At its Gemini at Work event on October 8, Google put dozens of named customer results(opens in a new tab) on stage. Commerzbank says manual document review fell from 20 hours to one. Bradesco says review time dropped from an hour to five minutes, with more than 10% in financial efficiencies. Wesfarmers says an internal agent at Bunnings saved half a million staff hours, and NTT DOCOMO says 450,000 hours a year came back to its people.
These are named companies, and they are good evidence that agents work on document-heavy, rules-bound tasks. They are also, almost all of them, time and speed measures picked by a vendor for its keynote. Only Bradesco attaches a financial figure. Hours saved become money only when someone cuts a contract, skips a hire, or does more paid work with the same team. None of the hours figures says which of those happened.
A CIO article from October 1 shows the gap. West Monroe's CIO reports a 40% cut in yearly managed service provider costs(opens in a new tab) from an IT support agent. LaunchDarkly's CIO reports about $50,000 in annualized savings on tier 1 support, plus about $1 million in avoided cost, which is not the same as cash saved. The same article walks through an illustrative tier 1 desk: an agent closes 40% of 10,000 monthly tickets, worth €60,000 on paper. Reopened tickets send €9,000 of work back to people, checking the agent's work costs about €15,000, and the net lands near €36,000, roughly 60% of what the pilot promised.
That desk is a model, not a measured result. But every cost it names is real and usually missing from a pilot business case: rework, human review, and the hard cases that only show up in production. Google's own launch points the same way. It added spend caps that pause a project's agent when they trigger.
The practical move for agent programs heading into 2027 budgets is three lines per pilot. First the hours claim. Then the dollar line it is meant to hit: a vendor contract, an open requisition, overtime. Then reopen and review costs, measured in the pilot rather than estimated. Approve scale on the net figure. If the pilot never tracked reopens or review time, track them for a month before the budget meeting. A smaller number you can defend is worth more to a CFO than a large one you cannot.
Action items
Google's customers reported large time savings from AI agents this week. Very few put a dollar figure on them.
For the CFO
Ask which contract, requisition or overtime line each hours claim is meant to reduce, and book only that.
For the CAIO
Measure reopen rates and human review time in every agent pilot, and report savings net of both.
For the board
Treat hours saved as a leading indicator. The result is the cost line that moved.
Researched and drafted by an automated workflow, then reviewed and edited by a human editor before publication. Every source is linked. See how we use AI here.
Also worth knowing
- Google Cloud BlogGemini at Work 2026: Introducing Gemini agent(opens in a new tab)
Dozens of named customer results, mostly hours and speed rather than dollars. Google also added spend caps that pause an agent when hit. Ask every agent vendor for the same control.
- CIOWhere AI agents are showing real IT savings(opens in a new tab)
West Monroe reports 40% lower managed service costs from an IT agent. An illustrative model shows reopens and review cutting a tier 1 agent's savings to about 60% of the pilot figure.
- Microsoft Command LineIntroducing Microsoft-Decision-1, our model for fast decision-making(opens in a new tab)
A 9B scoring model for routing and classification at $0.042 per million input tokens, output free, on Foundry. Speed and accuracy claims are Microsoft's own, so test it on your data.
- Cloudflare BlogIntroducing Clef-omni with full multimodality, plus a faster Clef and a cheaper Clef-flash(opens in a new tab)
Clef-flash input fell from $0.09 to $0.038 per million tokens. Four vendors now list a model at $0.10 or less per million input tokens, so rebid any routing workload still on a general model.
- Infosecurity MagazineASOS Confirms Data Breach Linked to Stolen Employee Credentials(opens in a new tab)
One socially engineered employee login reached third-party customer platforms, and the attacker told the BBC an AI marketing tool was the way in. List which AI tools hold customer profiles.