five signals.
one week.
every fix was a call
nobody made._

sinch surveyed 2,527 senior decision makers across ten countries and found 74 percent had already rolled back or shut down a live ai customer communications agent after deploying it. the rate rises to 81 percent at the companies with the most mature guardrails, which is the detail worth sitting with. gartner, off its own survey of 321 service leaders, expects half the companies that credited ai for headcount reductions to rehire the same functions by 2027 under new titles.

deployment is the easy part. the part nobody budgets for is the day the agent gets a case it was never scoped for. the cost is not the license. it is the cut you made in month one, the severance, the rollback, the rehire twelve months later at whatever the market charges by then, and the customers who spent the gap talking to something that could not help them and could not say so.

the fake fix is a better model or a bigger platform, on the theory that the last agent failed because it was not smart enough. the ones getting pulled are not failing at language. they are failing at governance: no rule about what the agent may promise, no line where it hands off to a person, and nobody reading what it actually said. do the version that holds: before it goes live, write the three things it may never do without a human, money out, commitments, anything you cannot reverse. then write the trigger that hands the conversation to a person, name who reads its transcripts every week, and keep headcount flat until it has run one full quarter with no rollback.

mit put a number on what buyers have been feeling for two years: 95 percent of enterprise genai pilots deliver no measurable roi. an entire demo economy lives inside that gap, and it is well funded.

the condition is simple to name. pilots get designed to impress the room rather than to move a number, so they end the only way they can, in a deck about learnings. the cost is not the pilot budget. it is two quarters of calendar, the integration hours your own people burned, and the internal credibility the next attempt now has to buy back at a premium.

the fake fix is another pilot with a newer model, which restarts the same clock with the same missing scoreboard. do the version that holds: before anything starts, write the one number the pilot has to move and the date it has to move by. then ask the vendor for a customer twelve months in production and call that customer yourself. a vendor with real deployments hands you the reference by end of day. one selling demos schedules another demo.

somebody has published a ranked comparison of 34 courses on how to start an ai agency in 2026, which means that market now has enough supply to warrant comparison shopping.

the condition it creates lands on your desk rather than theirs. the person pitching you an ai build may be six weeks out of a course that taught them to white label someone else's tool and put a retainer on it, and the pitch sounds good because the pitch is what the course actually sold. the cost is a build that works in the demo, breaks on your real data, and cannot be repaired by the person who sold it, because they never wrote it.

the fake protection is asking whether they have done this before, which was module three. do the version that holds: ask what the system does when the data is messy, ask them to name the underlying tool and what happens to you if that tool changes its pricing or its api, then ask for the last client still running it and call that client yourself. someone who built it answers in a sentence. someone who bought the answer sends a case study.

pwc's 2026 ai jobs barometer describes a labor market splitting in two, and the side gaining value is the human one: judgment, communication, the ability to be trusted with a customer or a number. 48 percent of hiring managers now put communication and collaboration at the top of what they screen for, which is an odd result for a year everyone spent buying software that writes.

the condition underneath it is simple. the tasks ai handles cleanly are the documented, repeatable ones, so what stays with a person is the exception, the judgment call, and the conversation nobody wants to have. the cost is a hiring bar you never updated: you keep screening for tool proficiency and output volume while the roles that remain fail on the parts you never tested for, and you find out in month five.

the fake fix is adding an ai skills line to the job post, which selects for people who can name tools rather than people who can decide. do the version that holds: take your two hardest open roles, write down the three decisions that person will actually own, build the interview around those decisions instead of the resume, and price the offer against the judgment rather than the tool list.

the volume moved from roughly 500,000 online in 2023 to an estimated 8 million in 2026. the version that reaches a company your size is not a political video. it is a thirty second phone call to whoever is authorized to move money.

the condition underneath it is that a familiar voice has always functioned as identity inside your business, so an urgent request in that voice skips every control you have, because the controls were written for documents and logins. the cost is a same day wire your bank cannot recall after the cutoff, plus the part nobody prices, which is a controller who spends the next quarter second guessing every legitimate request.

the fake fix is training people to listen for something off in the audio, and hearing the difference stopped being a human skill about a year ago. do the version that holds: write one rule that any request to move money, change bank details, or send payroll data gets verified by calling back on a number already in your records, never a number offered during the call. apply it to yourself louder than you apply it to anyone else, and tell your team in writing that nobody gets penalized for making the boss wait ten minutes.