At Orvera AI, we come out of contact center operations, with more than 18 years of it behind us, and we have watched enterprise artificial intelligence fail in that setting from close range. It almost never fails the way the headlines say it does. It rarely fails at the demo or go-live. It fails in month three, on a Tuesday, quietly, and by the time anyone reports it the number that moved has already moved back.
Sinch surveyed 2,527 senior decision makers across 10 countries this year, and 74 percent of them said they had already rolled back or shut down a live AI customer communications agent. The reversals came after go-live. The systems launched, then they stopped being right.
The industry's answer to the deployment gap is an engineer. Forbes called the forward-deployed engineer the most expensive job in enterprise in May. Anthropic and OpenAI both launched enterprise deployment ventures that month, days apart, worth roughly $5.5 billion between them. ServiceNow and Accenture announced a joint program on May 6. AWS committed a billion dollars to its own organization on June 30, and Microsoft put $2.5 billion and 6,000 engineers behind a new unit on July 2. We think the diagnosis is correct and the fix lands on the wrong phase of the work.
Here is why. An AI system in a contact center is calibrated to a snapshot: last quarter's reasons for contact, the policies in force on the day it was configured, a quality scorecard written for a process that is about to change. A contact center does not hold still for any of that. Marketing runs a promotion and the intent mix rearranges itself. Legal revises a disclosure and four flows are now giving an answer that was correct in March. A routing change makes the containment number look better than the operation actually is, and nobody notices for six weeks because the dashboard is green.
None of those are engineering failures. An engineer can fix any of them in an afternoon once a person decides what the right answer is. Deciding is the work, and the deciding never stops.
Forrester Research's Kate Leggett wrote the honest version of where that work goes. Last November she predicted that 30 percent of enterprises would create parallel AI functions that mirror human service roles, with managers to onboard and coach AI agents, operational teams to optimize AI performance, and specialists to unblock AI when it falters. Her advice to leaders was to start scoping those roles now.
We would put it differently, because we sit on the other side of the table. That prediction hands a vendor's gap to the buyer to staff. The software arrives, and the enterprise is asked to hire a function it does not have to keep alive a system somebody else built, on a floor where AI is not viewed as essential to their daily work.
The role the market has not staffed is not a strategist nor a second engineer. It is a person who can read a floor, who has sat in a calibration session, who knows how a scorecard drifts, who can look at an interval report and tell you that the containment rate is being flattered. Those people exist in large numbers. They are in outsourcers and in operations departments. They are not in AI labs, and the labs are not recruiting them.
We do not think this should be two people either. Palantir has paired a forward deployed engineer with a deployment strategist for years, and both halves of that pairing are scoped to getting a system built and put into use. The decisions that break a contact center deployment do not wait for a handoff. They arrive weekly and mixed together, so the person who catches them needs to be able to read the report and change the flow.
So we built the role as one role: engineers who have also run operations, on the account after go-live rather than rolling off when the implementation plan says so. We can do that because we run the system we deploy instead of handing it over, and because we came out of contact center services rather than out of software. We are aware that this is a convenient thing for us to believe.
That is why we would rather be judged on a standard than on a story. Ask any vendor, including us, what the same account's numbers looked like in month six against month one. Ask who was still on the account, by name and by role, 90 days after launch. Ask what happened the last time a policy change broke a live flow, how long it stayed broken, and who caught it.
Those questions are answerable. They separate the companies who ship a system from the companies who run one. And they will tell a buying committee more than any AI roadmap presented this year.
Ali Merchant is CEO of Orvera AI, provider of an agentic AI platform for enterprise customer experience.