Salesforce’s published guides price one customer request at a single action in one example and five in another [1]. That 5x spread means US$0.10 or US$0.50 for a single request, on the same rate card [1]. That spread is your blind spot. A bill tracked by agent steps cannot match a budget written as a flat share of spend.

The big picture:

CloudZero reports artificial intelligence (AI) hit 3.02% of the median cloud bill in September 2026 [2]. That 37-basis-point jump is the 14th straight monthly rise and the largest one yet [2].

At the same time, 252 finance leaders were asked for their estimate. The median guess put AI at 21% to 35% of infrastructure spend [2]. One figure is a hard invoice. The other is a rough guess.

I will take CloudZero at its word on the limits here. The firm notes its survey takers count AI packed into software tools, making 3.02% a strict floor [2].

What I keep circling is that neither metric tracks the real unit driving the cost: actions per request. Salesforce offers the clearest view of this new meter.

An action is a single step an agent takes, like updating a record or answering a product question [1]. A text action costs 20 credits, or US$0.10, while a voice action costs 30 credits, or US$0.15 [1]. A conversation costs US$2, and a Help Agent resolution also costs US$2 [1].

The vendor’s five published guides price a single request at 2, 3, 5, 1, and 4 actions [1]. That means the exact same task costs five times more at the top of the range.

Salesforce calls this credit a new unit of payment that ties cost to business value [1]. My read is that a price per step only ties cost to the number of steps. That step count lives inside the agent’s code, not the final business result.

Gartner notes that AI costs are unfixed and set at run time, unlike older tech tools [4]. That run time is the agent’s loop, where the final cost adds up long after you sign the deal.

By the numbers

  • 3.02% — Median bill share: September 2026 saw AI hit this mark across 422 firms, up 37 basis points from August [2].
  • 21% to 35% — Finance leader guess: The median estimate from 252 finance leaders, with 54% placing AI above 20% of that spend [2].
  • US$0.10 to US$0.50 — One request spread: Text actions on the Agentforce list price range from a one-action guide to a five-action guide [1].
  • 12.55% — Top quartile spend: AI’s share of cloud spend at the 75th mark, against 0.27% at the 25th [2].

What I’d watch:

I am watching three specific metrics here, in the exact order a finance team meets them.

  • Actions per resolution: Nobody prints this rate card. The first firm to report it holds the only metric that turns agent steps into business results.
  • Step cap per request: A hard limit on actions, written into the contract. This cap turns a technical design choice into a fixed budget line a planner can track.
  • Cost per resolved request: The final bill divided by finished work. A share of tech spend cannot be checked against an action meter. This unit-cost metric can.

Right now, 44% of chief executives struggle to show AI-driven financial wins [3]. Gartner advises them to measure results and unit costs instead of raw AI activity [3]. Unit costs need a bottom number. Actions per request is the one your bill already tracks.

The catch

I could be wrong about how typical those Salesforce guides are. They are marketing pictures the vendor chose, not an audit of live server traffic. A five-action request might sit closer to the norm than the extreme edge.

The huge gap between 21% and 3.02% is also partly a word problem. AI packed inside a software tool is real spend that never shows up on a cloud bill. Finance leaders may be describing the whole company, while the pulse measures only what it can see.

What survives both limits is the core gap. A budget built in shares, a bill metered in actions, and a board deck reporting results are three entirely different units.

The middle unit is climbing fast, with AI’s share of the median bill posting its largest monthly gain on record [2]. Fast growth in a unit nobody budgets is how a simple tracking error becomes a major risk.

At a glance

  • The Big Shift: AI vendors now bill by the action, and published guides show a single customer request can take anywhere from one to five actions to fix. The budget is still a flat share of tech spend, leaving a massive gap between what you plan and what you pay.
  • Why It Matters: AI reached 3.02% of the median cloud bill in September, marking the largest monthly jump on record. Meanwhile, finance leaders guess AI takes up 21% to 35% of infrastructure spend, and that gap between units makes financial returns nearly impossible to prove.
  • What I’d watch: The specific numbers that tie the billing meter directly to the finished work.
  • Actions per resolution: The exact count of agent steps behind one closed ticket.
  • Step cap per request: A contract limit on total actions that turns an open-ended loop into a clear line item.
  • Cost per resolved request: The total bill divided by finished work, providing the baseline for unit costs.
  • The Catch: The vendor guides are curated pictures rather than live traffic audits, and the gap between the bill and the guess exists because bundled AI is invoiced as something else.

Related reading

Sources

[1] Salesforce, “Agentforce Pricing” and the Agentforce rate card (list prices observed 2026-10-22) — https://www.salesforce.com/agentforce/pricing/ [2] CloudZero, “Your AI Economics Pulse for October 2026” (2026-10-13) — https://www.cloudzero.com/blog/ai-economics-pulse-october-2026 [3] Gartner, “Gartner Says 44% of CEOs Struggle to Demonstrate AI Financial Returns as Technology Becomes Central to Growth” (2026-10-20) — https://www.gartner.com/en/newsroom/press-releases/2026-10-20-gartner-says-44-percent-of-ceos-struggle-to-demonstrate-ai-financial-returns-as-technology-becomes-central-to-growth [4] Gartner, “Gartner Survey Finds Only 13% of CIOs Say Their Enterprise Has Realized Significant or Transformational Value from AI” (2026-10-19) — https://www.gartner.com/en/newsroom/press-releases/2026-10-19-gartner-survey-finds-only-13-percent-of-cios-say-their-enterprise-has-realized-significant-or-transformational-value-from-ai