Clients Want a Share of Your AI Savings. How to Price Now

The question has started arriving in renewal meetings: "You are using AI now. Where is our share of the saving?" It has reached the Big Four. KPMG International told its own auditor to pass on AI savings and the fee fell 14%. It has reached Indian IT services, where one chief executive says clients want the same work for 25 to 30% less. It is reaching in-house legal teams, six in ten of whom plan to push on how their law firms price. Agencies are not exempt; WPP's chief executive has said the group will move away from time and materials.

This post is the evidence, with sources and dates, followed by the arithmetic that makes the ask rational and the four pricing responses that exist. The short version: if you sell hours and AI halves the hours, the client is right that your price should fall, and the only durable answer is to stop selling hours. The discount they are asking for now is temporary. The productivity baseline it reveals is permanent, because every competitor who catches up resets it.

Every row below is a public statement or filing, with its date and an evidence label. Two widely repeated figures are not in the table because we could not find a primary for them: a claim that a large holding company links about a tenth of remuneration to performance, and a claim that a specific outsourcing deal splits AI savings between vendor and client. Both appear only in secondary reports.

The KPMG case is the cleanest because the number is in a public filing. The Financial Times report also carries the auditor's side of the argument in KPMG's own statement: AI can create efficiencies, but "developing and operating AI systems can generate additional costs". That sentence is the whole negotiation in miniature. The buyer sees the hours fall; the seller sees the tooling bill rise; the contract only prices one of them.

The numbers below are illustrative, not a client's. Take a deliverable that took a senior person 100 hours at $150 an hour. The agency builds a pipeline and the same deliverable now takes 50 hours of that person's time plus tool spend. Under time and materials the invoice halves. Under a deliverable price it does not, and the difference is the return on the pipeline.

Tool spend is not trivial at the top end. OpenAI wrote on September 22 that, valued at API prices, daily token usage inside the company "has exceeded $600 for the median researcher and $7,000 for researchers at the 90th percentile", in its GPT-6 Sol and Luna announcement. That is a frontier lab's researchers, not a marketing team's drafting tools, but it shows the direction: as agents take longer tasks, the compute line starts to look like a salary line, and a price that only counts hours cannot see it. Our post on two hires finishing the same task in half the time works the same arithmetic from the employee's side.

WPP's chief executive said at the group's February 2026 strategy update, as reported by Digiday, that outcome-linked fees are "the beginning of a more widespread commercial model evolution". Our June guide to agency pricing models lays out the full menu. This section is narrower: the four structures that answer the discount request directly, and what each one costs the seller. None is free. Every one moves a risk that hourly billing left with the client.

The objection to output pricing is old and worth hearing from someone who chose not to use it. IDHL, a British independent agency, considered it and has not used it to date.

The answer to that objection is not to abandon output pricing but to price the meaningful thing at a fixed fee and the measurable thing at a share, which is what WPP describes running side by side. Our guide to token-based versus outcome-based agent pricing covers the same trade-off for software.

Three executives describe the same sequence from different industries. TCS's chief executive told Reuters the company has so far offset AI-related downward pressure on revenue with new work, and that how far it stays ahead of that deflation will determine growth. WPP's chief executive, per trade-press reports of the first-half call, expects a short-term deflationary effect on pricing and says clients reinvest the saving rather than cut budgets. Omnicom's chief executive says clients are reinvesting savings "immediately into the marketplace".

Put together, the pattern is a lag, not a permanent discount. In the first phase, the agencies that automate first can either keep the margin or win work by passing it on. In the second, enough competitors have caught up that the lower cost becomes the market price, and the client's budget is spent on more output at that price rather than on less spend. The lag varies by industry: the audit and IT-services examples in the ledger are already in the second phase; much of marketing services is still in the first.

The implication for pricing is uncomfortable but clear. A discount negotiated against your current hours is a one-time transfer. A price set on the deliverable at the new cost level is the number you will be competing on anyway within a year or two. The agencies still quoting the old hours when that happens lose recurring work to the ones that repriced early. The ANA's compensation benchmark already showed the direction before generative AI: by its 2022 edition 82% of advertisers used fee-based compensation, and among advertisers spending $500 million or more, 53% used fixed or output-based fees, up from 5% in 2016.

The request usually arrives as one sentence in a renewal. The router below is how we answer it, depending on how the account is currently priced. The aim in every case is the same: agree that the efficiency is real, refuse to convert it into a smaller hourly invoice, and move the contract to a structure where both sides gain from the next improvement too.

Two things make that conversation easier. First, know your own numbers: the hours per deliverable before and after, and the tool spend, so the saving you offer to share is real. Second, have the new structure written before the meeting. Our AI transformation practice builds the delivery pipeline and the pricing model together, because one without the other is either a discount or a promise.