Each employment tribunal in Britain used to receive about twenty applications for interim relief a year. It now receives about twenty a month. Nobody hired anyone. Nobody changed the law. The only thing that changed is that writing one stopped being work.
Single claims are up 39 percent year on year. Disposals are down 12 percent. The backlog of single claims has grown 55 percent, and cases are now being listed into 2030. The judges' own explanation, in the guidance the Presidents of the Employment Tribunals issued on 23 June, is that claimants are using AI to help them file.
Read that as a British legal story and it is somebody else's problem. Read it as an operations story and it is a preview, because the mechanism has nothing to do with tribunals. A system built on the assumption that submitting something costs the sender real effort behaves very differently once it does not.
The number that matters is the ratio, not the volume
Every intake process you run has an implicit exchange rate. A customer complaint costs the customer twenty minutes to write and costs you ten minutes to resolve. A trade-in appraisal request costs a shopper five minutes and costs your used car manager fifteen. A chargeback costs the cardholder a phone call and costs your office an afternoon of gathering documents. Those ratios were never designed. They emerged, and every staffing decision you have made sits on top of them.
The ratio held because both sides were made of people. Drafting a coherent, policy-citing, three-page objection was genuinely hard, so most people who were mildly annoyed did not bother. That filter was doing enormous quiet work, and it was never written down anywhere as a control.
It is now priced. Objector.ai will produce a policy-cited planning objection for a full application for 45 pounds, with a 249 pound tier for residents who want to pool money against a larger scheme. The output is not junk. It cites policy, it engages with the officer's report, and it takes a planning officer just as long to read and answer as one written by a solicitor. Determination times in the UK planning system now average beyond 40 weeks against a 13 week statutory target, and permission grants have fallen to the lowest level since records began more than two decades ago.
Geoff Keal, who runs the national planning portal that handles roughly 95 percent of UK planning applications, put it plainly: people are "using AI to be able to provide better objection documents, much wider and much broader, which is slowing the system down." Note the word better. This is not a spam problem. Spam you can filter.
It is not a British problem and it is not a government problem
A paper published on 17 August by Chris Schmitz, Lewis Hammond and Alan Chan gives the pattern a name, agentic flooding, and a dataset: 84 potential cases across 11 jurisdictions. Their finding about where it lands first is the part worth copying down. Near-term risk is highest for services that are financially attractive but complex. Something worth money at the end, and enough procedural surface that most people previously gave up before reaching it.
Read that description again and count how many of your own processes it fits.
Warranty claims. Rebate and incentive submissions. Co-op advertising compliance. Chargebacks and representments. Trade-in valuation disputes. Post-sale we-owes. Every one of them is worth money to the person filing, and every one of them has historically been rationed by the fact that pursuing it was tedious. Not one of them is rationed by a rule you wrote.
What actually breaks first
Not the queue. The queue is the visible part and it is the last part to fail. Three things go first, in this order.
Your response time promise. Whatever you tell customers about how fast you answer was calibrated against historic volume. It is the first thing to break, and it breaks silently, because nobody is measuring the promise, they are measuring the average.
Your triage quality. When the volume of well-written submissions doubles, the staff answering them stop reading closely and start pattern matching. That is when a genuinely serious complaint gets the template reply, which is the failure that actually costs you money and occasionally costs you a lawsuit. The tribunals hit this exact wall: the judges' concern was not that the AI-assisted claims were bad, it was that the good ones and the hopeless ones now look identical on the page.
Your ability to tell the difference between one angry person and forty. If your intake is by email and by form and by phone and by review site, and none of them reconcile to a customer record, you cannot currently answer the question "is this forty complaints or one complaint filed forty ways." You will need to answer that question, and you will need to answer it under time pressure.
The instrumentation, and it is boring on purpose
There is no product to buy here. There are four measurements, and if you have never taken them, you do not know what your exposure is.
One. Count submissions per channel per week, and keep the series. Not resolution time, not satisfaction. Raw arrival counts, by channel, week over week. Flooding shows up as a step change in arrivals long before it shows up as a service problem, and you cannot see a step change without a baseline. Most operations have never plotted this.
