For about two years, the industry that sells machine conversations kept score with a word that sounds like success and is not: containment. It counted how many calls never reached a human being. Last week that word quietly lost its job.
At a summit broadcast on theCUBE, SiliconANGLE’s own studio, the analyst Zeus Kerravala of ZK Research put the new rule in one line: “Resolution and resolution quality is the new unit of value.” Agentic systems, he said, should be judged on “whether the customer’s needs were completed — and completed actually across the full journey.” His co-host, theCUBE Research’s Bob Laliberte, took the same view. The pair were closing out four days of conversations with Cisco, Talkdesk, Zoom and Five9.
Read that slowly, because it is an admission. For two years the machines were graded on what they prevented. From now on they are graded on what they finished. Those are not the same number, and for most businesses they have been drifting apart the whole time.
The word they dropped
Three days earlier, on the same broadcast, Zoom’s head of product for AI in its customer experience business, Ram Rajagopalan, said the quiet part plainly. “Many customers, even about a year, 18 months ago, focused on purely measuring containment,” he said. “Where I see conversations these days going is not just looking at containment, but end-to-end resolution… not just answering the basic inquiries, deflecting the call from going to a human agent, but actually completing the task that the consumers are calling in for.”
A call can be contained and still leave the person holding the phone with nothing. It was handled. It was answered. Nobody wrote down whether anything got done.
Adoption was never the hard part
On Monday, at the same event, Pedro Andrade of Talkdesk gave the sentence that should be pinned above every small business owner’s desk: “Adoption is easy. The orchestration is the hardest part.” Talkdesk’s own research, he said, found that 98% of companies have now deployed AI somewhere along the customer journey, while only 15% connect it across departments; the obstacles his respondents named were compliance at 50%, security at 48% and disconnected systems at 45%. The companies his firm classes as mature report about four times the improvement in customer sentiment scores — 22% against 5%.
Set the vendor’s numbers aside and keep the shape of them. Nearly everyone has switched something on. Almost nobody has checked whether the thing at the other end of it finishes.
The same discovery, in a plumbing magazine
None of this needs a summit. On Monday, Plumbing & Mechanical ran a piece by Chris Mechanic with a homework assignment in it, and it is the best hour a service business will spend this week. Pull up your phone log, he writes, and look at what happens to calls that arrive after 5 p.m., and on weekends. Citing the call tracking platform Invoca, he notes that about 26% of inbound calls to businesses go unanswered; in his own experience, nights and weekends are where the real hole is. Companies that run the exercise, he writes, typically find that a fifth to a third of their calls arrive outside office hours, that most of those are never answered, and that the next-morning callback converts at a fraction of the rate a live answer does.
Compare that with the implementation partner Joe Rittenhouse, on the same theCUBE panel, describing where he starts with enterprise clients: “Just a simple question of what do you do after hours? We’re usually staffed, we follow the sun from the East Coast to the West Coast, but then after hours we just have a general mailbox. And the answer is typically, I don’t know.”
A billion-dollar contact center and a two-truck plumbing shop have arrived at the identical blind spot, and neither of them found it in a dashboard.
The same magazine had gone further the week before. In a study of 100 plumbing and HVAC contractor websites across Dallas and Houston, Michael Carpenter found that 34% never responded to a web inquiry at all — “not slowly, not poorly; never.” The top tenth replied in an average of 23 minutes. The bottom group took more than six hours, in a category where, he writes, the first contractor to reply wins the job more than 70% of the time. Those are not lost jobs. They are jobs nobody knew they had.
Why the polished version fools you
There is a reason this goes unnoticed, and MIT Sloan Management Review named it on Monday: the capability mirage. Drawing on a joint study by the consultancies Anthrome Insight and Axialent, the authors describe organizations that look highly capable while real skill quietly erodes underneath polished machine-made output. Stephanie Antonian, who runs the AI product company Aestora, told them: “Everyone can produce a level of work that’s pretty good for basic tasks… It looks pretty good. But then you don’t know what’s underneath it, how resilient that piece of work is.” AI, the piece concludes, is not necessarily an improver of work. It is an amplifier.
Writing at Search Engine Journal that same Monday, Greg Jarboe put numbers under the same idea, and they are worth knowing because they are counterintuitive. In a METR study, 16 experienced developers working on 246 real tasks expected AI to speed them up by about a quarter; they finished roughly a fifth slower, and afterwards still believed they had been faster. A BetterUp Labs and Stanford survey of more than a thousand workers found that fixing machine-made work that looks finished but is not takes close to two hours each time. Workday’s research, he notes, found that for every ten hours AI saves, about four come back as rework.
The output looks done. The feeling says faster. The log says otherwise, and the log is the only one of the three that can be checked.
