AI in Planning Law: Drafting Representations and Appeal Statements

A planning law practice can use AI to draft representations and appeal statements safely by keeping the model away from policy memory. Paste in the policy text yourself, ask the AI to arrange arguments around only those extracts, and run every cited paragraph through a human citation ledger before filing.

The failure to guard against is specific: language models write policy references that look right. A plausible policy number, a paragraph that nearly says what you wanted, a decision name with the correct rhythm. None of it is checked by the model, so the checking has to be built into your process.

Follow me on Instagram@sagnikteaches

Where AI helps in a planning submission and where it must not

Planning drafting has two kinds of text. The first is structure and prose: turning a site history, a case officer's report and your analysis into an ordered argument. The second is authority: the exact wording and number of a development plan policy, a national policy paragraph, a supplementary document or a decision you rely on. AI is good at the first and unreliable at the second.

Connect on LinkedInSagnik Bhattacharya

So the rule for the whole workflow is that the model never supplies authority from memory. You supply the extract; it works with the extract. If a draft contains a policy reference that was not in the material you pasted, that reference is treated as invented until proved otherwise.

Subscribe on YouTube@sagnikteaches

Tasks that suit AI, in descending order of value for most practices:

• Issue extraction. Reading an officer's report or a refusal notice and listing each reason, the policy it cites, and the factual finding behind it.
• Restructuring. Turning a lawyer's rough notes into a statement with headings that track the reasons for refusal.
• Plain-language rewriting. Producing a version of a technical argument that a lay client or a residents' group can follow.
• Consistency checks. Comparing a draft against the plans schedule and the application description for mismatched floor areas, unit counts or heights.
• Representations to a neighbouring proposal. Drafting objection or support letters from a client's concerns, using only the policies you have supplied.

Tasks to keep with a lawyer: choosing the legal test, deciding whether a policy is out of date or carries reduced weight, assessing the strength of a ground, and any statement about what a decision-maker "must" do. Those are matters of professional judgement, and the broader question of what to hold back is covered in which legal tasks a small firm should never hand to AI.

Redacting client data before anything goes in

Planning files carry personal data in unexpected places: neighbours' objection letters with home addresses, a client's finances in a viability appraisal, health details in a personal-circumstances argument. Redact before you paste, even when the tool's terms say it does not train on your content, because redaction costs almost nothing and protects you if a setting is wrong.

A workable redaction pass takes about ten minutes per file:

1. Replace the client with a role label such as Appellant, and the site address with the Site.
2. Replace objector names with Objector 1, Objector 2 and keep a key in your case system, not in the prompt.
3. Remove application and appeal reference numbers; put them back in the final document by hand.
4. Delete anything privileged: counsel's opinion, without-prejudice exchanges, your own advice notes.
5. Keep the policy extracts, plan names and factual findings, because the model needs those to be useful.

On the tool side, use a plan whose terms exclude training on your content. Business tiers of ChatGPT and Claude do not train on business content by default, whereas consumer plans have a settings switch you must turn off yourself. Read the current terms before you rely on that, and see how to stop AI tools training on your business data for the settings route.

The citation ledger: the check that stops invented policy

The ledger is a table kept alongside every AI-assisted draft. One row per authority cited. It has five columns: the citation as it appears in the draft, the document it should come from, the page or paragraph you opened, whether it supports the point, and your initials with the date. A row is only complete when a person has opened the source document and read the paragraph.

The second and third rows are the kind of result the ledger exists to produce (the entries are illustrative). Numbering drifts when a national document is revised, and a model trained before the revision will quote the old number with confidence. A decision that looks like support can turn out to rest on site-specific facts.

Three rules make the ledger work. First, the person who checks is not allowed to rely on the AI to confirm its own citation; asking the model "is this paragraph real?" is not a check. Second, the ledger is completed against the current version of the document, so you note the version date. Third, any citation you cannot verify within a few minutes is deleted from the draft, not softened.

