Starbucks Retired Its AI Inventory Tool, Then Went After $400 Million of Software Spend

Story one: the tool that got retired

In September 2025, Starbucks finished rolling out an AI inventory system called Automated Counting to more than 11,000 company-operated cafés across North America. It was built by NomadGo, a startup out of Redmond, Washington, and it used iPad Pros with computer vision, spatial computing and augmented reality to scan backroom shelves and count milk, syrups and coffee bags. The launch claims were counts up to eight times faster than manual, at 99% accuracy.

On 18 May 2026, baristas found a line in their weekly newsletter telling them Automated Counting was retired and they should go back to counting the way they count everything else. Peel the QR codes off the shelves. Nine months, start to finish.

The reporting since (Reuters first, then a detailed Fast Company reconstruction) points at a set of failures that will feel familiar to anyone who has run a category through a proof of concept:

• The 99% was a lab number. Controlled tests are not 11,000 backrooms with mixed lighting, stacked boxes and staff in a hurry. In production the system confused visually similar items and missed things.

• The prep work moved the cost, it did not remove it. Stores had to align inventory in neat rows at fixed distances for the scans to work. Some locations had to buy a new iPad Pro at around $1,000 to run the app.

• Retraining could not keep up with the business. NomadGo's CEO has been candid that computer vision struggles when the inventory itself keeps changing. Seasonal cups and limited-time packaging could take up to six weeks of retraining, and his developers sometimes found out about new SKUs when they hit the shelves.

• The backend was never going to cooperate. People who worked on the tool pointed at a 1990s IBM AS/400 system still sitting under Starbucks store data.

And the operational killer: if staff have to verify every output, the tool has not saved a task, it has added one. That is the test I would put in front of any AI business case now, before anything else.

Note what this story is not. Automated Counting was bought. It was a vendor product, running vendor models, retrained on a vendor's schedule, by a vendor's engineers. When Starbucks told NomadGo on 3 April that it was pulling the plug, the startup cut a large part of its 30-person workforce within days, including the technical team that ran the Starbucks integration. A capability that 11,000 stores depended on daily sat inside about a dozen people at another company.

So the Starbucks AI failure is not evidence that building is dangerous. It is evidence about what happens when the model, the training loop and the data pipeline all live outside your walls.

Story two: the $400 million

In July 2026, Bloomberg reported on an internal Starbucks presentation. The company spends roughly $400 million a year on software. CTO Anand Varadarajan had already told employees there was clear room to cut it. The plan: use AI-assisted development to build replacements in-house, starting with the Microsoft system that tracks inventory and the IBM platform that manages maintenance workflows, targeted for late 2027. A point of sale system to replace Oracle is also in scope, and leadership is reviewing every contract and service. Near term they are aiming at roughly $30 million out of the enterprise technology budget, including about $10 million of software.

Microsoft dropped 2.4% and IBM 5.2% on the morning the story ran. The market understood the message faster than most procurement functions did.

What actually changed

For twenty years the buy decision won by default because building meant hiring dozens of engineers, and that arithmetic never survived a business case. AI-assisted development has moved that line. Not everywhere, and not for everything, but in the middle of the stack it has genuinely moved: inventory tracking, maintenance workflows, internal approval tooling, the long tail of workflow apps you are paying six figures a year to rent.

Infrastructure, payments and anything where compliance is the actual product will keep being bought. I have no argument with that.

I would also not pretend the build path is free. The honest counter-arguments are worth saying out loud, because the vendors will say them for you if you do not:

• Maintenance debt arrives in year three, after the people who wrote it have moved on.

• Internal builds rarely come with a support contract, an SLA or anyone to escalate to at 2am.

• Legacy backends do not disappear because you built a nicer front end. The AS/400 is still there.

• Some organisations will build badly, spend more than the licence, and quietly go back to buying in 2029.

The part that matters for procurement

Both Starbucks stories point at the same variable, and it is not price and it is not build versus buy. It is who controls the loop.

This is why I keep arguing for on-premise and self-hosted AI as the default posture for any organisation with real regulatory exposure or real operational dependency. Not out of ideology. Out of continuity.

When the model runs where your data already lives, you set the retraining cadence instead of waiting six weeks for someone else's roadmap. You keep the labelled data, which is the genuinely valuable asset in any of these deployments. You do not have your supply chain visibility deleted by a strategy change at a 30-person supplier. And in regulated environments you skip the entire argument about where the data went, because it did not go anywhere.

Practical things to put in the next AI contract or business case, regardless of which way you go:

1. The double-work test. If a human has to verify every output, there is no saving. Ask for the production error rate, not the sandbox accuracy.

2. Retraining commitments as an SLA. New SKU, new packaging, new process: how many days until the model handles it, and who pays?

3. Ownership of the labelled data and the fine-tuned weights. Write it down. Assume the relationship ends.

4. Continuity and escrow. Ask what happens if the vendor loses its biggest client or its technical team. NomadGo is not a hypothetical.

5. Total cost including the change on the floor. Hardware refreshes, shelf reorganisation, training time, extra manual checks. Automated Counting failed on this line, not on the licence line.

6. The exit. Nine months from full rollout to retirement is now a realistic scenario, not a pessimistic one. Price the exit at signature.

Starbucks is not confused. It ran a bought AI tool, watched it fail in production, and concluded that the answer is to own more of the stack rather than rent it. That is the same conclusion, arrived at expensively.

The build versus buy debate is the wrong frame. The question is whether the thing your operation depends on every morning is yours.

Sources: Fast Company, "Starbucks bet big on an AI tool; 9 months later, it pulled the plug"; GeekWire, on NomadGo being blindsided; Reuters, on the retirement notice; Bloomberg, on the internal presentation covering the $400 million software spend; BusinessWire, NomadGo's September 2025 deployment announcement.