Threll.ai
Context and handovers

Context dies at the handover. A new layer is not the cheapest fix.

Two numbers sit side by side in the same Genesys report: 95 percent of consumers expect not to have to repeat themselves, and 48 percent of companies pass nothing at all from the virtual agent to the human. The industry's answer is a coordination layer. That is right for some – and an expensive answer to a question most companies can ask much earlier.

7 min read
A woman in a dark coat stands at a hotel reception counter after dark, seen from behind in three-quarter profile, her phone in one hand resting on the counter and a folded sheet of paper in the other, while a receptionist behind the desk talks on the phone with his back half turned; a brass lamp on the counter is the only light

Two numbers sit side by side in the same report, and one of them explains the other.

The Genesys State of Customer Experience report for 2026 landed in early September, built on responses from 5,811 consumers and 1,560 leaders across more than twenty countries. In it, 95 percent of consumers say they expect their information to carry across channels so they do not have to repeat themselves. And 48 percent of companies still pass no data at all from the virtual agent to the human who takes over.

Almost half are not passing too little. They are passing nothing.

Two days later, the same company presented its answer to that problem. The answer is a new layer.

What was launched

At Xperience 2026, Genesys presented four new parts of Genesys Cloud: an AI Control Plane, a component it calls Contextual Intelligence, and two named Navigator and Orchestrator.

The first two are available now. Contextual Intelligence is meant to act as a persistent enterprise memory, linking what happens in the conversation to the customer's identity, history and wider journey. The Control Plane is the governance part: where AI agents are allowed to act, what they are allowed to see, and when a human has to be brought in.

The other two are a roadmap. Navigator is meant to replace static menus and routing rules with a judgement about what the caller is actually trying to achieve, and is expected to reach general availability between 1 November 2026 and 31 January 2027. Orchestrator is meant to hold the state of the customer's case and coordinate the next step across AI agents, employees and enterprise systems. It is expected between 1 February and 30 April 2027.

The diagnosis is right, and it is well put. Alex Ball of Genesys describes the risk as the customer being left with the feeling of dealing with “five different companies” inside one brand, because each department automates its own workflow separately. That is exactly what happens.

It is the answer that is worth a closer look.

Context is not lost because the systems cannot talk to each other

The usual explanation for why the customer has to repeat herself is technical: the systems are not connected, the data sits in silos, the integration is missing.

Sometimes that is true. But 48 percent is too high a number for it to be the whole explanation. A voice agent that has already heard what the case is about is holding that information in plain text at the moment it transfers the call. Passing it on requires no integration with any back-office system. It can be written into a ticket, read aloud to the person taking over, or sit in a field on their screen.

The reason it does not is usually that nobody decided what should go there. The handover is one of the places in a setup where skipping it produces no error. The agent transfers, the call continues, the numbers look fine. The only person who notices is the customer – and she does not complain about it. She tells her story once more, slightly shorter.

A coordination layer is an answer to a number

A layer that remembers across systems is the right solution when a case can live in many places. That is the whole point of it.

But the number of places is not a law of nature. For a large bank running twelve AI agents from five vendors, it is a given quantity that has to be managed. For most companies it is a decision somebody makes – usually without noticing that they are making it.

That is why the order matters. The question of what the agent has to be connected to before it can go live is in practice the same question asked six months earlier: how many places should a customer case be able to live? Answer “everywhere”, and you have in the same breath ordered the need for something to bind those places together.

The boundary most companies actually have

It is worth being concrete about who we are talking about.

A company with twenty or two hundred employees rarely has twelve AI agents. It has one, on one number, and one boundary that matters: the one between the agent and the human who takes over.

And almost half pass nothing across it.

In practice it looks like this. The caller has spent forty seconds explaining that the invoice relates to an address she moved out of in May. The agent transfers. The human says “hi, how can I help?” Nothing has broken, no log records an error, nobody gets an alert. The customer has simply spent another forty seconds saying the same thing again.

That is not an orchestration problem. It is one field and a decision about what goes in it. What the human learns at the moment of transfer is one of six decisions that determine whether escalation works at all – and it is cheap to settle now and expensive to postpone.