A worked example: a householder-style refusal appeal

An illustration, with generic details. A client wants to appeal the refusal of a two-storey rear extension. The refusal notice gives two reasons: harm to the character of the street scene, and loss of daylight to a neighbouring window. The case officer's report is nine pages, the local plan extracts you hold run to four pages, and the neighbour's objection is two pages.

Step 1, redact (10 minutes). Apply the redaction steps above. The client becomes the Appellant, the neighbour becomes Objector 1, and the address becomes the Site.

Step 2, extract the issues (15 minutes including review). Paste the refusal notice, the officer's report and the two policy extracts, and use a prompt like the one below.

You are helping a planning lawyer prepare an appeal statement.
Use ONLY the documents pasted below. Do not use policy numbers or
decisions from your own knowledge.

Task: for each reason for refusal, list
1. the policy the decision notice cites (quote the policy text I pasted)
2. the factual finding the officer relied on, with the report page
3. any gap between the policy wording and the officer's reasoning
If something is not in the pasted documents, write "NOT IN SUPPLIED
MATERIAL" rather than filling it in.

[REFUSAL NOTICE]
[OFFICER REPORT]
[POLICY EXTRACTS]

A realistic output lists two reasons, quotes the design policy extract word for word, and points out that the officer's report treats a 1.2 metre set-back as "overbearing" without referring to the policy's own test about street rhythm. That last item is the useful part. It is a gap for the lawyer to assess, and it is what a close reader would have found in an hour. Expect the model to miss subtle points about weight; that is your job.

Step 3, draft by ground (30 minutes). Ask for a first draft of Ground 1 only, using your bullet notes and the extracts. Review it before requesting Ground 2, so that an early error does not carry through the whole statement.

Step 4, ledger (20 minutes). A statement of this size cites perhaps eleven authorities. Opening each source and completing a row takes about two minutes apiece. In this illustration the ledger catches two problems: a policy number copied from the wrong version of the plan, and a sentence that says a policy "requires" something when the text says "should have regard to".

Step 5, lawyer's rewrite (as long as it takes). The draft is a scaffold. Your analysis of weight, the strength of the daylight evidence and the tone towards the inspector are written or rewritten by you. In this example the preparation comes to about 80 minutes before your own drafting, against what would otherwise be a blank page.

A second document type: representations to a live application

Representations to a live application are shorter, are read by a case officer who has dozens, and rise or fall on whether each point is tied to a policy. A good AI-assisted draft is a numbered list of concerns, each paired with the policy wording you supplied and one sentence on the harm.

Draft representations OBJECTING to the proposal described below, on
behalf of Objector 1 (a neighbouring occupier). Use only the policy
extracts I supply. Format: numbered points, each with
(a) the concern in one sentence, (b) the policy extract it engages,
quoted exactly, (c) the specific harm in plain words.
Maximum 600 words. Do not cite any policy not in the extracts.
If a concern has no matching extract, list it separately under
"UNSUPPORTED CONCERNS" so the lawyer can decide.

[PROPOSAL DESCRIPTION]
[CLIENT CONCERNS AS BULLETS]
[POLICY EXTRACTS]

The "unsupported concerns" instruction is the quiet hero here. Without it, a model tends to pair every concern with a policy, inventing the pairing where none exists. With it, you get an honest list of points that you either find a policy for or drop.

Sample output and what you would fix in it

Illustrative output for a concern about noise from a proposed rooftop plant enclosure might read: "Point 3. The plant would operate through the night and sits 4 metres from a bedroom window. Policy extract: 'Development should not cause unacceptable harm to the amenity of neighbouring occupiers through noise.' Harm: sleep disturbance."

What you fix: the "4 metres" came from the client's estimate, not a drawing, so you check it against the plans or remove it. The words "through the night" are an assumption from the client's notes; you confirm the proposed operating hours from the application documents. The quoted policy sentence is verified against the extract you pasted. Then it goes into the ledger like any other authority.

