The briefing

November 19, 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.

Starbucks switched off its AI stock counts after about nine months, and the verdict came from the store staff doing the counting. Today is about who answers when the tool is wrong; at five US banks it is now the customer. Correction: our Starbucks report of 18 November wrongly said the chain will cut its range to 6,000 or 6,500 items. That figure is BJ's Wholesale Club's, and we have removed it.

By the numbers

  • 5 banks Citi, PNC, Citizens, JPMorgan and Bank of America hold customers responsible for their AI agents, TechCrunch reports.
  • About 9 months From debut to switch-off, Supply Chain Dive reports; store staff called the counts unreliable.
  • 8 December The second Tuesday of the month; send it to a pilot group first, with a named person to undo it.
  • 100,222 VulnCheck's count by 13 November, more than double all of 2025; every fix is a change that can break something.
The counters overruled the cameras

Starbucks dropped AI stock counts after nine months; its CEO now credits daily deliveries

Starbucks ended the camera-based tool that counted stock in its stores about nine months after its debut, Supply Chain Dive reported on 4 June 2026, citing Reuters. Employees had called it "unreliable". On 18 November the trade title reported what came next. Chief executive Brian Niccol told the fourth-quarter earnings call that inventory availability is at an "all-time high", with daily deliveries reaching nearly 90% of company-operated coffeehouses in North America.

Why it matters: The fix Starbucks now talks about is trucks, not cameras. The tool, announced in September 2025, used computer vision, software that recognises items in camera images, to count part of each store's stock. Starbucks told the trade title Restaurant Dive it had "moved to a single, consistent process across all inventory counts". The verdict came from the people doing the counts. Starbucks shared what it said were internal employee comments, and one read: "Thank you for trusting the partners over unreliable spatial recognition to handle these counts." Starbucks calls its employees partners.

The catch: Nothing here can be benchmarked. The reports give no cost, no error rate, no number of stores and no vendor. Reuters reported that the tool occasionally miscounted or mislabelled items, and Starbucks did not comment on that. The "all-time high" comes with no figure and no comparison with the months the tool was live.

What to do: Before you sign for camera-based counting, ask the vendor for its miscount rate on your own items, measured against hand counts on the same shelves. In a pilot, keep the hand count as the official record, and set a date on which the store or warehouse staff decide whether the tool stays. Then ask your inventory planner how many of last quarter's stockouts came from wrong counts and how many from late deliveries.

Looking ahead: Niccol said Starbucks will meet its restocking goals in the next six to 12 months. It has been testing a model that restocks a store within 24 hours. The report does not say whether its tests of demand forecasting use AI.

Our coverage →Supply Chain Dive: Starbucks inventory availability at 'all-time high' with daily deliveries, CEO says (18 Nov 2026) ↗Supply Chain Dive: Starbucks ditches AI inventory system after just 9 months (4 Jun 2026) ↗Supply Chain Dive: Starbucks targets 24-hour inventory replenishment (6 Aug 2026) ↗

Your agent, your loss

Citi and PNC now hold customers responsible for what AI agents do in their accounts, hacks included

Citi changed its Online User Agreement this week so that customers who give a third-party AI agent access to their accounts "are responsible for activities conducted through those agents", a Citi spokesperson told TechCrunch on 18 November 2026. A third-party agent here is an AI assistant from another company that logs in to the bank for the customer. TechCrunch reports the update goes further: if the tool leaks the login or is hacked, the customer answers for whatever the hackers do.

Zoom out: PNC changed its terms on 18 November, and Citizens emailed customers the same week, TechCrunch reports. JPMorgan and Bank of America announced similar policies in October. That makes five large US banks.

The catch: TechCrunch's report covers personal banking. It does not say whether business accounts carry the same wording, and no bank gave a count of agent logins or disputed payments.

What to do: Tell your finance team in writing this week that no bank, card or payment-portal login goes into an AI tool. Ask your bank's relationship manager whether your business account agreement carries the same clause. If a tool needs your transactions, ask its vendor for a connection that can read but not pay. Anthropic says its Personal Finance tool for Claude works that way, connecting through the data provider Plaid, TechCrunch reported on 5 November.