The patience involved is also shorter than many assume. In the same report, 84 percent say they will give a virtual agent up to three attempts to resolve an issue. Fewer than 20 percent will give it more. That makes it an expensive place to spend one of them: on asking the customer to repeat what she has just said.

How many places can a customer case live with you today?

Fifteen minutes where we count the places together, and write down the four or five facts your human needs at the moment of transfer.

Book Meeting (opens in a new window)+47 24 12 66 30

15 minutes · no obligation · pick your own time

The strongest counterargument

There are four objections to the reasoning above, and none of them is weak.

The number is not yours to set. The customer emails on Monday, calls on Tuesday and writes in the chat on Thursday. None of those channels was chosen by you, and each one is a fresh boundary where context can fall out. You are free to decide that a case should live in one place – but the customer has not read that decision.

Consolidation is not a realistic first step. The billing system is the billing system. Genesys reports that only 31 percent of CX infrastructure is fully cloud-based on average, unchanged from 2025, and that 73 percent of leaders consider integration with the rest of the business systems critical. Asking somebody to tidy up their system landscape before automating anything at all is, in practice, asking them to wait three years.

Governance does not scale down to zero. As soon as more than one AI acts on the company's behalf, somebody has to be able to say what each of them may see and do. That need does not disappear because the company is small. It is the same question that arrives the moment you let an AI drive your phone line: which permissions it was granted at sign-in, and what requires a confirmation before it happens.

And the advice has a shelf life. “Keep the number low” works for as long as you have one agent. Genesys reports that 40 percent of CX organizations are already using agentic AI, and the entire point of the launch is that some of them are past that stage. Anyone who waits until they have five agents before thinking about coordination is thinking about it too late.

All four hold, and the last two hold hardest. The need to say what each AI may see and do arrives long before the fifth agent, and it does not go away because the company is small. A decision about the number is also worth little if it only holds until the next department buys its own tool without asking anyone. So the conclusion is not that nobody needs orchestration.

What follows

The conclusion is that orchestration is a consequence of a choice made earlier – and that three things hold whatever gets bought in 2027.

The handover record is yours to define. No vendor knows what your particular human needs to know in order not to start from scratch. It is usually four or five facts: what the case is about, what the agent has already confirmed, what it could not do, who the customer is if it knows, and why the call was transferred. That list does not exist in any product. You write it, and it is just as necessary inside a purchased coordination layer as it is in a field on a screen. Try it the simplest way first: have the agent read the five points aloud to whoever takes over, for two weeks, before anybody builds anything.

Measure what the layer is meant to fix, not what is easy to count. A coordination layer is bought to reduce how often the customer has to explain herself again. That is the number to follow. Containment and answer rates flatter the agent without saying anything about whether the case was resolved, and they look just as good the day after a customer has told her story three times.

Settle the number before you buy something to coordinate it. Not as an architecture exercise, but as one sentence in a meeting: how many places should a customer case be able to live here? If the answer is two, the rest of this is a conversation you can have in three years. If the answer is nine, it is worth knowing now, while it is still cheap to change.

Threll.ai builds voice agents in Norwegian, Swedish and Danish. A layer that remembers across five systems is a good answer to having five systems. It is not an answer to why there are five.

Frequently asked questions

From the products, for now, yes. From the question, no. The point is that the number of places a customer case can live is something you decide while you still have one agent – and that the decision is nearly free to make now and expensive to make over once you have five. Write the answer down in a single meeting, and let it govern what the next agent is allowed to own.

Four or five facts are enough: what the case is about, what the agent has already confirmed, what it could not do, who the customer is if it knows, and why the call was transferred. The last one is the one most often forgotten, and it is the one that sets the tone of the next thirty seconds. No vendor can write that list for you.

Sometimes. But the agent is already holding the important part in plain text at the moment it transfers, and writing that into a ticket or reading it aloud to the person taking over requires no connection to a back-office system. When almost half still do not do it, the reason is more often that nobody decided what should go there – skipping the handover produces no error message.

Count how often the customer has to explain herself again, not how many calls the agent handled alone. In practice: listen to ten transferred calls a month and note how often the human opens with a question the agent already has the answer to. Containment and answer rates do not move at all when this goes wrong, which is exactly why they are poor warning lights here.

Related Articles

Continue reading more articles

See all articles