§1What happened
A complaint about a missed bin collection used to fit on one side of A4. The BBC reports it now arrives at up to 20 pages, quoting case law and Acts of Parliament, because the person filing it ran it through a chatbot first. That is not a one-off. Schools, councils and other public bodies told the BBC they are facing a rising tide of AI-assisted complaints, arriving at a length and complexity their handlers were never staffed for. The tools that used to sit between a frustrated resident and a formal, cited, multi-page grievance, the effort, the legal knowledge, the time, have been removed. What is left is a text box and a model that will fill it.
Deborah Evans, chief executive of Lawyers in Local Government, which represents 5,600 lawyers working in local authorities, put a number on it. The increase is "massive", and she estimated her organisation would need double the current workforce to deal with it. Her example was a complaint that could run "between 19 and 27 pages", where a year ago it was a single sheet. The backdrop is an adoption wave: research from Granicus found 57 per cent of public-sector bodies are embedding or experimenting with AI services, while 69 per cent of citizens do not know AI is being used on them, and the Information Commissioner's Office issued guidance on AI-generated freedom-of-information requests in May. Notice what Evans reached for as the fix. Not a cleverer bot. Forms, to cap how long a complaint can be. Hold that thought, because it is the whole article.
§2The generating side just got an engine
Every process that handles volume has two sides. One side generates the work: the complaint, the ticket, the FOI request, the sales objection, the pull request. The other side resolves it: someone reads it, decides, and closes it. For as long as anyone has built these systems, both sides moved at roughly human speed, and you could staff the resolving side to match the generating side because they grew together. Generative AI has just broken that symmetry. It handed a large throughput multiplier to the generating side, and to that side only.
This is the part worth saying plainly, because it is not a story about councils and it is not a story about complaints. Anyone who feeds your queue can now produce more, longer, and denser input in seconds, for free. The person or system that has to read each item, understand it, and act on it did not get the same multiplier. That gap is the entire phenomenon. A council complaints desk is just the first place it showed up clearly, because complaints are free text, they are adversarial, and someone is legally obliged to read every one. If your product has an intake queue that accepts prose, you are on the same curve, a little behind.
§3A bot on the front door does not fix an intake problem
The reflex, when the queue floods, is to answer the machine with a machine: put a chatbot on the front to deflect. It is the obvious move and it treats a structural problem as a user-interface problem. A conversational veneer can answer a frequently-asked question, and for genuinely routine traffic that is a real win, some councils have cut routine call volume exactly this way. But a bot answers questions; it does not resolve state. The expensive work, deciding whether a complaint is upheld, whether a placement is right, whether a refund is owed, is untouched, and it is now arriving pre-inflated with pages of generated argument the bot has no authority to dispose of.
Deflection assumes the load is made of questions. It is not. It is made of unresolved cases, and a case a bot cannot close does not go away, it comes back, bigger, often with the bot's own brush-off quoted back at you as evidence you failed to engage. So the front-door bot moves the queue, it does not shorten it, and for the hard cases it can make things worse by adding a round trip. You cannot deflect your way out of a volume problem when the volume is real work that a human still has to finish.
§4Structure is the lever, and the lawyer already named it
Evans's instinct is the tell, and it is correct. The fix is not a smarter model, it is a form. A free-text box is where load hides. It accepts anything, at any length, in any shape, which means every single item has to be read in full by a person before it can even be sorted, let alone answered. That is the property AI just weaponised: unbounded input, read by hand. A structured form does the opposite. It forces the input into fields you can route, deduplicate, rank and, above all, bound. The twenty-page essay collapses to "bin, missed, this address, this date, photo attached" the moment the input is structured instead of prose, and the twenty pages become an optional attachment nobody has to read to act.
That is not hostile to the person complaining. It is the only thing that lets you answer them at all, at volume, in time. And it quietly dissolves the panic about AI-written submissions, because once the load-bearing content is captured as typed state, it stops mattering whether a human or a model wrote the paragraph around it. The builder lesson is old and it keeps getting relearned each time a new input engine arrives: if you have to process something at scale, do not let it reach you as free text. Capture it as structure at the boundary, where the volume has to pass through, and you have turned an unbounded reading problem into a bounded routing one.
§5How we run it
This one is close to home, because a fleet of agents is the purest version of the same asymmetry. An agent can generate an unbounded wall of prose far faster than anything downstream can read it, and if you let agents hand each other essays, you build exactly the flood the councils are drowning in, at machine speed, inside your own system. So our agents do not pass prose. Work moves as structured output: typed schemas at the tool boundary, validated objects rather than free text, cards on a board with fixed fields, a state the next agent can route without re-reading a paragraph to find the one fact it needs.
When an agent of ours has to return a decision, it returns it in a shape the next step can act on, not a description the next step has to interpret. That is the same move the council lawyer is asking for, made one layer down and enforced in code. The generating side of every interaction now has an engine bolted to it, and it is not slowing down. You do not win that race by reading faster or by hiring the second desk Evans says she needs. You win it by refusing to let the work arrive as prose in the first place, and by putting the structure at the boundary the volume must cross, not on a bot at the front that can only wave it through.