The briefing

November 18, 2026

A five-minute read for the people putting AI to work in support, sales, finance and operations: one story explained, two more in brief, what else happened, one number, one field note. Written by Wren, every source linked.

Microsoft now bills its Copilot agent features by use and published no rate for any of them. The one number on offer comes from a study Microsoft paid for: $96 a month for a heavy user, against a $30 seat. Today is about who sets the ceiling on an AI bill. Salesforce capped its sellers at $100 a month, and a CIO found that two of his three headline results from AI agents weakened under his own checks.

By the numbers

  • $96 a month The $30 seat plus $66 of usage, in a GigaOm study Microsoft funded; it rests on 37 prompts and some assumed prices.
  • $100 a month From a memo Business Insider printed; technical roles get $500 or $1,000, pooled by team.
  • 34% lower The one result of three that held when a CIO re-tested 25 agents; he gives no ticket counts or cost.
  • $16.32 a month Ramp's October figure for US firms that pay for AI; its top 10% spend $800.84.
The meter has no rate

Microsoft puts Office inside the Copilot seat and bills its agent features by use, with no rates given

Microsoft made its new Copilot app generally available on 17 November 2026, the first day of its Ignite conference. Word, Excel and PowerPoint inside the app are included in the Microsoft 365 Copilot licence, which lists at $30 per user per month, paid yearly. The features that carry out multi-step work for staff are billed by use, and the Ignite material gives no rate for any of them.

Why it matters: Your seat count no longer tells finance what Copilot will cost. Microsoft's list of "usage-based AI services" includes Copilot Cowork, where staff hand over multi-step work; Copilot Code, which builds dashboards and small apps from a typed description; and Autopilot, an agent that runs recurring work without being asked. The Today Feed, a list of priorities that refreshes all day, "requires Copilot credits", Microsoft's unit for billing usage.

The catch: The only figure comes from a study Microsoft says it "commissioned and funded" with the analyst firm GigaOm, published on 16 November. It models an advanced user, doing 140 tasks a month with 20 handed to agents, at $96 a month. That is $66 of usage on top of the $30 seat ($96 minus $30). The study rests on 37 prompts run against a fictional company and on list prices "or on an assumed price where list pricing isn't published". Microsoft tells readers to treat the findings as "directional".

What to do: Ask your Microsoft 365 admin to set a spending limit and alerts in Agent 365 Cost Management before anyone switches on Cowork, Code or Autopilot. It is a page in the admin centre, in preview, that shows what is driving spend. Ask your Microsoft account team in writing what one Copilot credit costs and how many credits each metered feature uses. Until the answer arrives, budget your heaviest users at $96 a month, not $30, and mark the figure as Microsoft's assumption.

Looking ahead: Autopilot is reaching customers in Frontier, Microsoft's preview programme, first. Microsoft says Copilot Code becomes generally available "in the coming weeks".

Our coverage →Microsoft: Ignite 2026 Book of News (read 17 Nov 2026) ↗Microsoft Copilot blog: Copilot at Microsoft Ignite, by Jared Spataro (17 Nov 2026) ↗Microsoft Copilot blog: What AI at work really costs, by Nicole Herskowitz (16 Nov 2026) ↗Microsoft: Copilot enterprise plans and pricing (read 18 Nov 2026) ↗

Capped like travel expenses

Salesforce caps AI spend at $100 a month for sellers and up to $1,000 for technical roles, memo says

Salesforce has set monthly AI spending limits by role in its sales, customer success and professional services organisations, according to an email from chief operating officer Miguel Milano that Business Insider printed on 17 November 2026. Non-technical sellers get $100 a month. Technical roles get $500 or $1,000. "Last quarter, some individuals spent thousands of dollars per month on tokens," the memo says. Tokens are the units of text that AI models bill by.

Why it matters: The memo is a working template. Leaders manage each team's budgets "similar to T&E", travel and expenses, "so one person's spike can be offset by others". Going up to 20% over, once, needs the second-line manager. Up to three times the limit needs the third-line manager and a return-on-investment case, and lasts 90 days. Above that, a division head approves with finance.

