A proxy for effort is an incomplete measure of cost.
Andy spends his working week in Excel. He knows macros and formulae, he is quick, and the work is consistent and good. Betty spends hers with an LLM, prompting it, checking what comes back, discarding as necessary. Her work is also good, but whether she delivers more than Andy isn't measured. Both of them book 40 hours.
The timesheet records those two weeks identically, and in cost-per-salaried-hour terms it is perfectly accurate for both Andy and Betty.
This record of effort is carrying three separate jobs:
- It's the basis of the invoice, it influences what the customer should pay
- It's the cost record, it tells you the engagement's staff cost and margin
- It's the feedback loop into the business for planning, operations and estimating
The same number used to provide a decent answer to all these questions and is behind the Monday timesheet chase to ensure everyone has booked exactly 40 hours.
I wrote last time1 about why communicating a price to customers based on effort may become less relevant (see also Josh Santiago's piece2 on why this won't be universal). The other two are internal, no customer sees them, but they bear thinking about.
Take the cost record. A timesheet measures presence, and that was tolerable while presence and cost were nearly the same thing, because an engagement's cost was overwhelmingly salaried people's time. But no-one works an exact 40-hour week and Betty's week costs more than Andy's. Licences, inference, compute, none of it on her timesheet, and likely not allocated to the specific engagements (although some3 thinking4 has been done in this area). By the time any of it reaches the management accounts it sits in a central software OpEx line.
Estimating has the same problem from another angle. "How many hours did it take" was only ever a proxy for "how hard was it". Andy's 40 hours are mostly production. Betty's 40 hours are mostly checking: reviewing, judging, deciding what to discard. Estimate the next engagement off typical hours and a blended rate and you can't tell which of the two you are estimating into. And the timesheet asks people to self-report, which is the weak point: in one controlled trial5, experienced developers using AI tools took 19% longer than those without, and finished believing they had been 20% faster.
The strongest objection is materiality. If Betty's tooling costs a couple of hundred pounds against an engagement of several thousand pounds, why build attribution machinery for a small percentage of the cost base? This is probably reflective of where most firms are today but the cost won't stay immaterial. Infosys publishes its costs by nature on the same basis every year: between the year to March 20216 and the year to March 20267, employee costs slipped from 55.3% of revenue to 53.2% while the software line went from 4.2% to 8.8%. Payroll grew about 1.7 times over those five years and the software line 3.7 times. It's a 300,000-person IT services business rather than a fifty-person consultancy, so I'd read it as direction rather than a benchmark. Nor is the machine side reliably getting cheaper per unit of work. Gartner8 expects generative AI's cost per resolution to pass $3 by 2030, above many offshore human agents, because agentic work burns five to thirty times the tokens even as token prices fall.
There's a shape to this that PS firms won't be used to. Betty's licence renews every month whether she uses it or not, and her token spend rises and falls with what she does. So the cost of her work is part fixed subscription, part metered usage, sitting in a central line and belonging to no particular engagement. SaaS businesses have been splitting costs like that across customers for years. PS is yet to catch up.
PS firms will always need to know the cost of completing an engagement, whether the customer sees it or not. What Andy and Betty show is that a self-reported timesheet is a weak instrument for this (arguably this was true even before AI arrived). As the balance of cost shifts from human to agentic, what the business needs is a cost line the timesheet can't show, and a unit of delivery that I haven't yet seen defined.
References
- https://lbconsulting.tech/posts/are-billable-hours-losing-meaning.html | Are billable hours losing meaning?, Lee Ball
- https://www.santiagocompany.com/insights/the-new-economics-of-professional-services | The New Economics of Professional Services, Josh Santiago
- https://www.finout.io/blog/tracking-ai-costs-per-customer-and-per-feature-in-2026 | Tracking AI Costs Per Customer and Per Feature in 2026, Finout Writing Team
- https://tanujgarg.com/blog/unit-economics-ai-cost-per-token-inference-customer | Unit Economics for AI: Calculating Cost Per Token, Per Inference, and Per Customer, Tanuj Garg
- https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ | Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, METR
- https://www.sec.gov/Archives/edgar/data/1067491/000106749121000030/exv99w06.htm | Independent auditor's report on audit of consolidated financial results 2021, Infosys
- https://www.sec.gov/Archives/edgar/data/1067491/000106749126000038/exv99w06.htm | Independent auditor's report on audit of consolidated financial results 2026, Infosys
- https://www.gartner.com/en/newsroom/press-releases/2026-01-26-gartner-predicts-genai-cost-per-resolution-for-customer-service-will-exceed-offshore-human-agent-costs-by-2030 | Gartner Predicts GenAI Cost Per Resolution for Customer Service Will Exceed Offshore Human Agent Costs by 2030, Gartner