Four ways AI drafting goes wrong in planning work

Superseded policy quoted as current

The model recites a policy that was replaced when the plan was adopted. It reads correctly and is wrong. The ledger's "current version" note exists for this reason; always check the adoption date of the plan on the authority's own site.

Decisions that do not say what the draft claims

A model asked for "appeal decisions supporting a rear extension of this type" will produce names, dates and reasoning that may belong to nothing. Never ask for authority from memory. Search the decision portal yourself, pick the decisions, paste the paragraphs, and then ask the model to compare them with your facts.

Overstated tone

Drafts often assert that a proposal "clearly" breaches a policy. Inspectors notice. Ask the model for a version that states the policy test, the evidence and the conclusion in neutral terms, then add emphasis only where your analysis warrants it.

Facts drifting between sections

Over a long document the model may alter a dimension from 6.4 m to 6.5 m. After the draft, run a separate consistency prompt: list every number, date and dimension in the statement and compare them with the plans schedule you paste.

Before and after: a paragraph the ledger and a rewrite improved

Seeing the change on one paragraph makes the workflow concrete. Here is an illustrative first-draft sentence from the model, followed by the version that survived review.

First draft: "The proposal clearly contravenes the local plan's design policy and national policy on good design, as confirmed in several recent appeal decisions, and the Council's refusal was therefore plainly justified."

That is a bad sentence for an appellant, and the problems are typical. It concedes the point the client is appealing against. It cites "several recent appeal decisions" that nobody pasted in. It claims a breach of two policies without quoting either. Run through the ledger, nothing in it can be verified.

Reviewed version: "The design policy asks that development respects the established rhythm of the street. The extract in Appendix 2 sets out that test. The proposal is set back 1.2 metres from the neighbouring flank wall, which retains the gap between buildings shown on drawing 04. The officer's report does not assess the proposal against the test itself."

The second version is shorter, quotes its source, points to a drawing and states the gap in the officer's reasoning. Every factual claim in it can go into a ledger row. This is the pattern to aim for: the model supplies the shape of the sentence, you supply the evidence.

Using the model to review your draft as an opponent

Once a statement is drafted, a useful second pass is to ask the model to argue against it. This does not need new authority, so it carries little citation risk. Paste the statement and the officer's report, and ask for the three strongest points the council could make in response, using only the pasted material.

Read the appeal statement and the officer's report below. Acting as
the local planning authority's representative, list the three
strongest points you would make in response, each tied to a
paragraph of the statement. Use only the pasted material. Do not
add policy references of your own. End with any assertion in the
statement that has no supporting document.

A realistic output flags that the statement relies on a daylight claim without an assessment, that it describes the street as "mixed in character" without any photographs, and that Ground 2 repeats Ground 1. Each is a prompt for you to add an exhibit, cut a paragraph or soften a claim. The model is not a substitute for counsel's view, but it is a fast way to find the soft spots before a decision-maker does.

Who does what in a three-person team

A small planning practice does not need a formal process, but it does need named owners, or the ledger is skipped when deadlines bite. One workable split for a partner, an associate and a paralegal:

• Paralegal: assembles the supplied-material folder, performs the redaction pass, and prepares the first ledger with citations pulled from the draft.
• Associate: runs the drafting prompts, completes the ledger checks against source documents, and writes the analysis of weight.
• Partner: reads the final statement and the completed ledger together, and signs off. A statement without a completed ledger does not go to the partner.

In a two-person practice the same person may wear all three hats, so insist that the ledger is completed on a different day from the drafting where possible. Fresh eyes catch a wrong paragraph number far more reliably than the person who has just typed it.

A quick sum on time and cost

Take illustrative figures for one householder-style appeal statement. Without AI, a fee-earner might spend five hours: one reading the file, one planning, two drafting, one polishing. With the workflow above, redaction takes 10 minutes, issue extraction 15, drafting by ground 30 to 45, the ledger 20, and your own rewrite perhaps an hour and a half. That is around three hours, a saving of about two. At an individual plan price of $20 a month the software cost per statement is trivial; the saving only holds if the ledger step is never skipped. If a single wrong citation costs you an hour of repair and a dent in credibility, the 20 minutes is cheap insurance.

