Upstream tech, translated for coffee operators.

Brewstack — Issue #001

brewstack.ai

Lead: Starbucks just killed an 11,000-store AI inventory deployment. The next pitch your POS rep makes for the same kind of tool gets one answer — skip.

Skip the AI-inventory upsell your POS rep is about to pitch.

Starbucks just killed an 11,000-store deployment that promised 99% accuracy. Cafés and roasters should refuse the same pitch — for the rest of 2026, at least.

For independent cafés and roasters of every size on every POS — Square, Toast, Lightspeed, Clover, Cropster — the next vendor call you take in Q3 will include an AI-inventory add-on. Skip it. POS stands for point-of-sale — your checkout system, the box that runs the register. On May 19, Starbucks killed the most expensive bet in retail-AI this year: NomadGo's LiDAR-based inventory counting system, retired across 11,000+ North American stores. Nine months earlier, NomadGo had claimed 99% accuracy and "up to eight times faster than manual methods." In production, baristas had to recount every scan.

That math is the whole story. If a tool requires you to verify every output, it has not removed a job; it has doubled one. Starbucks framed the retirement as "a decision to standardize how inventory is counted across coffeehouses." The internal newsletter dated May 19 was plainer: "Starting today, Automated Counting will be retired. Beverage components and milk will now be counted the same way you count other inventory categories in your coffeehouse." One shift supervisor put it on the record: "It started off not particularly accurate and got less accurate over time."

Read those failure modes carefully. The system confused similar milk types — whole, 2%, oat, almond, the four jugs sitting next to each other in every café in North America. It failed to recognize a peppermint syrup bottle in a promotional video Starbucks itself had uploaded. The most-resourced retail-AI deployment of 2026 broke on the easiest case: jugs and bottles in the back-of-house. If it broke at Starbucks scale, it will break at yours — without the vendor support and the engineering team.

This is Shipping > Adopting in one paragraph. AI inventory counting in 2026 has not cleared the bar of shipped tech that saves operators real time or money. It has cleared the opposite one: the largest production deployment of the year failed publicly, with employee quotes and a peppermint-syrup video to prove it.

Add a second data point: Andon Café, Stockholm, mid-April through late May. Andon Labs gave Google's Gemini agent a $21,000 budget and Slack-based command authority over a real café. Six weeks later, the agent had spent more than $16,000 against $5,700 in sales — and had ordered 3,000 unused rubber gloves, four first-aid kits, 6,000 unused napkins, and canned tomatoes that never made it onto the menu. Different vendor, different stack, same failure pattern: AI-agent autonomy applied to operational decisions burns money faster than the operators noticed. Stack Drift has a new shape — the AI add-on that bloats the contract and the count, without moving margin.

For most subscribers, the right answer this week is Skip. The next pitch your Toast, Square, Lightspeed, or Cropster rep makes for an "AI-powered inventory" bundle: pass. The next email promising "smart shelf" or "computer vision counts" — vision software that uses a camera to identify products — gets archived. The price-anchor pitch ("only $79 per location") doesn't matter; the cost is not the line item, it's the recount labor, the trust loss the first time a barista calls out a miscount, and the contract you can't exit in month four. Skip is the cheap, correct call for at least the rest of 2026.

For the smaller group of multi-unit cafés or roasters who already signed an AI-inventory add-on in 2025, the verdict is sharper. Rip out — but use the leverage. Send your vendor a single line citing NomadGo, and ask for pilot data from three same-size operators at six months of production use. If they can't produce it, claw back the renewal. If they can, the data tells you whether to stay.

The narrow exception — Build — applies only to roastery-with-retail operators running Cropster with a developer on contract. A Cropster API pull alerting when a green-coffee bin drops below threshold is a counting workflow worth shipping in-house. That isn't AI inventory; it's an inventory query. The distinction matters.

Verdict: Skip the AI-inventory upsell for the rest of 2026. Rip out if you already signed, and demand 6-month pilot data from three same-size operators before any renewal. Build only if you're on Cropster with developer access — and only the query, not the vision system. The verdict flips when a vendor can produce 12 months of audited production data from three independent operators showing labor saved net of recount time. Until then, the math is decided. Starbucks ran the pilot for the whole industry. Read the result.

