Diagnose low agent-assist adoption by separating access, exposure, relevance, trust, and workflow friction rather than treating usage as a performance score. An agent cannot use a suggestion they never see, and declining an unnecessary suggestion can be the correct decision. Investigate the workflow before setting a usage target.
Level AI describes agent-facing guidance. Google Cloud documents proactive knowledge suggestions. Product availability does not establish that agents in your operation can find, understand, or use the assistance when it matters.
What should an adoption funnel measure?
Start with eligible agents and contacts, then check whether the feature was available, whether relevant assistance appeared, whether the agent inspected it, and whether it helped the task. Keep unavailable or unsupported cases outside a behavioral judgment, while still reporting them as coverage gaps.
| Stage | Diagnostic question | Possible next step |
|---|---|---|
| Access | Could the agent open the feature? | Resolve eligibility or configuration |
| Exposure | Did useful assistance appear? | Review delivery and triggers |
| Understanding | Did the agent know its purpose? | Offer task-specific practice |
| Trust | Could the agent verify the output? | Review evidence and correction paths |
| Workflow fit | Did use interrupt required work? | Repair interface friction |
These are diagnostic categories, not promised analytics features. If event telemetry is unavailable, use a small, transparent observation exercise and agent interviews rather than inferring behavior from incomplete counts.
How should agents participate in diagnosis?
Ask agents to walk through recent ordinary contacts, including one where they used assistance and one where they did not. Invite explanations before suggesting a cause. Distinguish unnecessary help, wrong help, delayed help, unfamiliar controls, and concern about how usage data will be interpreted.
Show what feedback reaches the implementation team and report what changed. A feedback mechanism without a visible response can leave recurring problems unresolved. Give experienced agents and new hires a voice; they may need different support for the same contact type.
When is coaching the right response?
Use coaching when the tool is available and appropriate but the agent needs practice applying it. Choose a concrete task, demonstrate the verification step, and let the agent explain when to ignore a suggestion. Do not ask a manager to coach around a known integration defect or misleading answer.
Measure the effect of the intervention on useful work, not raw clicks. Compare correction burden, confidence in the process, and reviewed outcomes. Keep usage observations separate from disciplinary conclusions unless the organization has an appropriate, transparent evaluation process.
How should you choose the next intervention?
Assign the issue to the owner best able to change it: training, knowledge, product configuration, or desktop integration. Recheck the original example after the change. Use relevance analysis for noisy prompts and desktop acceptance testing for workflow failures.
How should a team diagnose a low acceptance rate?
An illustrative pilot has 100 eligible opportunities. Guidance is displayed for 60; agents inspect 30; 12 suggestions are used. Reporting “12% adoption” alone hides whether the other 88 opportunities lacked access, useful guidance or a reason to use the tool. These counts are constructed to explain the funnel, not product results.
| Observed loss | First investigation | Candidate intervention |
|---|---|---|
| Eligible but no display | Integration delivery, trigger and access logs | Repair exposure before coaching usage. |
| Displayed but unread | Desktop placement, timing and current workload | Test visibility and interruption behavior. |
| Read but declined | Agent reason and independent relevance review | Repair content or accept a justified decline. |
| Used but later corrected | Source support and saved case outcome | Address trust and answer quality before expanding. |
In the example, usage among displayed suggestions is 12/60; usage among inspected suggestions is 12/30. Neither is an accuracy measure. Add correction burden and task outcomes, and let agents explain justified non-use. Keep the eligibility definition stable across pilot versions so improving the percentage cannot be achieved merely by excluding difficult contacts.
What are the limits of this guide?
Sources were checked October 20, 2026. No adoption study was conducted, and this guide proposes no universal adoption target. Event data, interviews, and observations each have limitations. Confirm the telemetry available in your implementation, explain the purpose of monitoring to agents, and avoid interpreting a missing click as evidence that someone failed to help a customer.