ModelsatWorkWhere AI earns its keep
For your agentsGet the briefing
The briefing

November 12, 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.

Airbnb says AI now resolves about 40% of its support issues, and its own filing shows the expense line that holds support still rose 8% to $393 million. Today is about what the AI bill is made of. Zendesk wants to charge by outcome and has published no price, and Salesforce now sells its assistant by the seat at $50 a month. Each number needs a second figure beside it before it reaches your CFO.

By the numbers

  • $393M, up 8% Airbnb's quarterly filing, in the quarter it says AI resolved about 40% of support issues.
  • Down $26M Airbnb's filing credits AI for the fall; payroll in the same line rose $21M.
  • 500 actions a month $50 seat ÷ $0.10 an action, from Salesforce's pricing page (20 credits at $500 per 100,000).
  • $1.50 to $2.00 Zendesk's pricing page, per request its AI closes; no rate is published for the new Outcome plans.
  • 7 test cases Northern Trust and Microsoft's footnote; no error rate on live clients is given.
Resolved 40%, spent 8% more

Airbnb says AI resolves 40% of support issues, and its support expense line still rose 8%

Airbnb told shareholders on 5 November 2026 that about 40% of its customer support issues in the third quarter were resolved without a human agent, up from a third the quarter before. It said support cost per booking fell more than 17% in a year, "driven in large part by improvements to our AI assistant". Its quarterly filing for the same period shows the expense line that holds support, "operations and support", rose 8%, from $365 million to $393 million.

Why it matters: A CFO reads the budget line, and Airbnb's filing shows three movements inside it. Spending on outside support partners and other third-party services fell $26 million, "due to lower agent contact volume resulting from increased use of artificial intelligence". In-house payroll in the same line rose $21 million on higher average headcount. Customer relations costs rose $33 million on "higher make-good payouts and related case reserves", which the filing does not connect to the assistant in either direction. The saving the filing records landed with outside partners, and the line still grew.

The catch: The letter defines neither "resolved" nor "cost per booking", and gives no dollar figure for the second. It does say the 40% counts all support issues, "not just those that begin in the AI assistant". It does not say whether a customer who gave up counts as resolved. Neither document gives the cost of building or running the assistant. CX Dive reports that chief executive Brian Chesky said on the earnings call that "customer service NPS is as high as it's ever been", a customer satisfaction score; the report quotes no number.

What to do: Ask finance for three lines in dollars, for the quarter before your assistant went live and for the latest one: outsourcer spend, in-house support payroll, and refunds or goodwill payouts. When a vendor quotes a share of issues resolved by AI, ask whether it is a share of every issue, as Airbnb's is, or only of conversations the assistant touched. If your saving depends on an outsourcing contract, have procurement check its volume minimums before the forecast reaches the CFO.

Our coverage →Airbnb Q3 2026 shareholder letter (Form 8-K, Exhibit 99.1, filed with the SEC) ↗Airbnb Form 10-Q for the quarter ended 30 September 2026 ↗CX Dive: Airbnb reduces customer support cost per booking 17% ↗

Unlimited seats, no price

Zendesk announces unlimited-seat plans billed by outcome and publishes no rates

Zendesk said on 10 November 2026 that its new Outcome AI Platform plans, sold as Zendesk Unlimited, come with unlimited seats. Customers "set an Outcome Budget and pay for eligible issues resolved and business work completed". The press release gives no rate per outcome, no minimum budget and no start date.

Why it matters: Zendesk's pricing page, read on 12 November, still calls its pricing "primarily seat-based". It lists plans at $55 and $115 per agent per month, paid yearly, and charges $1.50 for each request its AI closes without a person on a committed volume, or $2.00 pay-as-you-go. A business case that paid for the AI by cutting seats does not carry over to a plan with no seat fee.

The catch: "Zendesk evaluates each outcome for quality and completeness", the release says, so the seller decides what is billable. It describes no way to dispute a count.