And the number you cannot get from outside
One bridge to yesterday, and then we are past it. If you read this column on Tuesday, you know the outside is barely measurable: nobody can yet tell you reliably whether an assistant names your business. Sean McCrohan, who runs technology at the call tracking firm CallRail, gave Search Engine Journal the plain version on Monday: clicks arriving from citations in AI answers account for roughly 1% to 2% of calls among his company’s customers — about double what it was in January — and following an assistant’s work all the way to a ringing phone is very hard right now, because the AI platforms do not yet hand out the tools that ad networks do.
That is precisely the argument for turning around. The outside numbers are modeled, small and disputed. The inside numbers are exact, free, and already sitting in your phone system, your inbox and your chat widget.
Three checks an owner can run this week
1. Break your answer rate out by hour, not by month. Your phone system or call tracker can show calls by time of day. Find the share that arrives outside your hours, then your answer rate inside that band alone, then how many of the next-morning callbacks turn into an actual booking. That is the Plumbing & Mechanical exercise, and it takes an evening. Most owners discover their worst hour is the one they never look at.
2. Pick one journey and write the number down before you change anything. The analysts closing that summit gave the same advice they would give a bank: take one bounded, high-value path, record the baseline, then measure. Rittenhouse’s version was blunter — “You just can’t boil the ocean… eat the pizza piece by piece.” If you cannot say today what share of inquiries end in a booking, you will not be able to tell next month whether anything you bought helped. And Kerravala’s warning applies to every size of business: “If you’ve got a broken process, you’re going to get to that bad destination faster.”
3. Read your own logs for the customer’s words, not yours. McCrohan’s point on that panel was that people describe their problem in words that may not match the words on your website — and where the two do not meet, you simply never come up. His co-panelist Steve Wiideman added that most businesses never read their own site chat logs, and that working through calls and voicemails turns into a content plan on its own. The most valuable sentences about your business this week were spoken by your customers, into your own recordings, and nobody has listened to them.
How we think about it
The old measure asked whether anything reached you. The new one asks whether anything left finished. A machine that answers everything and completes nothing is worse than a voicemail, because the voicemail at least admits it.
That is the standard we build AI employees against, and it is the order we argue for: hire the AI employee first, then a person for what is left.
Avery, our AI Administrator, answers the message that arrives at night or on a holiday, works out what the person actually wants, and carries it to a booking or to a request that lands on your desk instead of dying in an inbox. He does not invent anything that is not in your base.
Ruby, our AI Customer Support, answers where those conversations land for a small business: the direct messages of your connected channels and the comments under your posts. She answers out of the rules you wrote, and where a rule is not written, she says she does not know and brings the question to you — because an invented rule is more dangerous than a silence.
Neither of them picks up a telephone, and we are not going to pretend otherwise. What they do is close the gap between “somebody wrote to us” and “the thing got done,” in the channels where a small business is actually written to.
You do not learn a dashboard to run them: you describe your business in plain words, in the ChatGPT, Claude or Gemini app you already use. And before you hire anyone, you can watch our AI employees working in public on two demo storefronts, Tampa Pasta House and Casa Lista.
The industry spent two years counting what it managed to avoid. This week it started counting what it managed to finish. Your business can make the same change tonight, for the price of one evening and no software at all.
Sources
- SiliconANGLE / theCUBE, Sept. 14, 2026, 22:38 UTC, Cheryl Knight — “Contact center AI ROI shifts toward resolution quality and governed execution”
- SiliconANGLE / theCUBE, Sept. 11, 2026, 14:20 UTC, Kristen Nicole — “How conversation to completion moves contact centers past containment”
- SiliconANGLE / theCUBE, Sept. 14, 2026, 21:28 UTC, Cheryl Knight — “AI adoption outruns readiness as contact centers chase end-to-end resolution”
- Plumbing & Mechanical, Sept. 14, 2026, Chris Mechanic — “The Astounding Cost of Missing Calls”
- Plumbing & Mechanical, Sept. 8, 2026, Michael Carpenter — “Why Plumbing Contractors Are Losing Leads They Don’t Know Exist”
- MIT Sloan Management Review, Sept. 14, 2026, Melissa Swift, Teryluz Andreu and Dolores Hernandez — “How AI Creates a Capability Mirage”
- Search Engine Journal, Sept. 14, 2026, Loren Baker — “How To Connect AI Search Visibility To Local Leads”
- Search Engine Journal, Sept. 14, 2026, Greg Jarboe — “AI Isn’t Delivering Marketing Efficiency, It’s Repeating Programmatic’s Broken Promise”
Note on sources 1–3: these are SiliconANGLE’s own theCUBE broadcasts of “The AI ROI in Contact Center Summit,” carried with the publisher’s disclosure; the speakers quoted are named analysts and named vendor executives speaking at that event.