The catch: A Salesforce spokesperson told Business Insider the company has no companywide individual budgets and leaves AI usage to team leaders. The memo's evidence is one sentence: a pilot team was "well under budget by month two with no drop in output". It gives no spend figure and no measure of output.

What to do: Ask finance to pull last quarter's AI spend per person from each vendor's admin console. Set a limit per role above what your heaviest tenth spent. Before the cap goes live, write down who approves 20% over and who approves three times.

Our coverage →Business Insider: Salesforce sets AI budgets after employees spent 'thousands of dollars per month on tokens,' memo says (17 Nov 2026) ↗

One of three held

A CIO re-checked three results from his 25 AI agents; only the 34% fall in support escalations held

Naren Gangavarapu, a technology executive who says he spent 18 months putting 25 AI agents into production at a large travel and cruise group, wrote in CIO on 18 November 2026 that two of the three results his team used to justify more investment came out weaker when he checked how they were measured. A 34% fall in Tier 1 support escalations held. Onboarding that fell "from 23 days to 3" included a backlog that was cleared once. A 24% rise in booking conversion ran alongside a marketplace rebuild and new pricing.

Why it matters: The escalation figure survived for three reasons he names. The six-month period was fixed to include peak season. Contact volume was flat year over year. Finance, not the agent team, reviewed counts taken from the system of record.

The catch: He does not name the company in the column. "We withheld volumes," he writes, so the 34% has no ticket count behind it. He gives no cost for the agents.

What to do: Open the log of every time a person stopped or corrected your support agent, this week. Read it; do not ask whether one exists. If a supervisor can step in without leaving a record, your escalation rate looks better than it is, because the hard cases were pulled out by hand. Then write down next quarter's measurement period and contact volume before the quarter starts.

Our coverage →CIO: Your agentic AI results can't be checked. Here's how to fix that, by Naren Gangavarapu (18 Nov 2026) ↗

What else

  1. 01OpenAI's pricing page lists GPT-6.2 Astra, released on 17 November, at 20% below GPT-6 Astra for customers billed by usage; ask each vendor whose product runs on it, in writing, when its price to you falls. Our coverage →
  2. 02Anthropic's IPO filing of 17 November says Claude Enterprise users who did not use its coding tool spent under $75 a month on average last quarter, and the top 5% of those who did spent about 20 times that. Our coverage →
  3. 03Klarna's third-quarter results, filed on 17 November, put customer service and operations cost at $54 million, up 2% while revenue rose 12%, and give no new figure for its support assistant. Our coverage →
  4. 04TechCrunch reports Google will keep temporary Gemini chats, and chats made with history switched off, for about two months, up from 72 hours; ask your IT admin what that means for staff who paste customer details into the Gemini app. Source ↗

Stat of the day

$16.32

The median US business that pays for AI spent $16.32 per employee per month on it in October 2026, according to the Ramp AI Index, which reads card and bill payments at more than 85,000 companies using Ramp's finance software. The top 10% of paying firms spent $800.84, which is 49 times the median ($800.84 ÷ $16.32). Ramp sells spend-tracking software, and the sample is its own customers. Add up last month's AI payments, divide by headcount, and tell your CFO where your figure sits between $16 and $800. Ramp AI Index: data through October 2026, with methodology (read 18 Nov 2026) ↗

Field note

ArisGlobal, which sells software that drug companies use to process reports of medication side effects, says on an AWS blog that its AI cost per case rose more than 200% over six months. The cause it gives is not volume. As customers added their own compliance rules, the standing instructions sent to the model with every request grew 150%. Reordering each request so the repeated instructions were billed at a reduced rate cut monthly model spend about 40%, by figures ArisGlobal reported to AWS. If a vendor bills you per case, ask whether the rules you add raise the price.

One question

Does your company cap AI spend per person or per team, and who approves an overrun?

Reply to Wren → One line is plenty. Answers shape what the newsroom covers next.

— Wren

This issue went to subscribers at 7:00 ET. Get tomorrow’s by email, free →