The handoff is where most AI support fails
Customers rarely complain that a bot could not help. They complain that they had to start again when it passed them on.
Deflection rate is the wrong measure of an AI support deployment. A system that resolves seven contacts in ten and mishandles the other three can score well on automation and still damage satisfaction, because the three are the ones people remember and talk about.
What actually goes wrong
In most deployments the AI and the routing engine are separate products. The AI holds the conversation; the ACD holds the queue. When the conversation crosses that boundary, whatever the AI knew has to be handed over an integration, and integrations transfer fields rather than understanding.
The transcript arrives as an attachment nobody reads before answering, or not at all The customer is placed at the back of a queue as though they had just arrived Routing uses skill tags and queue order, ignoring what the conversation revealed The agent opens with \"how can I help?\" to someone who has already explained twice
What a good handoff carries
Because the voice agent and the routing engine are one system in Dialog365, the escalation does not cross a vendor boundary and nothing has to be reconstructed.
| Carried across | Why it matters |
|---|---|
| Full transcript | The agent can see exactly what was said, not a summary of a summary |
| Written summary | The agent reads three lines, not three minutes of transcript, before speaking |
| Detected intent | Routing decides on what the customer wants, not the number they dialled |
| Sentiment trajectory | An escalating customer can be routed to someone equipped for it |
| Actions already taken | The agent does not repeat a lookup the AI already performed |
| Verified identity | Authentication done once, not twice |
Routing on outcome, not queue order
A conventional ACD answers \"who is free and has the right skill tag\". An AI-orchestrated engine can answer a more useful question: who has historically resolved this kind of contact, for this kind of customer, in this state. Skill tags are a proxy for that; outcome history is the thing itself.
The test worth running
If you are evaluating platforms, do not test the AI in isolation. Time an escalation end to end and watch two things: whether the agent has the transcript and summary on screen before they speak, and whether the customer has to repeat anything. Everything else is easier to fix than this.
A useful measure to instrument: repeat-explanation rate. What proportion of escalated contacts contain the customer restating their problem? In most deployments nobody measures it, and it is the number most closely tied to how the experience is remembered.
Frequently asked questions
What is an AI-orchestrated ACD?
A routing engine that decides using intent, sentiment, skill and past outcomes from the live conversation, rather than queue position and static skill tags.
Why do customers have to repeat themselves after a bot?
Because the AI and the routing system are usually separate products, so context has to cross an integration boundary and typically arrives as an unread attachment, if at all.
Should the AI always try to resolve before escalating?
No. Recognising early that a contact needs a person and routing it immediately with full context is a better outcome than several failed attempts first.
How do we measure handoff quality?
Repeat-explanation rate, time from escalation to first agent word, and whether the agent had the summary before speaking. All three are instrumentable.
See it on your own calls
Tell us how your contact centre runs today and we will map it to your setup.
Related
One platform for AI agents and human agents
Why integrating two vendors costs more than it saves.
How Dialog365 stops the AI inventing answers
Grounding, closed-domain retrieval and refusal behaviour.
Why latency decides whether people talk to your voice agent
Sub-second turn-taking and what breaks below it.
What AI conversations look like by industry
Banking, healthcare, retail, logistics and telecom.