transformationplaybook.ai
AI Wire · 24 AUG 2026

The Real Cost of an AI Agent Is the People Watching It, McKinsey Finds

A hand resting on a dark keyboard in low light.

Photo: Nubelson Fernandes on Unsplash

22 - 29 Aug 2026
70-75%

of the variable cost of running a customer-service AI agent in banking is human oversight, not token costs, McKinsey found.

Token costs are just 20-25% of the bill. The lever on agent cost is process design, not model procurement. Gartner shows service leaders funding AI growth by cutting the labour their agents will need back as reviewers.

Where a banking service agent's money actually goes (McKinsey)

Token / model costs22%
Human oversight73%

Funding the AI budget out of a flat function (Gartner, service leaders)

Overall function budget growth2%
AI spending growth38%

The week in three lines

  1. McKinsey published four AI pieces in seven days. The question has moved from whether AI works to what a unit of AI work costs, and what finishing the job takes.
  2. Two firms landed on numbers that don't sit comfortably together. McKinsey finds oversight, not tokens, is 70-75% of running cost; Gartner finds service leaders funding a 38% AI spend rise by cutting labour and overhead out of a budget growing just 2%.
  3. McKinsey introduces 'completed work ROI': the fully loaded cost to finish a unit of work, such as a resolved claim or a closed sale, across humans, agents and systems combined. Few organisations can answer what one finished unit of work costs them today. That's the gap the metric is built to expose.

The papers

McKinsey QuantumBlack24 Aug 2026Agent cost

The real cost of running an AI agent is the people watching it

McKinsey's practitioner guide to agent economics finds token costs are typically just 20 to 25% of an AI agent's variable running costs, while human oversight accounts for 70 to 75%. A single-agent banking service workflow costs $20,000 to $30,000 to run; a conversational onboarding agent for 2,500 customers a year costs $10,000 to $15,000.

Why it matters

Process design, not model procurement, is the lever on agent cost. What matters is what proportion of runs get reviewed and by whom: McKinsey's own onboarding example puts 10 to 20% of runs in front of a risk or functional expert.

Read the source
The playbookFree toolkit
McKinsey Industrials26 Aug 2026Agent cost

AI's return comes from decisions, not headcount

McKinsey argues that since SG&A typically runs 5 to 12% of costs, labour savings alone can't explain a 20% EBITDA improvement from AI, so the value must come from faster decisions and better use of existing assets. Companies that redesign end-to-end workflows around AI report roughly 20% EBITDA uplift and three dollars of profit for every dollar invested.

Why it matters

This counters the idea that AI value is purely a headcount question: it points investment toward production scheduling, demand sensing and capital allocation instead. McKinsey's estimate that delaying a $50 million initiative by a year can turn it into a $75-100 million problem is a useful way to price delay.

Read the source
The playbookFree toolkit
McKinsey Quarterly28 Aug 2026Playbook

Twenty companies that made real money from AI, and what they had in common

McKinsey studied 20 large companies that created significant value from AI: EBITDA improved 20% on average, with $3 of incremental EBITDA returned per $1 invested. Two-thirds focused on three business domains or fewer; Freeport-McMoRan found 60% of its AI system reusable across plants, and DBS cut model deployment time from 15-18 months to 2-3.

Why it matters

Focus beats breadth: two-thirds of the winners worked on three domains or fewer, the opposite of how most AI use-case portfolios are built. Freeport's 60/40 reuse split is a practical planning number for sizing a central AI team against local ones.

Read the source
The playbookFree toolkit
Gartner26 Aug 2026Agent cost

Service leaders are raising AI spend 38% inside a budget growing 2%

Gartner surveyed 199 service and support leaders and found AI spending up 38% while the function's overall budget grew just 2%: the money is being redirected, not added. Leaders are moving spend away from labour and overhead and toward technology.

Why it matters

This is what an AI business case looks like with no new money behind it. Any organisation self-funding AI from a flat budget should weigh this against the finding that human oversight is most of an agent's running cost: it may be cutting the labour its agents will need back as reviewers.

Read the source
The playbookFree toolkit
Gartner26 Aug 2026Supply chain

The real test for an inventory technology is how much counting it removes

Gartner argues that barcodes, RFID and smart cabinets improve individual counting tasks while leaving the underlying operating model unchanged, whereas cameras combined with demand prediction and automatic replenishment remove the counting task altogether. Written for healthcare supply chains, but the logic holds across sectors.

Why it matters

This is a direct test for retail stock accuracy too: does a proposed technology make counting faster, or does it remove the need to count at all? Most proposals on the market do the first while being sold as the second.

Read the source
The playbookFree toolkit
MIT Sloan Management Review26 Aug 2026Playbook

The AI platform isn't finished, so invest in what outlasts it

Kevin Boudreau argues that generative AI adoption is already large, around 2.4 billion monthly users, but the technical and institutional architecture around it remains unsettled, unlike electricity, his benchmark for a mature general-purpose technology. His argument is academic rather than based on new data.

Why it matters

A steadying frame for any board facing AI-FOMO pitches: own the assets, such as proprietary data and customer relationships, that become more valuable as AI becomes abundant, rather than trying to own the AI itself. It's a fair test for any capital request that assumes today's model pricing and capability will hold for the life of the business case.

Read the source
The playbookFree toolkit

Also published

What nobody is saying

McKinsey says human oversight is 70 to 75% of the variable cost of a banking service agent; Gartner says service leaders are funding a 38% AI spend rise by cutting labour and overhead. If both are right, plenty of functions are cutting the same capacity their agents will need back as reviewers, and booking the saving twice.

McKinsey's own research argues both that AI could reach cost parity with human labour in around two years, against roughly 55 years for steam, if companies can repeat today's best results at scale, and that few have managed it yet. The firm treats that gap as an implementation problem and points to the operating model, which may well be right. It's also, conveniently, the argument that supports what consultancies sell.