The Stack — what else moved this week

1. Square's free-plan online card rate jumped from 2.9% + $0.30 to 3.3% + $0.30 on January 13 — a 14% per-transaction increase now five months baked into your P&L. `Indie cafés:` if your online volume is $5K+/mo, model the move to Square Plus ($49/mo per location) this week; the break-even is fast. AGMS

1. Square Vol 3 shipped May 7 with bar-tab preauthorization in Standard mode — no F&B plan upgrade required on iOS or Android app v7.3+. `Multi-unit cafés:` if you've been holding off on tab-based service because of the plan upgrade cost, turn it on now. Square Release Notes

1. The EU Commission's May 4 simplification package adds soluble (instant) coffee to EUDR scope — EUDR is the EU Deforestation Regulation — and the application date stays December 30, 2026 for large operators, June 30, 2027 for micro/small. `Roasters:` if you've been negotiating with a paid EUDR vendor, you have a six-month longer runway than feared; re-open the contract before signing a long-term seat. Comunicaffe

1. Klaviyo shipped a Model Context Protocol connector for Anthropic's Claude on May 11 — the first AI-agent integration that reads your email-marketing data directly instead of via screenshots or exports. `Roasters:` if you're on Klaviyo and pay for Claude already, the connector is worth a one-hour pilot this week; don't replatform your marketing stack for it. Klaviyo via Logical Position

1. Apple's Tap to Pay on iPhone is now live in 43 countries after a May 12 expansion to South Africa — letting any iPhone accept contactless cards with no terminal or extra hardware. Indie cafés running pop-ups, events, or wholesale tastings should treat this as a $0-CapEx upgrade; CapEx is capital expenditure, hardware you'd otherwise buy. Apple Newsroom

Field Note

Two named AI-ops failures hit production this quarter. Both broke on the same boring thing: counting and ordering. Here is the scoreboard, sourced from primary reporting — what each cost, what each promised, and what each missed. The Call's Skip verdict rests on this pattern, not on one bad week.

Starbucks + NomadGo (LiDAR inventory counting)

• Scale: 11,000+ North American stores
• Production duration: September 2025 → May 19, 2026 = ~9 months
• Vendor claim at launch: 99% accuracy, "up to eight times faster than manual methods"
• Real-world failure modes: confused similar milk types (whole, 2%, oat, almond); failed to recognize a peppermint syrup bottle in a promotional video Starbucks itself uploaded; required workers to verify every scan — doubling the count work, not removing it
• Cost: full chain-wide write-off plus recount-labor tax across every store-day during the deployment

Andon Café + Google Gemini ("Mona" agent)

• Scale: 1 café, Stockholm
• Production duration: mid-April → late May 2026 = ~6 weeks
• Vendor claim: full operational autonomy — hire staff, place orders, manage inventory via Slack
• Real-world failure modes: over-ordered bread some days, under-ordered others; bought 3,000 unused rubber gloves, four first-aid kits, 6,000 unused napkins, and canned tomatoes never used on the menu
• Cost: $21,000 starting budget → $16,000+ spent against $5,700 in sales (~$10K burn in 6 weeks)

The pattern: when the AI's output requires verification, the operator pays twice. Sources: Fortune, TechTimes, Futurism.

⭐ Do this week

If you run any café or roastery on any POS: when your vendor's next sales call mentions "AI-powered inventory," "smart shelf," or "computer vision counts," respond with one question — "Show me 12 months of audited production data from three operators my size." If they can produce it, listen. If not, end the call. Save the email; you will get the same pitch from three more vendors by Q4, and the same question disqualifies every one of them. Why now: every POS vendor's roadmap has an AI-inventory bundle in market by Q3; the pitches start landing this month. Cost: 30 seconds per sales call. $0.

Sign-off

A good operator stack is one where every line item still earns its seat a year later. Most AI add-ons sold to cafés in 2026 won't survive that test — they're sold on demo, not on data. Starbucks ran the pilot for the industry and the answer came back: not yet. Independent operators have the easier job. They get to read the result and skip the pilot.

Next Tuesday: a closer look at one stack layer where the verdict actually flips Buy — payments, where a quiet 14% rate hike on Square's free plan is now five months baked into café P&Ls.

Until next Tuesday.

— Stef. from Brewstack

Forwarded? Subscribe at brewstack.ai.