Companies Don’t Transform, People Do: Rishad Tobaccowala on Work, Agencies, and Reinvention

Rishad Tobaccowala is out with a book and a thesis, so some skepticism is warranted on the specific numbers. The claim that "top uses of AI are therapy, relationships, and medicine" is not what usage surveys show. Pew and Stanford's tracking both put writing, coding, and search at the top. The McKinsey "more AI agents than employees within a year" line has no source and no definition of what counts as an agentic employee. Take the framing; discard the stats.

Two claims survive that skepticism because they rest on documented cost data.

The corporate capability gap is real, even if overstated. Epoch AI's data shows the price to run a given level of AI capability has collapsed from roughly $50,000 to $296 in two years. When a company deliberately runs a six-month-old, sandboxed version of a model for safety and compliance reasons, while employees at home run the current public version, the gap is real and directionally meaningful. "Exactly half as capable" is made up. The direction is not.

What Tobaccowala misses: open-weights models like Llama, Qwen, and Mistral give enterprises a third option. Run current-generation capability inside your own walls, without shipping data to a single vendor or freezing on a stale version. That solves the safety worry that made companies sandbox in the first place, and it undercuts the argument for locking to one AI vendor entirely. If the frontier is a commodity, a premium vendor lock-in buys procurement comfort, not capability.

The billing model pressure is the more solid claim. Tobaccowala cites FT reporting that banks are already refusing standard law-firm associate billing rates. Agencies, law firms, and consultancies all sell headcount-time as their unit of value. When the input cost of producing an hour of that work drops 30 to 40% per year, the buyer can see the margin and will push to reprice. That math is not sensitive to whether his therapy statistic is accurate.

If you run or sell into a professional services business, the exercise is simple: model your revenues at a 10% annual price cut and find out what breaks. That pressure is arriving regardless.

Our call: By the Q4 2026 earnings calls of Publicis, WPP, and Omnicom (reported February to March 2027), at least one of the three will publicly describe client contract terms shifting from fixed headcount-based fees toward output- or outcome-based pricing. Confidence is medium. The repricing has already started in adjacent professional services, cost deflation is documented, and Publicis in particular has built its outperformance story on being ahead of this shift, giving it a strong incentive to say so on a call. Revisit by 2027-03-31.

What's being decided here. Nothing, technically. This is a strategy conversation, not a product release or a benchmark. So treat it as a set of testable claims about the AI market, and check whether they hold.

How hard is any of this to undo? The claims cost nothing to evaluate. The decisions they point at (locking your whole company to one AI vendor, or building a business on billable hours) are expensive and slow to reverse. That asymmetry is the whole point.

What sets the deadline? Nothing hard. Tobaccowala throws out a one-year clock on McKinsey going mostly-agentic, but that's his number, unsourced.

The Skeptic. Tobaccowala is selling a book and consulting on exactly this thesis, and the two loudest numbers here are the two you can't check. "Top three uses of AI are therapy, relationships, and medicine" is flatly not what usage surveys show. Writing, coding, and search dominate by volume. Pew and Stanford's own tracking say so. Therapy is real and growing, but it's not the top of the pile, and he needs it to be true so "alien intelligence" sounds inevitable. The McKinsey "more AI agents than employees within a year" line has no source, no definition of what counts as an agentic employee, and a strong incentive to dramatize. Take the framing, discard the stats.

The Researcher. One claim here is checkable and roughly right: the corporate-vs-individual capability gap. Epoch AI's own data, in the reading list, shows the price to run a given level of AI capability has collapsed, in their framing from a $50,000 car to $296 in two years. That's the mechanism behind Tobaccowala's point. If frontier models improve fast and a company deliberately runs six-month-old, sandboxed (walled-off from the open internet for safety) versions, the employee on the latest public model really is working with a meaningfully better tool. "Exactly half as good" is made up. The direction is not. That gap is the one claim in this episode a buyer should act on.

The Open-Source Advocate. The part Tobaccowala gets exactly right, without saying it: the model layer is a commodity. His words are that most corporate AI strategies don't differentiate because the tech underneath is a commodity and the advantage is human judgment. That's the open-weights argument in a suit. If the frontier is a commodity, paying a premium to lock into one vendor buys you procurement comfort, not capability. The counter he misses: open models (Llama, Qwen, Mistral) let a company run current-generation capability inside its own walls without shipping data to a single provider or freezing on a stale version. He frames the choice as "one old corporate model vs. seven fresh consumer ones." There's a third door, and it's the one that actually solves the safety worry that made companies sandbox in the first place.

The Compute Pragmatist. Follow the money on the billing claim, because that's the one with teeth. Tobaccowala says clients are already demanding 10% annual fee cuts and banks are refusing standard law-firm associate rates, citing FT reporting. The reason this bites is cost, not fashion. If the input cost of producing an hour of work drops the way Epoch's numbers say, then anyone selling that hour at the old rate is selling a margin the buyer can now see. Agencies, law firms, consultancies: all three sell headcount-time as the unit. When the thing inside the unit gets 30 to 40% cheaper to make each year, the buyer prices to the new cost. The old bill doesn't survive that arithmetic.

Where the real disagreement is. The Skeptic says most of this is a book pitch with cherry-picked stats. The Compute Pragmatist and Researcher say two of the claims survive that skepticism because they rest on cost math that's independently documented, not on Tobaccowala's say-so. The real split: is the "corporate AI is half as capable" gap a durable disadvantage, or a temporary artifact that IT departments close once open-weights models make current-generation capability safe to run in-house? Tobaccowala treats the gap as structural. It might be a 2026 problem that solves itself by 2028.

What it hinges on. Two facts. One, does AI cost really keep falling fast enough that billable-hours pricing gets repriced across professional services. Epoch's data says yes, and the FT reporting on law firms says the repricing has started. Two, does the corporate capability gap persist, or do enterprises catch up by adopting open models and faster refresh cycles. The first is the more solid bet.

What to actually check. If you buy AI for a company: find out how old the model your staff is allowed to use actually is, and whether the sandbox is protecting real data risk or just protecting the vendor contract. If you sell time (agency, consultancy, legal, creative), model your business at a 10% annual price cut and see what breaks. That's the pressure that's arriving whether or not Tobaccowala's therapy stat is true.

Prediction: By the Q4 2026 earnings calls of Publicis, WPP, and Omnicom (reported February to March 2027), at least one of the three holding companies will publicly report client contract terms shifting away from fixed headcount-based fees toward output- or outcome-based pricing.

Confidence: Medium. The repricing has already started in adjacent professional services, and cost deflation is documented.

Why: Tobaccowala says clients are demanding roughly 10% annual fee cuts and cites FT reporting that banks are refusing standard law-firm associate billing rates, which means the pressure is already live in the same headcount-time business model agencies use. The mechanism is cost: Epoch AI's data shows the price to produce a given level of AI work is collapsing fast, so any buyer paying old rates for work that is now cheaper to make can see the margin and will push to reprice. Holding companies talk about pricing structure on earnings calls constantly, and Publicis in particular has built its whole outperformance story on being ahead of this shift, so it has a strong incentive to say so out loud. The opposite outcome, all three staying silent on pricing model change through an entire earnings cycle while the pressure is this public, is the less likely world.

Revisit by 2027-03-31: We're right if Publicis, WPP, or Omnicom describes a concrete move toward outcome- or output-based client pricing on a Q4 2026 earnings call or in accompanying materials. We're wrong if all three continue to describe client fees only in traditional headcount, retainer, or commission terms.