Two. Write down the cost-to-answer for each intake, in minutes. Actually time it. Then put the cost-to-submit next to it. Any process where those two numbers are more than about five to one apart is where you will get hit, and the order in which you get hit is the order of that ratio.
Three. Make identity reconcile across channels. Not a CRM project. The narrow version: can you take a name and a phone number and see every touch that person has had with your store across every intake, in one place, in under a minute. If the answer is no, the "one person or forty" question is unanswerable and everything downstream of it is guesswork.
Four. Decide your ration before you need it, and write it down. The tribunals' answer was to publish guidance restating that interim relief requires a "pretty good chance of success," a standard deliberately higher than more-likely-than-not. They did not add staff. They raised the bar and said so in public, in advance. That is the move available to almost every process you run, and it only works if the standard exists in writing before the volume arrives, because a standard invented during a flood reads as a brush-off.
The part that will be tempting and is wrong
The obvious response is to answer AI with AI. Point a model at the inbound and let it draft the replies. For genuinely templated work that is fine and you should do it.
It fails on exactly the cases that matter. The submissions that hurt you are the ones where somebody has a real grievance and now has a competent instrument for expressing it. Automating the reply to those does not resolve them, it produces a fluent non-answer at machine speed, which escalates. You have then built a machine that manufactures escalation, and the sender's next submission will be better than your reply because they only have one to write and you have four hundred.
The asymmetry does not go away when both sides automate. It gets worse, because your side has a compliance obligation, a brand, and a legal exposure attached to every sentence, and their side does not.
Where this actually lands
The filter that was doing the work was friction, and friction is being removed from every direction at once, by tools that are cheap, good, and improving. There is no version of the next two years where submitting something to your business gets harder.
Which means the only variables you control are the two ends: how fast you can see arrivals change, and how clearly you have stated the standard something has to meet before it consumes a person's afternoon. Both are unglamorous. Both are cheap. Both have to exist before the week they are needed, because the tribunals are now booking hearings into 2030 and no amount of clarity purchased in 2029 helps them.
Twenty a year became twenty a month, in a system with no marketing department, no growth incentive, and no reason for anyone to file more except that it finally became easy. Your processes have all of those things pushing in the same direction.
Sources
- Freeths, "Employment Tribunal statistics: A growing backlog and the rise of interim relief applications", 26 June 2026. Source for single claims up 39 percent, disposals down 12 percent, backlog up 55 percent, 50,000 single claims filed, listings into 2030, and the quoted characterisation that each tribunal historically received about 20 interim relief applications a year and now receives that many monthly.
- The Law Society Gazette, "Tribunal tries to deter interim relief bids amid AI-related surge", 23 June 2026. Source for the guidance issued by Judge Barry Clarke, President of the Employment Tribunal in England and Wales, and his Scottish counterpart, and for the "pretty good chance of success" standard. Note that the Gazette renders the interim relief increase as a Britain-wide figure while Freeths renders it per tribunal. We have used the per-tribunal reading because it is the one that matches the Presidents' own wording, and flag the discrepancy rather than pick silently.
- Chris Schmitz, Lewis Hammond and Alan Chan, "Characterizing Agentic Flooding of Government Services", arXiv, 17 August 2026. Source for the term, for the dataset of 84 potential cases across 11 jurisdictions, and for the finding that near-term risk is highest for financially attractive but complex services.
- Business Matters, "AI-Powered Nimbyism", 20 May 2026. Source for Objector.ai's 45 pound and 249 pound pricing, for determination times beyond 40 weeks against a 13 week statutory target, for permission grants at their lowest since records began, and for Geoff Keal of TerraQuest. That article asserts a causal link between AI objections and planning delays without publishing supporting data, so we have used it for the pricing, the timings and the quotation, and not for causation.
- The Economist, issue of 8 to 14 August 2026, for first raising agentic flooding as a pattern rather than a set of unrelated complaints. Its rendering of the Schmitz dataset as 93 cases does not match the paper's own figure of 84, and we have used the paper.