That finding points to the engineering task ahead: complete the customer’s request and make a difficult case easier for a person to resolve. The useful unit of service is the finished task, tracked across channels. Conversation length tells us much less.
Field note: the chat window hides several services
Salesforce’s Customer Zero account describes its own use of AI alongside data preparation and human collaboration. A policy question, a disputed bill and a repair booking can all begin in the same chat window, yet they require different operating capabilities.
The distinction between knowledge and permission is decisive. An agent can know a refund policy while lacking authority to issue the refund. Giving it technical access does not settle eligibility. The system needs both the relevant evidence and an enforceable permission to act.
An unresolved request must stay visible even if the conversation has ended. Connecting the service record to subsequent calls and corrections lets the team see where apparent automation simply displaced work. That is where engineering attention belongs.
In the EU, Article 50 of the AI Act has applied since 2 August 2026, requiring providers of directly interactive AI systems to make the AI interaction clear to users unless it is already obvious. That disclosure belongs in the service design alongside the route to a person.
Field note: completion needs a receipt
Choose a request with a clear beginning and a verifiable end. For a change of delivery appointment, completion means that the new slot is accepted by the operating system and communicated accurately to the customer. A confident message that the change is being arranged is only an intermediate state.
The agent needs to distinguish a pending transaction from a confirmed one. If a dependency fails or returns an uncertain result, it should check the current state before retrying. Otherwise an attempt to be helpful can create duplicate bookings or contradictory communications.
Measure the whole request across channels. Include repeat contact and correction effort alongside completion time. Where the service has different customer groups or languages, check whether the apparent average improvement conceals a worse experience for some of them.
Field note: transfer the work already done
Escalation should carry the customer's objective, the evidence already gathered and the actions attempted. An adviser should be able to see which parts are confirmed and which remain uncertain. That reduces the need to begin the conversation again.
The receiving team also needs capacity and authority. A transfer button cannot compensate for an understaffed queue or unclear ownership between departments. Design the human part of the service with the same care as the automated part, including what happens outside normal operating hours.
This changes the adviser’s working day. A difficult case arrives with its history already assembled, the customer’s desired outcome and the precise decision that needs human authority. The adviser can begin with judgement instead of asking the customer to repeat the story.
Fix the interruption before the customer asks
Begin by asking the agent to prepare a response or transaction for review. Evaluate it against real request patterns, including missing information and contradictory records. Record why advisers reject or change proposals; those decisions reveal weaknesses that a simple answer score may miss.
Where performance supports wider authority, define the permitted action precisely. Set limits according to the type of request and the state of the account. Keep the controls in the service systems, with a record of what happened and a practical way to correct it.
Now extend that design to an interrupted delivery. The operating system identifies the missed connection, checks available alternatives and reserves an acceptable replacement within the customer’s agreed preferences. The customer receives a confirmed arrangement, with a direct route to change it. The service has removed the reason for the call.
We expect this kind of proactive resolution to become the more valuable ambition. The engineering connects event detection to permitted action and an observable result. Where approval is needed, the service prepares the choice and waits; it does not send a reassuring message while the underlying job remains undone.
The service team can then study prevented disruption alongside resolved requests. Repeated delivery changes may point to a scheduling problem that belongs with operations. A well-designed agent brings that pattern into view, allowing the organisation to improve the service itself rather than make apologies faster.
Start with one reason customers call back
This quarter, follow one common request from first contact through repeat calls and final completion. Agree the evidence of resolution before setting the next automation target.
A NectarGlobal AI opportunity evaluation can turn that request into a service design with a clear completion test and a useful human handover.