Cheap answers, expensive problems: measuring AI support across repeat contacts

AI support can make one answer cheaper while making service more expensive overall. A customer receives a fast response, remains stuck and returns through another channel. When a company measures each conversation separately, the return looks like new demand while the first interaction remains formally successful.

October announcements support considering a different economic unit: a resolved customer problem. Twilio released its Video Media SDK in public beta on October 15, providing backend access to conversation media tracks. Thought Industries announced Lens in beta on October 20, governing knowledge across sources. Zenarate announced its resolution-focused positioning on October 8. Google announced meeting recaps in Chat on October 19, including decisions and assigned actions.

These products are not a ready-made shared system. The editorial inference connecting them is that value emerges when observation, knowledge and completed work are linked. An additional channel earns its place by shortening the route to an outcome, rather than producing another record.

Contacts are easy to count, but customers need outcomes

Short handling time and the absence of a handoff can be useful operational signals. Neither establishes that the customer obtained access, received the delivery or got a fault fixed. A more appropriate unit links contacts by their cause and an observable result.

That does not mean merging everything one person says into an endless case. The unit needs boundaries: a product, order or device, an issue reason, expected action and a repeat-contact window. Someone can change a delivery address and ask about a new plan on the same day. Without separating causes, the system labels ordinary new work as a failure to resolve the earlier issue.

Zenarate's announcement is useful as a statement of the problem: it proposes reading related contacts as a resolution journey. It does not establish the performance of an individual deployment. A team can begin with a few common issue types whose outcomes can be checked independently of a conversation score.

A more expensive channel can reduce total cost

Video may cost more than a short chat while making a fault or incorrect setting visible. Twilio's server-side media access makes that context available to an application. The SDK is in beta; recognition quality and suitability need separate testing. Access to the video does not establish the cause of a fault or replace a specialist.

The economics depend on work the channel removes. If someone spends ten minutes describing equipment, a brief visual check may prevent another call and the wrong replacement part. If a precise help article resolves the issue, a mandatory video call adds expense and friction instead.

Tie the channel choice to missing evidence. An operator may need to see a particular setting, confirm a change or inspect an item. Asking for a camera feed just in case lacks that connection and creates additional data processing without a defined purpose.

New knowledge has to change the next action

Lens describes conflict detection and governance across existing knowledge sources. That is potentially useful when teams give incompatible instructions. Fixing a document does not itself close outstanding customer tasks, however. There must be a route from the identified cause to an owned action.

Google's announcement concerns workplace meetings, rather than a packaged customer-support module. Its adjacent principle is useful: a recap includes a decision and an owner for the next step. In service operations, an equivalent record can link a cause to a product fix, refund, revised instruction or follow-up contact. A deadline and completion signal should not disappear inside a conversational summary.

This also changes the investment question. A recurring fault may justify improving the product rather than training an assistant to explain it more smoothly. When related contacts are visible together, the team can compare the cost of resolving an underlying cause with the cost of repeatedly handling its symptoms.

A hypothetical cost comparison

Assume an AI interaction costs one unit and a specialist conversation costs eight. These are hypothetical assumptions, not platform prices. In the first path, a customer completes three unsuccessful chats and then reaches a specialist. Direct interaction cost is eleven units, before frustration and internal handoffs.

In the second path, the initial chat recognizes that selecting a replacement part requires visual inspection and routes the issue to a specialist. If the eight-unit conversation completes the job, the total is nine. Earlier escalation makes the autonomy metric worse while lowering cost per resolved problem. If the inspection does not help, that advantage disappears; the pilot should reveal this too.

Observe the cost of all related contacts, recurrence for the same reason and confirmed completion. Comparing a model's price with an operator's wage misses knowledge retrieval, work in adjacent teams and the consequences of a wrong decision.

Start with one recurrent problem and a small manually labelled set of contact chains. A service handling rare, straightforward questions may not need new multimodal infrastructure. Planning a customer portal helps define visible self-service outcomes. The service operations project provides a relevant setting involving connected calls, CRM and work handoffs. To evaluate AI economics in your service, describe a recurring customer problem and how its resolution is currently confirmed.