ICYMI
• The Better AI Gets, The Harder It Is To Use – Five compounding challenges mean capability and deployability are moving in opposite directions.
TL;DR
• AI has cut the cost of writing a complaint or request to a company to about a penny a letter. Answering still costs pounds and is sometimes a legal duty.
• Inbound correspondence volumes are increasing. Graduate vacancies draw on average 140 applications each, and UK ombudsman complaints rose 54 per cent in 2024/25.
• AI makes low-merit and high-merit submissions look more alike. Polish is becoming a weaker signal of substance.
• The response is to upgrade your correspondence engine: re-price the perimeter where lawful, structure the intake around facts rather than prose, write for machine readability and build legitimate finality into every channel.
Before 1840’s postal reforms, the cost of mailing a letter in Britain was based on distance travelled and the number of sheets, with the bill paid by the recipient at the door. For example, a three-sheet letter from London to Edinburgh cost 39½ pence, and the receiver could simply refuse to pay. Following the reforms, a half-ounce letter could be sent anywhere in the United Kingdom for one penny, prepaid by the sender. Annual volume rose from 76 million letters in 1839 to 168 million in 1840, and nearly 350 million by 1850. The Post Office spent the following decades industrialising to cope, until London was taking deliveries up to twelve times a day.
Britain’s postal system behaved as any economist might expect. We are now seeing the same pattern repeated in every inbound channel at once, as AI radically cuts the cost of inbound. Sending a polished complaint, grievance or information request now costs pennies. But answering still costs pounds.
Four linked and compounding shifts are increasing the burden on receivers:
1. Falling cost to sender. The marginal cost of each inbound item is collapsing. Five years ago, a solicitor would charge £150 to £500 to draft a formal letter, and a claims firm would take up to 40 per cent of any award for its help. Today Garfield.Law, the first AI-only law firm authorised by the Solicitors Regulation Authority, drafts a compliant letter before action for £7.50. It costs about a penny to generate a 1,000-word letter at today’s frontier AI model prices, and far less on budget models. Costs will likely continue to fall as low-cost AI models become more capable.
2. Increasing inbound volumes. When the cost of an additional application, complaint or grievance approaches zero, people send more of them, to more organisations, and retry more often. The Institute of Student Employers counted 140 applications per graduate vacancy in 2024, up 59 per cent in a year. Volumes will likely continue to grow as consumers begin to deploy AI agents alongside chatbots.
3. Increasing difficulty identifying legitimate claims. AI raises the presentational standard while lowering the effort required to reach it. Grievances arrive lawyer-grade, citing the staff handbook accurately. That does not mean every AI-assisted submission is weak. It means a polished document is no longer evidence that the sender invested effort or that the claim has substance.
4. Cost to resolve stays high. Generating a reply may get cheaper, but taking and evidencing a decision is harder to automate. An individual can subscribe to an LLM in a few minutes, but an organisation needs to upgrade its entire machine for correspondence, paying the Deployment Ratio along the way. The Financial Ombudsman’s cost per resolved case rose from about £920 in 2019/20 to £1,116 in 2023/24. The receiver must still verify the facts, apply policy and stand behind the decision.
The four shifts compound most sharply where there is a legal requirement to respond. Much inbound can be triaged, deflected or ignored, but some cannot. The ACAS Code requires an employer to engage with the substance of every grievance however it was drafted. Behind the grievance sits the tribunal, where open single claims rose 42 per cent in the year to March 2026 and some cases are already listed for 2029. Monitored central government bodies received a record 94,526 FOI requests in 2025, and the ICO says an AI-drafted request citing invented decision notices must still be answered.
For illustration, let’s take a bank fielding 100,000 complaints a year, mid-table among the FCA’s reporting banks. Estimating an hour of handling per complaint, that is a 60-person operation and roughly £3.5 million a year in staff costs. About one case in twelve goes on to the ombudsman at £680 each, adding £5.7 million in fees. Call it £9 million a year, before any redress is paid. Now double the inbound, which is the direction of travel, and the bank must either hire another 60 handlers or find a new way to handle complaints.
Building a Better Machine
Faced with this set of trends, firms need to rebuild their machinery for responding.
• Increase the price to the sender, where the law allows. Ask for evidence up front, such as transaction IDs, dates and a signed declaration that the facts are true. Require professional representatives to declare when they use AI. On contested claims, pick up the phone, as human interaction increases the sender’s costs. Apply charges where possible. When the ombudsman charged representatives £250 a case, referrals fell from 37,100 a quarter to 4,300.
