AI agents have quickly become one of the most valuable applications of AI in customer operations. Sierra, Decagon, Cresta, ElevenLabs and others are building genuinely strong products, and their growth reflects the real value enterprises are seeing from automation.
As adoption accelerates, though, CX leaders have started asking a harder question. One executive put it particularly well:
That is exactly the right question to be asking.
Containment tells us whether an interaction ended without a human handoff, and resolution tells us whether the AI believes it completed the task. Both are useful measures of how the technology performed. What neither one tells us is what happened to the customer afterward, or what happened to the business.
Did the customer's problem stay solved? Did they call back? Did the purchase go through? Did they stay, buy another product, or eventually come back after leaving? And did total cost to serve actually decline?
Those outcomes determine whether AI is creating business value, and measuring them requires looking well beyond the AI interaction itself.
The Customer Journey Continues After the Session Ends
Every technology in this stack sees one part of the customer journey. An AI-agent platform has deep visibility into the interaction it handled, including what the customer asked, how the agent responded, which workflow ran, and whether the session ended without requiring a person.
The customer, of course, keeps going. They may call eleven minutes later, open a ticket tomorrow, complete the purchase, dispute a charge, cancel a service, buy something else, or stay with you for another five years. Those events live across CRM, contact center, billing, digital and operational systems, and no individual platform holds the complete picture. Which is why a question as simple as whether the AI is working turns out to be surprisingly hard to answer.
That leads to a principle I feel strongly about. An enterprise cannot rely on its AI vendor as the sole arbiter of the business value that vendor is creating.
I want to be precise about what I mean, because this is not a criticism of Sierra or Decagon or anyone else in the category. They see their own interactions extremely well, better than their customers ever could. What they cannot see is the downstream journey where the business outcome actually shows up. And when a vendor's economics depend partly on how a successful resolution gets defined, that is one more reason to measure the outcome yourself.
Anyone who has managed a BPO relationship already knows the discipline. Trust the partner. Measure the business outcome independently.
Containment Is Not the Same as Resolution
Consider an AI interaction recorded as successfully contained. If the customer calls back two hours later and spends twenty minutes with a human agent resolving the same problem, the original interaction looks successful operationally and looks considerably worse economically. The company absorbed the AI cost and the human cost, and the customer spent additional effort getting to the same place.
Now consider the reverse. An AI agent recognizes a complicated cancellation request, understands the customer's history, and transfers to a skilled retention agent with the right context, who resolves the problem permanently and keeps the customer. Containment failed. The customer journey and the economics both succeeded.
That distinction matters because optimizing the wrong metric produces the wrong behavior. The objective is maximizing successful customer outcomes and lifetime value at the lowest appropriate cost to serve, and containment is a proxy for that at best.
Measure the Entire Economics of AI
Cost per successful resolution is total cost to serve divided by the cases actually resolved correctly and durably, without avoidable repeat contact. It tells you what the operation spends to produce an outcome that holds. Customer lifetime value captures the other side, through retention, cross-sell and win-back. What we are trying to do is maximize customer value while continuously reducing the total cost required to produce successful outcomes.
Once AI interactions are connected to CRM, billing, digital, operational and financial data, leaders can start asking considerably more valuable questions. Which AI interactions actually stay resolved. Which contact drivers generate repeat demand after apparent containment. Where automation materially reduces total cost to serve, and which AI experiences improve conversion. How the experience affects lifetime value through retention, cross-sell and win-back. Where a human outperforms AI economically.
And the one I find most valuable of all. Which customer problems should never have generated an interaction in the first place.
Imagine thousands of customers contacting an AI agent because a billing event posts later than they expect. The AI may resolve those interactions beautifully and post excellent containment numbers. The highest-return intervention available is probably not improving the AI at all. It is changing the billing workflow, rewriting a confirmation message, or notifying customers proactively before they have any reason to ask.
Every interaction you eliminate permanently reduces cost to serve, and it usually improves the customer experience at the same time. That is a very different economic outcome from automating the contact.
Decide Where AI Actually Belongs
Once you can measure the whole journey and its economics, the next question gets more interesting. What is the best way to handle each type of customer need? I think there are four answers.
