Three AI Agents, One Town Planning Consultancy, 64% More Applications Submitted

In September 2026, a five-person Bristol town planning consultancy was preparing planning applications for 14 clients a month and turning away work because they had no capacity left. The bottleneck was not expertise — it was time. Thirty-one hours a week were disappearing into planning statement drafting, heritage impact assessments, and pre-application correspondence before a single site visit happened. Three AI agents changed the maths. Within eight weeks, the practice was processing 23 applications a month, had recovered 26 hours of professional time weekly, and was running their agent stack for £80 per month. <figure> <img src="https://images.unsplash.com/photo-1486325212027-8081e485255e?w=1200&q=80" alt="AI agents for town planning consultancy UK — three AI agents help a Bristol planning practice submit 64% more planning applications by automating planning statement drafting, heritage reports, and pre-application correspondence" width="1200" height="630" loading="lazy" /> </figure> <h2>The 31-Hour-a-Week Admin Problem</h2> <figure> <img src="https://images.unsplash.com/photo-1450101499163-c8848c66ca85?w=1200&q=80" alt="Town planning consultancy admin burden — planning statements, heritage impact assessments, design and access statements, and pre-application letters consuming 31 hours weekly before AI automation was introduced" width="1200" height="800" loading="lazy" /> </figure> The practice — three chartered town planners and two planning assistants — handled residential extensions, change of use applications, listed building consents, and strategic planning work for property developers and homeowners across the South West. Each job followed roughly the same process: site visit, client brief, desk research, drafting, submission, and then weeks of correspondence with local authority planning officers. The problem was not the planning judgement. That part took, on average, three to four hours per job and was where the team's expertise actually mattered. The problem was everything surrounding it. A planning statement for a modest change of use application took four hours to draft from scratch — site context, planning history, policy compliance, material considerations. A heritage impact assessment for a listed building job took five. A design and access statement another two. Multiply that across 14 active applications per month, add the pre-application correspondence with councils, the conditional discharge requests, the responses to planning officer queries, and the fortnightly status reports to clients — and you had 31 hours a week of skilled professional time spent on document production rather than planning practice. The principal had tried document templates. They helped for the first paragraph of each section and were mostly ignored after that because each site, each council, each policy context was genuinely different. The planners were not wrong to write from scratch each time. They were spending expert hours on the scaffold around their expertise rather than on the expertise itself. <blockquote>The planners were not wrong to write from scratch each time. They were spending expert hours on the scaffold around their expertise rather than on the expertise itself.</blockquote> <h2>The Three-Agent Architecture</h2> <figure> <img src="https://images.unsplash.com/photo-1518770660439-4636190af475?w=1200&q=80" alt="Three AI agent architecture for town planning consultancy — Planning Document Writer, Pre-Application Intelligence Agent, and Project Status Agent working together to automate the planning workflow end to end" width="1200" height="800" loading="lazy" /> </figure> The build took eight weeks. The principal wanted three things: faster document drafts, better handling of planning officer correspondence, and a system that told clients what was happening without the team having to write status emails. Those three wants became three agents. Agent 1: The Planning Document Writer. This agent produces first drafts of planning statements, design and access statements, heritage impact assessments, and conditional discharge requests. The planner fills in a structured brief — site address, application type, key planning considerations, relevant local plan policies, heritage designations, any notable planning history. The agent uses that brief, combined with a knowledge base of national planning policy, local plan extracts, and the practice's document library, to produce a complete first draft in under four minutes. The drafts are not final documents. They are structured, policy-referenced starting points that cover every required section. A planner reviews and amends where professional judgement is needed — usually 45 minutes of editing rather than four hours of writing. The saving per document is between 75% and 85% of the original drafting time. Agent 2: The Pre-Application Intelligence Agent. Planning applications generate correspondence — planning officer queries, requests for further information, neighbour objection summaries, committee reports. This agent monitors the practice's email inbox for incoming council correspondence, classifies each item, extracts the specific questions or concerns raised, and drafts a structured response with policy references and suggested answers. The planner reviews the draft response, confirms or adjusts the planning judgement, and sends. Where the query requires a new drawing or specialist consultation, the agent flags it as requiring external input and creates a task in the practice's project management system. What previously took 90 minutes per query — reading, researching policy, drafting, checking — now takes 20. The agent handles the structure. The planner handles the substance. Agent 3: The Project Status Agent. This