What to do: If you renew within two quarters, ask your account manager in writing for the Outcome rate card, the minimum budget and the definition of a billable outcome before you sign another seat-based term. Have an analyst pull last quarter's seats, total tickets and AI-resolved tickets, so you can price both models the day rates arrive.

Our coverage →Zendesk: Zendesk Upends Enterprise Service with Outcome AI Platform, Unlimited Seats (press release, 10 Nov 2026) ↗Zendesk pricing page (read 12 Nov 2026) ↗

500 actions a month

Salesforce sells its Agentforce Coworker assistant at $50 per user per month

Salesforce said on 4 November 2026 that companies can buy Agentforce Coworker, the assistant a seller or service manager hands a job to in plain language, for $50 per user per month. Its pricing page, read on 12 November, also sells metered Flex Credits at $500 per 100,000.

Why it matters: The page prices an action, one step the assistant takes, at 20 credits. That is $0.10 an action ($500 ÷ 100,000 × 20), so $50 buys about 500 actions a month, or 25 a working day. Salesforce's own example has 100 users each handling three cases a day at 60 credits a case: $1,800 a month on credits. The same 100 people on seats cost $5,000.

The catch: The page lists "All Coworker searches included" with the seat. It does not say whether actions draw credits on top, or what happens past the 30 MB of company data each user may make searchable a month. Salesforce counts more than 185,000 Coworker users and gives no cost for any customer it names.

What to do: Run Coworker on credits for a month with one sales or service team and count actions per person. Buy seats only for people well above 500 a month. Get Salesforce to state in writing whether the seat covers actions.

Our coverage →Salesforce: Agentforce Coworker usage soars 21x in three months (announcement, 4 Nov 2026) ↗Salesforce: Agentforce pricing (read 12 Nov 2026) ↗

What else

  1. 01DXC and AWS say DXC's OASIS agents for managed IT have cut resolution time 95% on average across 56 measured use cases since March 2026, with every agent starting read-only; their jointly written post gives no cost. Our coverage →
  2. 02Zendesk's consumer survey, published in the release that announced its new plans, found 84% say human help should always be available; the same release says customers can automate up to 84% of service requests. Our coverage →
  3. 03DaVita's CIO Madhu Narasimhan tells CIO that employees have created more than 10,000 agents since a company-wide AI platform launched this spring, and that staff from clinical, legal and cybersecurity teams review each use case. Source ↗
  4. 04A post on the AWS blog says customers of treasury software firm Trovata have seen more than $1 million in annualised savings and over 5 hours a day saved per accounting team; the figures are the companies' own. Source ↗

Stat of the day

7 test cases

Northern Trust and Microsoft said on 9 November 2026 that an AI agent cut the data preparation for onboarding a new institutional client by 83%, from about two and a half weeks to three days, with analysts reviewing its output. The 96.5% quality score in the same release comes from seven test cases built from anonymised past client workbooks, a footnote says. The release does not say how many live onboardings sit behind the three-day average, or what the work cost. When a vendor quotes a quality score, ask how many cases it covers and whether any were live. Microsoft Source: Northern Trust and Microsoft Frontier Company joint press release (9 Nov 2026) ↗

Field note

Mike Tomasco, technology chief (CDIO) at Fujifilm Biotechnologies, tells CIO how employees there build their own agents. All 5,000 staff may apply to a restricted sandbox; about 50 work in it. A finished agent is tested by coworkers, then reviewed for regulation and duplication, and typically reaches production six to eight weeks after the idea, he says. The legal team's contract-review agent, which proposes edits for the lawyers to check, is due in production in early December. After release the company measures adoption and decides whether the agent stays. "There's no sense in having something run that's not creating value," Tomasco says.

One question

When you report what a support ticket costs, which lines are in the figure: outsourcer spend, your own payroll, refunds, or the AI's running cost?

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

— Wren

ShareLinkedInXBlueskyHNEmail

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