• Structure the incoming information. Stop reading prose. Whatever arrives, have AI distil it into structured data that your systems and people can process easily. Senders write what they like. Structure helps you to attend to the facts (e.g., harm, legal risk, new evidence) and ignore distractions (e.g., length, tone, who is copied in). Structure also makes it easier to spot patterns such as a coordinated campaign. Give legitimate AI agents a front door of their own, rather than letting them flood the free-text queues.
• Write for the machine reader. The decision to escalate or to give up is increasingly made by the sender’s AI. A well-evidenced answer is more likely to result in a dropped case, while a thin or templated one invites the next round. Track escalation rates by reply type and experiment to see what works best.
• Audit the promises, rewrite the rulebook. Part of how and when you respond to inbound is set by law. But the rest is often based on legacy standard operating procedures that pre-date AI. Update the rulebook with new definitions of what counts as a complaint, what counts as new evidence, how many reviews are needed and when a decision is final.
• Decide well, then stop. AI doesn’t just reduce the cost of the original request. It also reduces the cost of subsequent challenge. New workflows should be designed to survive challenge. Capture clear reasons, evidence and an audit trail of decision-making. Then hold the line. Fairness is not unlimited responsiveness.
• Mutualise the machine. Every firm in your sector faces the same flood, yet each builds its own defences. The financial services industry proved out this model with Cifas, a shared platform for fraud intelligence. Share intake standards, pool precedents, respond to fee consultations together.
Before 1840, the receiver could refuse a letter at the door, unpaid and unread. You don’t have that option. On the channels that matter most, you are obliged to accept, to read and to answer, at your own cost, however cheaply the thing was made.
Monday Morning Actions for Executives
• Gather trend data for inbound channels. Not only cost and volume, but also the ratio of genuine to spurious claims.
• Find the AI already arriving. Sample last month’s inbound for machine drafting and estimate the signal rate per channel.
• Assess your response. Measure how much work AI is already doing for you in inbound channels, and what more it could do.
Board Questions for the Back Pocket
• Have we modelled our exposure if inbound volumes explode?
• Are we tooling the answering side as fast as our counterparties are tooling the sending side?
• Which of our regulators and ombudsmen are themselves drowning in AI-generated inbound, and what does that do to the speed and quality of the decisions we depend on?
Sources & Notes
Several figures below are practitioner surveys or platform self-reports rather than audited data, and the bank example is the author’s own construction. Each is flagged where it appears and should be read as directional.
• The Penny Post. Before 1840, British postage was charged by distance and number of sheets and normally paid by the recipient on delivery, which meant an unwanted letter could be refused (The Postal Museum, “Rowland Hill’s postal reforms”). A three-sheet letter from London to Edinburgh cost 39½ pence, falling to 2 pence under the reform (W. Walker Hanlon and co-authors, “A Penny for Your Thoughts”, NBER Working Paper 30076). The Uniform Penny Post took effect on 10 January 1840; Post Office records report 112,000 letters posted on the first day, around three times the previous year’s count. Annual volume rose from 76 million chargeable letters in 1839 to 168 million in 1840 and nearly 350 million by 1850 (General Post Office figures via The Postal Museum). London eventually received up to twelve deliveries a day (London Museum).
• Before AI: the sender’s price. Published firm price lists and cost guides put a bespoke solicitor’s letter at £100 to £500 plus VAT before 2023, with £150 to £500 the defensible band for a formal letter of claim. Indicative market ranges, not an audited average. Claims management companies charged large shares of redress: the FCA noted some consumers paid over 40 per cent before its fee caps of 15 to 30 per cent by redress band took effect on 1 March 2022.
• Garfield.Law. The Solicitors Regulation Authority authorised Garfield.Law, the first purely AI-based law firm in England and Wales, announced May 2025. Services start at £2, with a debt-protocol-compliant letter before action at £7.50. SRA chief executive Paul Philip called the approval “a landmark moment for legal services in this country”. Garfield’s own comparison figure for a traditional solicitor’s letter (£150 to £400) is vendor-sourced, though it aligns with independent price lists.
