Conservation of Habitats and Species Regulations 2017 or the Wildlife and Countryside Act 1981, a developer left with a stopped site, and a negligence claim aimed at whoever signed the report. Your professional indemnity cover and your licence-holder's standing with the regulator are what is at risk. The software vendor carries none of it.

Survey timing cannot be reasoned around. Bat activity, great crested newt, breeding bird and botanical surveys all have windows. Government guidance to planning authorities is that an application should not be decided until the necessary surveys are done, by a suitably qualified person, using the right methods, at the right time of year. If a tool ever seems to be helping you around a seasonal constraint, something has gone wrong.

The non-negotiables

• Apply your acoustic rule to your language model. Let it do the first pass, check every result yourself, and put only the checked result on the record.
• No AI in the chain that produces a protected-species conclusion, a condition score or a biodiversity unit figure. Drafting the report around those findings is fine. Producing them is not.
• Open every citation, reference and designation. Assume invention until you have checked.
• Client and site data goes into a business-grade account with training switched off. Survey data records where protected species are. That is sensitive information about a client's land, and in the wrong hands it could lead someone to a badger sett.
• Say where you used it. CIEEM's approach to generative AI in its own membership process is transparency rather than prohibition, and CIWEM advises stating AI use in reports. It is a defensible approach for report production, and a much stronger position if a report is ever scrutinised.
• Use it deliberately, not everywhere. Your own footprint is small but not nothing, and efficiency has a way of inflating total use. Point the tool at the admin and drafting that free real hours, and leave it off the work it cannot do anyway.

A sensible first month

1. Pick one recurring document, such as the PEA or a standard survey report: high volume, known structure, easy to check.
2. Write your house structure and standard sections down once, properly, and give that to the tool as its template. Draft quality comes from the context you give it, not the model you pick.
3. Use a business-grade account, not a personal login.
4. Draft only from your field data. If the notes do not support a sentence, it does not go in, and a model that invents one has just told you the notes were thin.
5. At the end of the month, ask two questions: how many hours came back, and did a single error reach a client? If the second answer is yes, stop and find out why before you scale it.
6. Note where you used it, so you can say so if a report is ever questioned. A line in your quality records is enough.

What does not change

The value of an ecology consultancy is a competent professional who went to the site, formed a judgement, and will put their name to it. That is the part becoming scarcer, more valuable and harder to replace, exactly as the simplest, smallest-site end of the work is being automated and deregulated. It is also the part AI cannot do.

So the practices that come out of this well will not be the ones that adopted the most AI. They will be the ones that were clearest about which parts of their work must never be handed to a machine, and used it hard on everything else.

What to do next

Take one report type, run it for a month, and count the hours. If it would help to work out where AI belongs in your practice, and where a person has to stay in charge, book a call. Before you spend anything, it is also worth knowing how to tell a good adviser from a good salesperson.

This is general information, not legal or regulatory advice. Biodiversity Net Gain rules, protected-species offences and a consultant's professional duties apply differently from one project to the next, so take proper advice on your own case.

Sources and further reading

• Biodiversity Net Gain: one year on, Home Builders Federation, April 2025. Source for the 98% figure among the small and medium-sized builders surveyed, the 94% and 90% figures across all respondents, the proportion of councils without in-house ecological expertise, ecologist losses from councils, and the £23,000 average annual council spend on external consultants.
• Understanding biodiversity net gain, GOV.UK. Source for the 10% requirement, the statutory and small sites metrics, and the exemption for developments of 0.2 hectares or below from 6 August 2026.
• Biodiversity net gain: exempt developments, GOV.UK. Source for the 0.2 hectare and related exemptions from 6 August 2026 not applying where on-site priority habitat is impacted.
• Solick, Hopp, Chenger and Newman, Automated echolocation classifiers vary in accuracy for northeastern U.S. bat species, PLOS ONE 19(6), 2024. Independent. Source for the variation in classifier accuracy, the low sensitivity for Myotis species, and the recommendation that a qualified analyst verify automated classifications before regulatory and permitting decisions.
• Ayinde v London Borough of Haringey and Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin), 6 June 2025. Source for the fabricated citations put before the court and the regulatory referrals that followed.
• Protected species and development: advice for local planning authorities, GOV.UK. Source for the expectation that surveys are done by suitably qualified people, using appropriate methods, at the right time of year, before a decision is made.
• State of the Profession Survey 2025, CIEEM. Source for the shortage of experienced senior ecologists and the concentration of workload pressure in the April to October survey season.
• Energy and AI, Executive Summary, International Energy Agency, 2025. Source for data centres at around 415 TWh, roughly 1.5% of global electricity in 2024, the projection to about 945 TWh by 2030, and AI as the most important driver of the growth.
• Measuring the environmental impact of AI inference, Google, August 2025. Source for the median Gemini text prompt at 0.24 watt hours, 0.26 millilitres of water and 0.03 grams of CO2, on a comprehensive measurement boundary.
• How much energy does ChatGPT use?, Epoch AI, February 2025. Source for the roughly 0.3 watt hours per typical query estimate, and for the older 3 watt hour "ten times a web search" figure being an overestimate based on outdated assumptions.
• Greenhouse gas reporting: conversion factors 2025, Department for Energy Security and Net Zero, June 2025. Source for the average petrol car emissions factor used in the survey-drive comparison.
• The risks and benefits of AI in environmental impact assessments, CIWEM. Source for the recommendation that use of AI be stated clearly in the report.
• Norman et al., the computational carbon footprint of ecology, Frontiers in Ecology and the Environment, 2026. Source for the "conscientious computation" argument that method choice, not AI adoption in the abstract, is the environmental lever.
• BTO Acoustic Pipeline, British Trust for Ornithology. Source for automated classification of bat and other species calls from static-detector recordings.
• Living England habitat map, Natural England (GOV.UK Algorithmic Transparency Record). Source for national machine-learning habitat mapping from satellite imagery.
• An automated wildlife monitoring network for Natura 2000, Basic and Applied Ecology, 2024. Source for the reported cost savings achieved partly through less frequent site visits.
• Action 2030, CIEEM. Source for travel being treated as a headline category in the profession's own carbon planning.