Independent
Software Developer
HR Manager
Team Leader
Accountant
Copy Writer
Customer Service
Virtual Assistance
The Rollback Nobody Priced In
Sinch’s 2026 customer-experience research found that 74 percent of enterprises have shut down or rolled back a live AI customer-service agent. Gartner’s September survey put a number beside it: only 22 percent of enterprises have scaled AI across multiple business units. Salesforce, on September 11, told the market every one of its Q4 top-10 wins included Agentforce. The three numbers describe the same trade, seen from three positions on the P&L.
For nine Wednesdays this column has tracked the money moving from labor services into AI. Last week’s turn was HSBC’s observation that Philippine services employment has continued to grow through the transition. This week the arithmetic runs into the buyer’s production environment, which, on the evidence, is where most of the sale is failing to hold.
Sinch’s finding, published this year, sits in a research thread that also includes work from Zendesk, Gartner, and the Harvard Business Review. Each of those studies, in different language, describes the same pattern: a large majority of enterprise AI deployments in customer service do not survive the year after go-live. That is not a small caveat to the productivity story. It is the story’s other half.
The rollback number
Sinch’s headline — 74 percent — measures a specific event, not sentiment or intent. Each enterprise counted put a live AI agent in front of customers, and then took it back down. The research covered organizations across financial services, healthcare, retail, and technology, and the pattern held across geographies.
Gartner’s parallel survey, of 1,303 enterprises with annual revenue above $50 million, framed the same phenomenon from the other end. In Gartner’s own language, “only 22 percent of organizations have successfully scaled AI across multiple business units.” Eighty-five percent of the functional leaders in that survey said they plan to spend more on AI in 2026 than they did in 2025. Gartner’s forecast on where the additional spend lands is that 40 percent of agentic-AI projects will be canceled by 2027, with runaway costs, unclear return on investment, and governance disputes cited as the primary reasons.
The market has not slowed down. The rollbacks are trailing the deployments, not preventing them.
Where deployments tend to break
The failure pattern, in the post-mortems that have been published, is unromantic. An agent that closes 90 percent of pilot journeys often closes 30 percent of the live ones. Knowledge bases described as “sufficient” during procurement turn out, on inspection, to be six months stale on the articles agents actually touch. Escalation paths sized for edge cases collapse under normal case volumes. Governance the buyer expected to be handled by the vendor tends to be a compliance disclosure the buyer, in fact, owns.
None of that reflects on the underlying model quality. It reflects on the operational work required between a working model and a working deployment, and that work has historically been what enterprise IT is best at. HFS Research’s August report called the resulting shape “Services-as-Software”: the buyer pays for software, but the vendor still has to do services work to make it hold. The 74 percent number is largely what happens when that services work is skipped or under-scoped in the sales cycle.
What Salesforce is actually booking
Salesforce’s Q4 disclosure, on the same September 11 briefing that unveiled its outcome-based pricing model, tells a genuinely different story. Every one of the top-10 customer wins in the quarter included Agentforce. Across the wider Agentforce base, Salesforce reports over $100 million in annualized cost savings and a 34 percent lift in customer-reported productivity. Named cases include Hibbett, whose AI now handles 90 percent of core shopper journeys after a six-week deployment, and Asana’s Piper agent, which drove four times the conversation volume after a 45-day rollout.
Two things about that base worth naming. First, the customers Salesforce cites tend to share a profile: a mature CRM data model, a clean knowledge base, and an in-house team that owned the deployment work end-to-end. That profile fits a smaller subset of enterprises than the market as a whole. Second, the population overlaps materially with the 22 percent Gartner measures — the buyers who successfully scaled AI. Both surveys, read in parallel, describe the same buyers at different resolutions.
What the two numbers reconcile to
The productivity number and the retention number, taken together, are less contradictory than they look. Enterprises that had done the operational readiness work — data model, knowledge governance, escalation design, in-house delivery capacity — got the productivity outcomes Salesforce cites. Enterprises that had not, and there are more of them, produced the rollbacks Sinch measures. Both scoreboards are accurate. They describe different populations of the same enterprise base.
For a BPO or IT-services buyer, the implication is that “does my vendor have AI” is no longer the useful renewal question. Every vendor does. A more useful question, if the numbers above are representative, is whether the buyer’s own team can build and maintain the operational scaffolding that lets an AI deployment survive contact with a live workload — and, if not, which vendor in the shortlist is willing to do that work as part of the contract.
Renewal season is the natural place to press the point. A vendor prepared to answer “which of your live agents survived twelve months in production on a workload comparable to mine” is a different vendor from one who can only show the productivity slide.
The question for your business
When your AI vendor pitches the productivity number, are they also willing to show the retention number on a workload that looks like yours?