A short routine to keep each file honest

1. Create the matter folder with a Supplied material sub-folder holding the policy extracts you will paste.
2. Redact, then paste only what the task needs.
3. Draft one section at a time with the "only the supplied material" instruction.
4. Complete the ledger before anyone outside the team sees the draft.
5. Save the final prompt and the ledger on the matter file, so a colleague or the client can see how the document was produced.

If your team drafts several statements a month, a shared prompt library with the "NOT IN SUPPLIED MATERIAL" instruction built in will save repeated set-up. For help deciding between a general assistant and a legal-specific product, see ChatGPT or a legal AI tool for a small firm. Costs for an individual plan are modest (ChatGPT Plus and Claude Pro are both listed at $20 a month on their pricing pages), but the real cost is the checking time, which is why the ledger is kept brief.

How to tell the workflow is working

Measure two things for the first ten matters. First, the number of ledger rows that failed (wrong number, wrong version, unsupported claim). If that count is falling, your prompts and supplied material are improving. If it is zero for every matter, you are probably not checking hard enough. Second, the time from instruction to a reviewable draft. A reasonable target for a statement of this size is a first draft in under two hours of fee-earner time, with the ledger included.

Also keep a record of near misses. A single line per incident ("model cited a superseded policy number; caught at ledger") is enough to show the team why the check exists, and it is the evidence you would want if a client asked how you control AI-assisted drafting. A wider introduction process is in what your staff need to know about AI.

One last practical point: disclose AI use in line with your engagement terms and any guidance from the body you file with, and keep the lawyer's name on the document as the person who stands behind it. AI changes the drafting speed, not who is accountable.

Questions planning lawyers ask before using AI on drafts

Can I paste a client's planning file into ChatGPT or Claude?

Only on a plan whose terms stop your content being used for training, such as a business or enterprise tier, and only after redaction. Strip names, addresses, application references and any privileged advice first. If the tool does not give you a written data-handling statement you can show a client, keep the file out of it.

Should the appeal statement say that AI was used?

Check the rules of the body you are filing with and your engagement letter. Where a tribunal or inspectorate has issued guidance on AI-assisted documents, follow it. Whatever the rule, you remain responsible for every sentence and citation, so the disclosure question matters less than the checking.

How many policy citations should I verify by hand?

All of them in anything that is filed. A sampled check is fine for internal working notes, but one wrong policy number in a statement of case costs credibility with the decision-maker. The citation ledger described above makes a full check take minutes, not hours.

Does a legal AI tool remove the invented-citation problem?

It reduces it where the tool answers only from a closed library, but it does not remove it. Any tool that summarises or paraphrases can misstate a paragraph. Treat tool output as a draft that still goes through the ledger check.

Further reads

• How to Stop AI Inventing Case Law: A Small-Firm Checking Routine — A general checking routine for any AI-assisted legal drafting.
• AI Implementation Plan for a Small Law Firm: The First 90 Days — A 90-day rollout for the whole firm.
• Per-Seat or Pay-As-You-Go? Legal AI Pricing for Small Firms — Per-seat against pay-as-you-go costs.
• How to Summarise Bundles and Transcripts With AI Safely — Safe summarising of long bundles before you draft.
• How Small Law Firms Use AI to Draft Letters and Routine Documents — Routine letters and documents in a small firm.
• AI Hallucinations Explained for Business Owners: Causes and Fixes — Why models invent references in the first place.
• How to Use AI in a Law Practice: A Four-Week Starter Plan — A week-by-week rollout for a small law practice: write the policy, pilot one low-risk task, verify every citation, then settle client disclosure and review.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: ChatGPT pricing and ChatGPT Business pricing pages, OpenAI Data Controls FAQ, Claude pricing page and Anthropic privacy articles on model training, all checked 4 November 2026; standard practice for citation verification.