This gives leaders a much better framework for allocating CX investment. A password reset is an obvious candidate for automation. A confusing recurring billing issue may be better eliminated altogether. A complicated technical problem often benefits from AI-assisted human resolution. And a high-value customer calling to cancel probably warrants an exceptional retention specialist. The goal here is being pro-outcome rather than taking a side in the AI-versus-human argument.
Connect Every Decision to Business Value
This is the point where AI measurement turns into capital allocation.
For every major contact driver, leaders should eventually be able to answer what it costs today, what share of it gets successfully resolved, how much repeat demand it generates, and what happens downstream to conversion, retention, cross-sell and win-back. From there they can see the resulting impact on customer lifetime value, and whether eliminating, automating, augmenting or keeping the interaction human produces the greatest economic return.
With that in hand, management can compare opportunities that otherwise look nothing alike. Automating one contact driver might save eight million dollars a year with no change in customer outcomes. Another automation might save three million in labor while destroying five million of expected lifetime value through higher churn. A human-assisted retention workflow may cost more per interaction and generate substantially more value through retention and cross-sell. And eliminating an unnecessary billing contact might improve the customer experience while permanently removing millions of dollars of cost.
Those are the economics a leadership team needs in order to decide where the next dollar of AI investment should go.
From Insight to Action
Measurement gets you to the starting line. Once the organization identifies an opportunity, something has to change. A prompt gets improved, a routing rule changes, knowledge gets updated, a policy is simplified, an agent gets coached, a workflow gets automated, or a confusing communication gets rewritten.
Suppose repeat contacts about promotional billing suddenly increase. The system identifies the increase and determines that most of it comes from customers whose promotional credit does not appear until their second billing cycle. It quantifies the incremental contact cost and the associated churn, then recommends explaining the timing proactively before the first bill arrives. The company changes the communication.
Then comes the question that matters most. Did it work? Did those contacts disappear, did repeat demand fall, did cost per successful resolution improve, did CSAT improve, did retention change? Now the organization knows something it did not know a month ago.
Build Organizational Memory
Run that loop long enough and you end up with something considerably more valuable than better reporting. You get a record of which interventions actually work.
Imagine knowing that the last three times billing confusion increased, proactive communication reduced contact volume by 35 to 50 percent. Or that moving a cancellation journey entirely to AI cut operating cost 28 percent and raised 90-day churn enough to make the whole intervention economically negative. Or that for high-value customers, human retention specialists generate enough incremental retention and cross-sell to more than justify their higher cost.
At that point the company has moved past collecting customer data and started accumulating institutional knowledge about how to improve the business. Every new intervention makes that knowledge base stronger.
Where Perch Fits
All of this requires an intelligence layer spanning the systems already doing their individual jobs well. AI-agent platforms should keep building great AI agents, CRM should manage customers, contact center platforms should route interactions, billing systems should process transactions, and QA should evaluate quality.
Perch becomes the independent intelligence layer across them. We bring interaction, CRM, operational, workforce, digital and financial data together around the customer journey and analyze it continuously, so management can keep answering the questions that decide where the operation goes next. What happened. Why it happened. What it did to cost per successful resolution and to lifetime value. Whether the interaction should be eliminated, automated, augmented or kept human. What to change, and whether the change worked. And over time, what the organization has learned from every prior intervention like it.
That is the difference between measuring AI activity and running a system that continuously allocates people, technology and capital toward better customer and business outcomes.
The Question I Would Ask
AI-agent platforms will keep producing better metrics. Containment, resolution, deflection, latency, cost per interaction and many more. All of them are useful.
I would add one question to every executive operating review. What happened to our customers and our economics as a result?
Did customers stay resolved? Did cost per successful resolution improve? Did they convert? Did they stay, expand the relationship, or come back after leaving? What happened to customer lifetime value? And can you distinguish between the interactions your AI should handle, the ones where humans create more value, and the ones your business should eliminate altogether?
Those answers are already sitting somewhere in your company's data. The companies that win with AI will be the ones that learn fastest where AI creates value, where humans create value, and where the best interaction is the one they removed entirely.
That is what we built Perch to help companies do. We can typically create the initial journey-level view in about three weeks using data a company already has. If you are working through what your AI agents are actually delivering, and where your next dollar of AI investment should go, I would be glad to show you what that looks like.
Amit Basak is CEO and Co-Founder of Perch Insights. This piece is part of Running on Perch, a series on what it looks like to operate a business with all of its data connected.