Our coverage →TechCrunch: More banks say customers using AI agents are on their own, even if they get hacked (18 Nov 2026) ↗TechCrunch: Claude can now analyze your spending with its new Personal Finance tool (5 Nov 2026) ↗

Patch, then outage

US officials blame patching, not an attack, for a week of outages at Microsoft 365, the FAA and Chase

"We are not under systemic attack," Secretary of State Marco Rubio said on 18 November 2026, after Microsoft 365, the Federal Aviation Administration, Chase and the National Security Agency's internal network failed within six days, Fortune reported. When companies patch vulnerabilities, he said, it can create a "short-term disruption in the way your system works".

Why it matters: There is more to patch each month. The security firm VulnCheck counted 100,222 publicly listed software flaws for 2026 by 13 November, more than double all of 2025, The Register reported. Microsoft's monthly update fixed 1,094 of them on 10 November. VulnCheck's Josh Bressers told The Register that hunting for flaws with AI "is driving a majority of what we're seeing".

The catch: No source puts a cost on last week. Fortune's reference point is July 2024, when one faulty CrowdStrike update cost the Fortune 500 an estimated $5.4 billion in direct losses, with cyber insurance covering 10% to 20%, by the insurer Parametrix's figures.

What to do: Microsoft's next monthly update falls on Tuesday 8 December. Before then, ask your IT lead which updates install themselves with no pilot group, and who can undo each one. Forward staff Microsoft's 16 November warning: emails with the subject "Outage" and a "RESTORE MESSAGES" button are phishing, sent to trick staff into clicking a malicious link. Ask your insurance broker whether a faulty update with no attacker is covered.

Our coverage →Fortune: Top cyber expert warns of a 'vulnpocalypse' after 4 major U.S. systems broke in 6 days (18 Nov 2026) ↗The Register: This year's bug count just passed 100K (13 Nov 2026) ↗Microsoft Security Blog: High volume phishing campaign uses Microsoft 365 outage lures to target multiple industry sectors (16 Nov 2026) ↗

What else

  1. 01Google released Gemini 4 Neon on 18 November in Gemini for business; at Google's listed rate it costs 50% more than Gemini 4 Carbon until 30 November and 25% less from 1 December, when Carbon's rate doubles, so ask any vendor billing you at Carbon rates what your bill becomes. Our coverage →
  2. 02The Decoder reports, citing the Financial Times, that KPMG ran more than 100 AI agents through a trial audit of a FTSE 250 company alongside the normal one and plans as many as 20 more; KPMG will not say whether audit fees fall. Our coverage →
  3. 03Twilio's chief marketing officer, Chris Koehler, told Business Insider his campaigns team now produces 15 account-based campaigns a quarter, up from three, with no more than 20% of marketers moved into AI-led pods and headcount planned at "about the same" next year; he gave no tool cost. Our coverage →
  4. 04HR Dive reports that the University of Kentucky granted an employee a religious accommodation in late September that excuses him from using AI at work; if you require AI use, route refusals on grounds of belief to your existing accommodation process. Our coverage →

Stat of the day

2.1 hours a day

Regular users of AI at work in the US say it saves them 2.1 hours a day, down from 2.3 a year earlier, Indeed Hiring Lab and Glassdoor reported on 18 November 2026. In the same survey, 56% of employed respondents did not agree that AI "makes or could make me more efficient". Only 45% of AI mentions in Glassdoor's employee reviews are positive, down from 82% in 2019. The hours are each person's own estimate; no output was counted. If your business case rests on hours staff say they save, swap in one counted output per team, such as tickets closed, before and after the tool. Indeed Hiring Lab and Glassdoor Economic Research: There is no AI Jobpocalypse (18 Nov 2026) ↗

Field note

Airbnb's authors write on an AWS blog that an AI coding tool cost far more once staff used it than its tests suggested. Each request cost about three times as much as on the model before it, and the average user sent about three times as many requests, because people started long sessions in which the tool worked on its own. The tests measured one task, they write; real use measured how people chose to work. A newer model later scored higher at about a sixth of the cost per task. Check spend per user a few weeks after any rollout, not only the pilot's cost per task.

One question

Has a tool your team trialled been voted out by the people who had to use it, and what replaced it?

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 →