THE KEY ANSWER
Choose a repeatable agent task: classification, finding a procedure, summarizing, or drafting a response. Ensure up-to-date sources and handover to a human. Measure problem resolution and repeat contacts, not just response speed.
Where does the agent really lose time?
Observe several typical cases from start to finish. How much time is spent understanding the problem, how much searching for data, and how much waiting for another department? A bot responding immediately will not help if the source of delay is the lack of current order status. First, find the part that can be improved without creating a new layer of confusion.
Demonstration example: before responding, the agent reviews a long thread and three systems. The assistant can prepare a summary, point out missing data, and attach the correct procedure. The human still leads the conversation. This scope allows assessing quality before delivering an automated response directly to customers.
Context and references: Google PAIR: People + AI Guidebook
Knowledge must have an owner
Collect valid answers and exception rules. Determine who updates information about returns, deliveries, and products. Contradictory materials should be resolved. The assistant should not randomly choose between an old instruction and a newer notice in another folder.
Separate general information from specific case data. A question about return conditions requires a procedure, while a question about status requires access to the correct order. The model should not guess status based on typical delivery time. If the integration does not respond, the user should see this situation and an alternative step.
Handover should save repetition
When human help is needed, pass the problem description, collected data, steps taken, and the reason for escalation. The client should not have to tell everything from the beginning. The agent should distinguish the client's statement from the system-generated summary.
Design boundaries: atypical, emotional cases, or those requiring a decision or change of terms. Do not try to keep every ticket in the automated channel. A transition to a human that is too difficult may improve the apparent "self-service" metric while worsening the experience and increasing subsequent contacts.
How to assess the benefit for the client and the team?
Measure time to resolution, response accuracy, repeat tickets, and draft review time. A fast first response is useful but may hide a lack of resolution. Compare similar case types and check whether harder cases fall outside the measurement.
In the pilot, collect reasons for correcting or rejecting suggestions. If the problem concerns tone, improve communication. If it concerns outdated data, fix the source. Implementation should reduce the effort of both parties. The agent must not become the person who endlessly cleans up errors of a tool that supposedly speeds up the department.
WHERE TO START
Bring this into your project.
- Choose a specific part of the agent's work.
- Establish the owner of procedures and updates.
- Pass context during escalation.
- Measure resolution, repeat contacts, and review time.
Choose one thing your process is missing today. It's a useful topic for your first conversation with the team.
QUESTIONS AND ANSWERS
Frequently asked questions.
Is it worth launching a public chatbot immediately?
Not always. An internal assistant allows identifying errors and value in a controlled scope. A public channel requires separate assessment of sources, client expectations, and handling of atypical situations.
Is the percentage of cases without a human the most important?
Not on its own. A high result may stem from difficult escalation. Check the quality of the resolution, repeat contact, and whether the client actually achieved their goal.
Sources and context
- Google PAIR: People + AI Guidebook ↗
The Google PAIR guide concerns the design of human-AI relationships. The agent work scenario and the set of metrics are a proposal by ALGOV.
Prepared by the ALGOV team. Current as of September 8, 2026. Examples describe possible scenarios, not results from client projects. How we create our guides.