• A penny to send. An author calculation, not a quoted price: a 1,000-word letter is roughly 1,300 output tokens, which at published API prices costs about $0.013 on a frontier model and under a tenth of a penny on the budget tier. Treat as an order of magnitude. The direction is settled: Stanford HAI’s 2025 AI Index records the cost of querying a GPT-3.5-class model falling from $20 to $0.07 per million tokens between November 2022 and October 2024, a more than 280-fold reduction.
• Graduate applications and screening. Institute of Student Employers, Student Recruitment Survey 2024: 140 applications per vacancy, up 59 per cent and the highest in over three decades, from 1.2 million applications for roughly 17,000 vacancies; the ratio held at 140 in the 2025 survey. LinkedIn told CNBC in October 2025 that applications were running at nearly 9,500 a minute; a platform self-report.
• Sticky resolution costs. FOS annual reports put its unit cost per resolved case at about £920 in 2019/20, £1,040 in 2020/21 and £1,116 in 2023/24. The standard case fee was £550 for six years to 2019/20, peaked at £750 between 2021/22 and 2023/24, was cut to £650, and is £680 for 2026/27 (FOS Plans and Budget). Ministry of Justice tribunal statistics show the mean time to clear a single employment claim rising from 19 weeks to 31 weeks in the year to late 2025. Rising unit costs have several causes, including inflation and case complexity; the trend, not the attribution, is the point.
• AI-drafted grievances. Paman Singh, principal associate at Weightmans, quoted in People Management on the “stratospheric rise” in AI-generated grievances taking hours to resolve. Irwin Mitchell’s survey of 200 HR professionals found 60 per cent had dealt with grievances they suspected were AI-generated, and 52 per cent said they were harder to resolve. A practitioner survey, self-reported.
• FOI. Cabinet Office statistics: monitored bodies received 94,526 FOI requests in 2025, the most since monitoring began in 2005. The ICO published guidance on AI-generated FOI requests in May 2026; Deborah Clark of the ICO cited higher volumes and greater complexity, and the guidance notes that AI-drafted requests may cite decision notices that do not exist yet remain valid requests. The section 12 cost ceiling is £450, or 18 hours, for central government.
• The grievance duty. The ACAS Code of Practice on Disciplinary and Grievance Procedures requires employers to engage with the substance of a grievance however it was produced; tribunals may adjust awards by up to 25 per cent for unreasonable non-compliance. See DAC Beachcroft, “Generative AI and grievances”.
• Tribunal backlog. Ministry of Justice Tribunal Statistics Quarterly and Law Society analysis: open single employment claims rose from around 45,000 in March 2025 to 64,000 in March 2026, an increase of 42 per cent. The Work Rights Centre’s May 2026 report records a claim lodged in January 2025 listed for hearing in 2029. The backlog has several causes; AI-assisted volume is one pressure among them.
• The bank example. Author arithmetic for planning rather than benchmarking. It assumes an hour of handling per complaint on average (simple cases close in days, complex ones run to the FCA’s eight-week limit), a loaded staff cost of £35 an hour (roughly £58,000 per full-time handler) and 1,650 productive hours a year, giving a 60-person operation for 100,000 complaints. Escalation of one case in twelve combines FCA aggregate complaints (around 3.6 million a year) with Financial Ombudsman volumes (305,726 in 2024/25); the datasets cover different periods and the 2024/25 representative surge inflates the rate. An illustration, not a benchmark.
• Ombudsman volumes. Financial Ombudsman Service annual data: 305,726 complaints in 2024/25, up 54 per cent on 198,798 the year before and the highest since the PPI peak. The standard respondent case fee is £680 from April 2026 (£650 in 2024/25 and 2025/26).
• Re-pricing precedent. From 1 April 2025 the FOS charges professional representatives £250 per referred case after ten free cases a year, reduced to £75 where the consumer wins. FOS quarterly data: in Q2 2025/26 representatives brought 4,300 cases against 37,100 in the same quarter a year earlier, and total new complaints fell by more than a third to 46,300. Part of the decline reflects the FCA’s pause on motor finance commission complaints rather than the fee; the FOS also reports withdrawn or abandoned representative cases falling from over a third in 2024/25 to 19 per cent, evidence the fee changed referral behaviour. From 1 April 2026 the representative fee is £260, reduced to £80 where the consumer wins.
• Mutualisation. Cifas, the UK’s not-for-profit fraud-prevention service, runs cross-sector data sharing among hundreds of member organisations; the working template for pooled defence against high-volume adversarial inbound. Sector-level answering infrastructure is the author’s extrapolation, not an observed practice.