By Damian Mathews & The Last Mile Team
Sooner or later someone is going to ask you to prove the AI worked.
Most contact centers will reach for containment rate, and that number will not survive the conversation.
The evidence for that sits in Forbes reporting on the 2026 Enterprise AI Strategy Pulse Survey, published on August 6. The sample skews to Fortune 500 and Forbes Global 2000 companies, which means this is the top of the market rather than the long tail.
Seventy-four percent of those enterprises now run at least one AI solution in production, and 93 percent are piloting or further along. Half of the ones already in production cannot consistently tell you whether any of it worked.
Adoption was never the hard part. Production is, because production is where somebody finally asks what the thing costs and what it returned. A pilot can be defended with a demo. A production system has to be defended with numbers a finance team already trusts.
Nearly everyone shipped it. Half cannot prove it.
The gap is already changing behavior. The narrowest deployments turn out to be the worst measured. Among enterprises running AI inside a single business function, 74 percent say ROI is either too early to measure or is not tracked at all. That is worse than the topline figure, and it is backwards, because one function should make attribution easier, not harder. A use case that goes live without a pre-deployment number is unmeasurable for the rest of its life. There is nothing left to compare it against. We covered an early version of this divide in The 20% Club.
It also found where the problem now lives. Only 21 percent of AI ownership still sits with a chief AI officer or a center of excellence. Thirty-seven percent sits with functional and line-of-business heads, which is the right home for it, since value is created inside the function. But ownership moved into the business faster than the accounting did, and the space between the two is where the ROI question goes to die.
Read that as an operator rather than an analyst.
What separates the survivors is visibility. They could see what was happening clearly enough to make the argument when someone asked.
In a contact center, that is exactly where things break down.
Containment rate is the number most operations report upward, and it is the least trustworthy figure in the building. Containment records where a conversation ended. It says nothing about whether the customer’s problem ended with it.
A customer who gets stuck in a loop and gives up is contained. A customer who abandons and calls back tomorrow on another channel is contained twice.
So the metric climbs while the experience degrades, and it is the number being used to defend the spend. A measure stops being a good measure the moment it becomes a target, which containment did years ago. Michael Fisher walked through the mechanics of that in Ardeo Ergo Sum.
There is a structural reason this is hard. The metrics contact centers rely on were built to measure human work. Handle time, transfer rate, containment. When AI takes the volume, those numbers keep reporting and quietly stop meaning what they used to mean.
Klarna is the cautionary version. The company cut customer service headcount on the strength of its automation numbers, then quietly rehired when quality slipped. The automation was real. The measurement was not good enough to catch what it was costing.
Three numbers that hold up under questioning: cost per resolved contact, repeat contact rate inside seven days, and the share of contained conversations where the customer came back on another channel.
Get those right and the AI case argues itself in front of a finance team. Get them wrong and you are defending a program with a figure a CFO can take apart in a single meeting.
This is why we start engagements with the reporting rather than the tooling. Before anyone talks about a platform, we look at how performance is measured today, which numbers people actually trust, and where the dashboards disagree. We call that starting point a CX Performance Blueprint, and most of what it finds lives in the definitions.
The pressure is only going in one direction. AI vendors now ship their own scorecards, and every one of them reports favorably on itself. Boards have moved from asking about activity to asking about return. And the organizations pulling ahead are not the ones deploying the most AI. They are the ones who built the visibility first.
If someone asked you tomorrow to prove your AI is working, which number would you reach for?
— Damian
Here’s what went down this week.
Bleeding Edge
Early signals you should keep on your radar.
Allstate is building Allie, an eight-component agentic platform in which a single component handles every customer interaction. CEO Tom Wilson described agent-to-agent processing layered onto 250 analytical models and hundreds of millions of interactions, with components built for reuse across the enterprise. No launch date was given. The signal worth watching is architectural: a carrier that size treating the customer layer as something to build rather than buy suggests the reference architecture for large CX estates may be shifting toward orchestration owned in-house.
Target named Chandhu Nair as its first chief AI officer, hiring him out of Lowe’s where he ran the AI transformation office, and appointed Purvi Shah as SVP of user experience on the same day. Nair starts August 24. Two roles announced together is the tell: Target is treating customer experience as an engineering function with an owner , a budget and a reporting line, rather than a program living inside marketing. Expect more retailers to copy the org chart before they copy the technology.
Leading Edge
Platform changes live now that need a decision this week.
Genesys Cloud removed AI Guides (version 1) in its August 3 release, the second deprecation to land on contact center teams in as many weeks after native Enhanced TTS support for selected Google and Microsoft voices ended on August 5. Anything built on Guides v1 has to be rebuilt rather than migrated. If nobody has inventoried which flows depend on v1 guides and which still call the retired voices, those two audits belong in the same sprint, because the remediation work overlaps.
Amazon Connect shipped the UpdateContactTaskTemplate API on August 7, letting supervisors and developers swap the task template on a contact already in progress instead of cancelling and recreating the task. That reads like plumbing until you count how many workflows currently handle a mid-task change by closing the contact, which breaks handle-time attribution and leaves orphaned records in reporting. Teams running task-heavy back-office queues on Connect should look at this before the next reporting cycle closes.
Future You’s Problem
Handle it now, or explain it later.
California’s AB 1609 went to the Senate suspense file on August 3, which parks the bill on cost grounds rather than killing it. The Right to Human Customer Service Act would require businesses above $500 million in revenue to make a good-faith effort to connect a customer with a human within fifteen minutes, prohibit representing a chatbot as human, and carry penalties of $5,000 for a first violation and $10,000 for each one after. Disclosure rules are one thing; a statutory time-to-human is an operational SLA written into law. Containment strategies that work by making the human path hard to find would need rebuilding, and that is architecture, not copy.
AI-attributed layoffs reached 205,000 US workers through August, already matching the full 2025 total, with cuts concentrated in customer service, data operations and finance back offices. Gartner’s standing forecast is that half the companies cutting service headcount on AI grounds will rehire for the same functions by 2027, under different titles. Both of those can be true at once, and the gap between them is the risk. Cutting before you can evidence sustained containment and CSAT means budgeting twice for the same capacity.
