In May 2024, Sonos pushed an app update which broke its own product. By the start of 2025, the CEO and several other executives had been pushed out. Around 100 staff were gone. Wirecutter had pulled its long-standing recommendation. The remediation was costed at twenty to thirty million dollars in short-term spend, with two product launches delayed and the associated revenue with them.
I do not write this as a Sonos post-mortem. Plenty of those exist. I write it because the pattern which took Sonos to this point is the same pattern most boards are signing off, right now, in the name of AI transformation. And the people who lost their jobs over it had spent the previous year being told the strategy was bold.
What Sonos did
The app rebuild shipped alongside the Ace headphones launch. The new app went live without a substantial portion of the features long-term users relied on. It also shifted the underlying architecture away from Sonos Net, the local mesh which had been the product's quiet engineering achievement for years, and onto a cloud-dependent model which introduced lag and surfaced network issues the old architecture had absorbed silently.
Customers noticed within hours. Sonos did not roll back. The decision to push forward through the complaints, rather than restore the working version while the new one was fixed, is the decision which made everything else inevitable.
A new app, on a new architecture, with reduced feature parity, released without the level of beta testing a hardware-tied software platform requires. Three changes at once, each significant, each capable of producing the failure on its own.
The trade the board signed off
Read the public communications from this period and the trade is obvious. New revenue from the Ace launch was prioritised over existing user trust. New architecture was prioritised over reliability. New interface was prioritised over feature continuity.
The justification in each case was the same. We have to move forward.
Boards approve trades like this every quarter. The trade looks reasonable on paper. The risk language sits in an appendix somewhere. The rollback plan exists in name. The user research is summarised as positive.
The mistake is not in approving the trade. The mistake is in believing the trade has been properly priced. Sonos paid twenty to thirty million dollars in remediation. They lost share price, market valuation, and reputation which does not come back at the same speed it leaves. They lost a CEO.
The cost was not the cost of building the new app. It was the cost of building the new app while breaking the old one.
Why this is the AI story
The pattern playing out across enterprise AI rollout looks like Sonos did in early 2024. New capability launched alongside existing services, on new architecture, with reduced parity, without the beta testing the previous generation of software releases would have demanded.
The rationale is identical. Move first. Capture the productivity gain. Iterate after launch.
Three things should give a board pause before signing off on this approach.
First, AI rollouts almost always replace something which already works. Customer support bots replace human agents. Drafting assistants replace process. Predictive systems replace existing decision aids.
The thing being replaced has known failure modes, established trust, and a feedback loop the organisation has spent years tuning. The replacement has none of these. The Sonos lesson is what happens when the replacement ships before parity is established.
Second, automated testing is no substitute for user testing. Sonos had already replaced dedicated test expertise with developer-written unit tests, a trend the industry had flagged as a risk well before the app launch. AI rollout intensifies this.
The systems being shipped have probabilistic outputs. They behave differently for different users in ways unit tests do not surface. Boards approving AI rollouts on the basis of automated test coverage are approving the same trade Sonos approved.
Third, the rollback plan needs to be real. Not a statement one exists. A working version of the previous service, maintained in parallel, with a defined trigger for switching back, owned by someone with the authority to pull the switch. If the board cannot point to this person and this trigger, the rollback plan is theatre.
What changes in the boardroom
A C-suite reading this should take three things from the Sonos experience and apply them to the next AI sign-off.
The first is to ask what is being broken to make the change. Not what is being added. What is being lost or reduced or made less reliable. If the answer is nothing, challenge it. Something is always being traded.
The second is to ask who tested this with real users, in their working environment, before the live release. Not in a controlled pilot. Not in a focus group. In their daily working context, with their normal workload, on their normal devices. If the answer is no one, the board is approving a launch on automated tests alone.
The third is to ask, plainly, what triggers a rollback. If this question cannot be answered in a sentence, the rollback plan does not exist. The remediation will be reactive, expensive, and visible. The CEO who has to manage the response is the one whose name appears in the press release explaining the departure.
The Sonos board did not have these answers in May 2024. By January 2025, they were paying for the absence in the most expensive way available. The next AI rollout your board signs off this quarter is, structurally, the same decision. Do not assume the price has been paid.