agent runs every weekday morning. It checks the relevant council planning portal for each active application, retrieves the latest status, decision notices, and any new documents uploaded by the authority, and generates a client status update for each active job. The update goes to the planner for a 30-second review before sending. For straightforward updates — "your application is validated and awaiting officer allocation" — the planner approves with one click. For updates that require explanation or action, they add a sentence and send. Clients who previously received a fortnightly status call now receive a twice-weekly email update. Response rates on client surveys improved. The team stopped spending Monday mornings chasing portal logins and drafting "nothing to report" updates. <h2>The Numbers After Eight Weeks</h2> <figure> <img src="https://images.unsplash.com/photo-1590283603385-17ffb3a7f29f?w=1200&q=80" alt="Results from AI implementation at Bristol town planning consultancy — 64 percent more applications submitted per month, 26 hours of professional time recovered weekly, monthly cost of £80 to run the three-agent stack" width="1200" height="800" loading="lazy" /> </figure> Eight weeks after the three agents were live, the principal ran the numbers. <ul> <li>Applications submitted per month: 14 → 23 (64% increase)</li> <li>Professional time recovered per week: 26 hours</li> <li>Monthly running cost: £80 (n8n hosting + Claude API + planning portal integrations)</li> <li>Average document draft time: reduced from 4 hours to 45 minutes</li> <li>Pre-application query response time: reduced from 90 minutes to 20 minutes</li> <li>Client status updates: fortnightly → twice-weekly, at no additional staff time</li> </ul> The 26 recovered hours did not become downtime. The practice took on five new clients — four residential developers and a housing association — that they had previously been turning away. The revenue impact was direct. The agent stack cost £80 per month to run. The additional five clients generated substantially more than that in monthly fees within the first billing cycle. The principal noted one outcome they had not anticipated: the consistency of documents improved. Every planning statement now covers the same policy sections in the correct order. Every heritage impact assessment follows the same analytical structure. The agents do not replace the planning judgement — but they ensure the judgement is never lost inside a document that omits a material consideration. <blockquote>Every planning statement now covers the same policy sections in the correct order. The agents do not replace the planning judgement — but they ensure the judgement is never lost inside a document that omits a material consideration.</blockquote> <h2>What Made It Work</h2> <figure> <img src="https://images.unsplash.com/photo-1503387762-592deb58ef4e?w=1200&q=80" alt="Three success factors for AI agent implementation at a UK town planning consultancy — knowledge base assembled first, human review at every output stage, and sequential agent deployment starting with the highest-volume task" width="1200" height="800" loading="lazy" /> </figure> Three decisions made the difference between a build that worked and one that would have failed. First, the knowledge base was assembled before the agents were built. The practice spent two weeks uploading local plan policies for their five most common local planning authorities, the National Planning Policy Framework, their own document templates, and a library of previously approved applications. Agents without context produce generic outputs. The investment in the knowledge base was what made the Planning Document Writer produce documents specific enough to be genuinely useful rather than generic text requiring complete rewriting. Second, every agent output has a human review stage before it goes anywhere. No document leaves the practice without a planner reading it. No correspondence is sent without approval. This is not a limitation — it is the design. The agents handle the production. The planners handle the professional responsibility. The <a href="/blog/human-in-the-loop-ai-agents-uk">human-in-the-loop architecture</a> is not a workaround for AI limitations. It is the appropriate design for a regulated professional services context, and it is what gave the principal confidence to deploy agents handling documents that affect their clients' planning permissions. Third, the build started with the highest-volume, most repetitive task — planning statement drafting — and proved the model before adding more agents. The temptation in any AI build is to automate everything simultaneously. The practice started with one agent, ran it for two weeks, measured the saving, and then built the second agent with the confidence of having seen the first one work. This sequenced approach is why the <a href="/blog/ai-pilot-trap-uk-production-gap">pilot trap</a> was avoided — because the first agent was production-grade before the second was started. The practice is now building a fourth agent to handle Section 106 agreement drafting — a task that previously required a specialist to work from a local authority standard template, adapting obligations to the specific application. Early tests show the same pattern: strong first drafts in under five minutes, 70% time reduction per agreement, and no reduction in the quality of the planning judgement embedded in each one. Town planning is a knowledge profession built on policy, precedent, and professional judgement. None of those things can be automated. The admin that surrounds them can be — and in this practice, it has been. If you want to understand whether the same architecture applies to your firm, <a href="/contact">book a free 30-minute call</a> and we will work through it with you.