Agentic AI sits on almost every CX roadmap, and any vendor can demo an agent that reschedules, refunds and apologises in twenty minutes flat. The comfortable story is that real projects stall because organisations are timid. As we discussed in our recent CRMKonvo with Carsten Ratzlaff, founder of Ratzlaff CX & AI Advisory, who spent almost twenty years at Oracle, the slow project is rarely the problem. The fast one is. Agents do not stall on models. They stall on decisions nobody in the organization ever had to make.
TL;DR
If you want to watch the full CRMKonvo, please go ahead here (optimized for smartphones) or here (optimized for tablets/computers).
Else, be my guest and continue to read.
Or do both …
Vendors Sell Capability; Customers Have to Grant Permission
Carsten spent his last Oracle years as director of business development and strategy for CX in DACH and Europe, so he has watched these deals from the selling side. His summary is blunt: vendors sell what an agent can do, and customers have to decide what it is allowed to do. The demo runs on clean data, following the happy path. The customer's reality runs on permission and with less than clean data. Which data may the agent trust? Which may it access? Which action may it take? Who is responsible when it gets it wrong?
Add legacy systems and five to seven departments, each with its own definition of a customer, and the gap between demo and operation is less than trivial. Business life is mostly exceptions; the demo is built to avoid them. Let’s be clear about it: the demo answers a question nobody on the buying committee is actually asking.
Automating the Official Process Automates Your Worst Service
Where do the exceptions live? According to Carsten, mostly in the heads of a few experienced people. Ask three service leads how much goodwill an agent may grant, or when a complaint goes to a manager, and you get three answers, at least. The technology is ready to obey. Oracle now talks about an Enterprise Operating Envelope of goals, permissions and escalation limits. But before an agent can follow a rule, someone has to write it down, and someone has to own it.
His sharpest line hurts: "If you only automate the official process, you automate, from my point of view, the worst version of your service." It reminded me of a builder who once defended shoddy work as "it’s up to code". The official process is like the building code: the bare minimum one needs to do. Good service lives in the exceptions, in the colleague who has known the customer for fifteen years.
I pushed back on one point. Process mining records what actually happens in the system, exceptions included, which makes the reliance on any one expert's head less crucial. Carsten's counter was that few companies mine their processes yet; they automate what is already documented. Both statements are true.
That leaves an awkward question that we only brushed against when Carsten called the employee side "underrated". The people who know the exceptions are regularly the very people the efficiency business case is aimed at. Why would the fifteen-year veteran write down what makes them hard to replace, for a project whose RoI is measured in their headcount? Capturing exceptions is a knowledge transfer exercise, and knowledge transfer runs on trust.
A Stopped Pilot Is a Lesson; a Bad Live Agent Is a Liability
Carsten does not worry about slow projects. He worries about fast ones, and the Fivetran 2026 Agentic AI Readiness Index shows why. Nearly 60% of organizations invest millions in agentic AI, only 15% consider their data foundation fully prepared, although 41% already run agents in production, which basically says that many agents run on shaky foundations their own companies do not really trust.
A failed pilot is visible and contained: budget, time, a few bruised egos. A live agent that is confidently wrong does real damage in front of customers, and no dashboard reports it. Stopping, in Carsten's words, is good management as long as it rests on evidence rather than politics. His test for pulling the plug consists of three questions: what is the agent allowed to do, how do we know it is doing it right, and who can stop it? If you cannot answer all three, or your internal KPIs improve while customer outcomes get worse, something is very wrong.
On data, he rejects both extremes. Data is never ready, and waiting for perfect data means waiting forever. The useful question is narrower: which data does this specific agent need for this specific decision, and is it good enough? "Our data is not ready" is so broad that it lets everybody off the hook. But then: What does “good enough” mean?
Your Dashboard Measures the Soloist, Not the Orchestra
Handling time falls when the agent is quickly wrong. Self-service rates rise when the customer gives up. Most CX metrics were built for teams working cases, not for an orchestra of agents handing customers to each other. Gartner found that only about a quarter of service leaders show positive financial returns from their AI use cases. Carsten takes this as a steering problem, and not a measurement problem.
The root cause is that a cost project gets sold as a CX project and then measured only by cost. Savings from removing waiting and repetition are measurable. Savings from making a human harder to reach are a less so. His test to identify whether it’s a CX project is simple: is the customer's issue resolved faster, or only deflected faster?
However, this argument has a gap. Making customers happy at all cost is not a business model and needs a counterbalancing metric. What the right one is, is "difficult to say", however, it needs to be based on a customer centricity strategy. A strategy is not a metric. Carsten did name first-time resolution as the outcome to track. The counterweight he did not name could be cost per resolved issue rather than cost per contact. It punishes both the expensive human loop and the cheap deflection that returns as a second call.
You Can Buy the Instruments, Not the Conductor
To the orchestration vendors, Carsten's says: "You can buy the instruments, but you can't buy the conductor." Platforms deliver routing, agent-to-agent protocols and monitoring. Deciding which agents exist, what they may do, what scales and what gets stopped is a leadership job across marketing, sales, service and IT. A vendor will optimize its part, never your whole customer relationship. Without that conductor, agent silos simply replace application silos: every vendor ships its own agents, every department buys its own.
CRM, in his view, is not losing its job; it moves from the stage to the foundation. Agents do not need a user interface. They need trusted customer data, rules and a record of commitments. According to Carsten, our next CRM is an architecture rather than a system. IT owns the platform; the business owns the rules and exceptions. And as Ralf noted, an agent doing wrong things is a liability that should land with the CEO, not the CIO.
The tension sits right here. When Ralf asked whether leaders have actually built the skills for this, Carsten was candid: they may have become a little more efficient, but he is not sure they have improved significantly. The entire prescription rests on a conductor the guest himself is not sure exists. His lightning-round answer makes the stakes plain: the most dangerous failure is not bad data or excess autonomy, both fixable, but nobody being accountable. Orchestration is a management job wearing a product label.
Pragmatic Playbook for Enterprise CX Buyers
Carsten's Monday-morning advice was to stop collecting pilots, pick the one customer moment that hurts most and name one owner. There are three ways to hold vendors and yourself to that.
Make the demo run your exceptions, not their happy path. Bring several real cases out of your veterans' heads: goodwill above the usual threshold, a partial delivery, a customer with an open complaint. Ask the vendor to show the permissions, escalation thresholds and the handoff to a human with full context, configured live. If the handoff is hand-waved, so is the product.
Scope the data to the decision, then put names on it. Reject "our data is not ready" as an excuse in either direction. For each agent, list the data that one decision needs, validate it is good enough, and name a data owner and a rule owner before anything is built. Demand an auditable record of every action the agent takes against the customer record.
Write the kill switch before the go-live date. Answer Carsten's three questions in writing: what the agent may do, how you know it is right, who can stop it. Baseline first-time resolution, repeat contacts and cost per resolved issue before you build, and retire average handling time as a target. Keep a human review on any decision that goes against the customer, whether rejecting a claim, refusing credit or processing a cancellation, until the rules behind it are written down.
Your customer does not care how agentic your AI is. They care whether your promise is kept and whether anyone makes them say it twice.

Comments
Post a Comment