This is the short version of an argument I’ve set out at length in a paper of the same name - the evidence, the sector detail, and worked examples you can adapt. It’s here if you’d rather have all of it. Otherwise, read on.

The reason it takes hours is not the length. It is that before anyone can respond, somebody has to establish which of its assertions are true. The complaint cites legislation, references regulatory codes, sometimes invokes case authority, and reads as though a solicitor wrote it - and some of that will be accurate, some will be irrelevant, and some will be invented. A fabricated authority costs more to disprove than a real one costs to address, and the entire burden of disproof sits with you. Effort used to be the rationing mechanism and, worse, a proxy for merit: a long, legally literate complaint told you the sender was serious, had researched their position, and was unlikely to go away. Every triage process ever designed assumes that. It has just stopped being true.

So the instinct is to fight AI with AI. Build the filter, tighten the triage, answer at machine speed. It feels like an arms race, and organisations are budgeting for it as one. It isn’t a stupid instinct, either - when demand rises you automate, and that has worked for thirty years.

It will not work here, and the reason is structural rather than technical. An arms race is only worth entering if escalation costs both sides something comparable. Look at what it actually costs each side.

Every AI capability the complainant deploys is free, or costs about twenty dollars a month. It improves automatically, overnight, without being asked. It requires nobody’s approval, no integration, no assurance and no explanation to anyone.

Every AI capability you deploy needs procurement, integration with systems that were never designed for it, model risk assessment, data protection review, governance sign-off, staff training, and the ability to defend the whole thing to a regulator afterwards. That is months of work and a business case, repeated every single time the capability needs upgrading.

Your escalation is capital expenditure on a governance leash. Theirs is a subscription that upgrades itself. That is not an arms race - an arms race is symmetrical. This is a tax, and only one side is paying it. Every dollar you spend widening the gap gets closed for free at the next model release.

So organisations reach for the tools they already know. The three most obvious are, unfortunately, the three most likely to fail:

Detect the AI ones.

They cannot be detected reliably, and the way detectors fail is the problem. They flag plain, regular, structured writing - which is how people write when English is not their first language, when they are unsure of themselves and stick to safe constructions, and how a great many dyslexic and autistic people write, because consistency is what makes writing manageable. A detection-based screening policy would systematically deprioritise exactly the people the complaints process exists to serve, and would be indefensible the first time anyone examined it.

Triage algorithmically.

More defensible, and it has a place - but the regulatory bar is already published, and it is about evidencing that your automated triage actively identifies vulnerable people rather than filtering them out.

Answer at machine speed.

This is the one that costs most. Part of what people respond to in a good complaint response is the evidence that a human being spent time on them. Automating the reply spends the relationship, and the credibility of every future response you send. It is the one asset here that does not get cheaper as the models improve, and you cannot buy it back.

Now the part almost nobody has noticed. Your process assumes a person who reads your guidance, finds your form and describes what happened in their own words. That person is disappearing. They open a chatbot first and arrive at your front door with the output already written.

The process no longer begins where you think it does. It begins several minutes earlier, somewhere you have no presence at all - and every characteristic that makes the resulting complaint expensive was decided in that gap.

You cannot reach into that gap. You can stop pretending it isn’t there. Thankfully, there are three things you can do to mitigate this, none of them a product, all available to be implemented right now at minimum cost:

Publish guidance on complaining well.

The ombudsman schemes worked this out first. The Parliamentary and Health Service Ombudsman publishes suggested prompts for people using AI to complain, including one line that does more work than any filter ever built: don’t ask AI to add legal arguments or reference laws to make your complaint sound more official. That sentence removes the most expensive characteristic of the AI-written complaint at source. It reads as helpfulness because it is helpfulness - an analysis of 1.1 million US consumer complaints published in Nature Human Behaviour this year found that AI assistance improved the chance of a favourable outcome purely by improving presentation, without changing a single fact. Sit with that, because it isn’t really a finding about AI. Presentation was already deciding cases - the machine just handed the advantage to the people it had been costing.

Ask one clarifying question at intake.

Thank them for the detail, ask what the central issue is and what outcome they want, commit to a date. It makes no accusation, requires no detection, and treats the complainant as capable. The person with one real concern will tell you what it is, and the enormous submission collapses into a paragraph you can usually answer.

Use the proportionality you already have.

There is a widespread and expensive belief that a long complaint requires a long answer. It does not. The obligation is to handle complaints promptly and fairly, not to engage exhaustively with every assertion in an incoherent document. What that needs is structural rather than heroic: a policy permitting summarisation, templates for setting scope, and neutral documentation of your reasoning.

That is the whole intervention, and here is why it is urgent. Among HR directors who saw an AI-assisted grievance reach tribunal, 86% reported that none of those cases succeeded. Read carelessly, reassuring. Read properly, alarming - the organisation absorbed the full cost of complaints that overwhelmingly failed. Merit and cost have come uncoupled. You now pay the same either way, which means the only variable left is how much of that cost you design out at the front.

All three moves are the same move, and it is worth naming. They put the person back into their own complaint.

Because a complaint has always been a remarkable act. Someone with no power and no leverage writes down what happened to them, and an institution is obliged to read it and answer. What is drifting now is the person out of it - first the words become the machine’s, then the scope, and on the agentic path eventually the intention too, with the agent filing before anyone has noticed anything is wrong.

Every intervention above interrupts that drift. The prompt guidance asks people to say what happened to them, in their own words. The clarifying question asks what actually matters out of everything the machine produced. Both are the same instinct: get the human back in the room.

Which leaves the only real question. Is the work of hearing the person inside the paperwork a cost to be engineered away, or the whole point of answering at all?