Choose a use case that suits an agent
Good first use cases are repetitive, well defined, and low in risk: order status, appointment changes, password and access questions, simple eligibility questions, or internal help desk requests. They have clear correct answers and a straightforward way to hand off to a person when needed. Avoid starting with high-stakes decisions, ambiguous requests, or topics where policy is unclear. Choose one use case, define success in measurable terms, and expand once it works.
- High volume, low complexity, and a known right answer.
- A clear definition of when the agent must hand off to a person.
- A baseline measure, such as handling time or resolution rate, to compare against.
Knowledge: the content an agent will quote
An agent is only as good as the knowledge it can draw on. Review the articles it will use: are they current, consistent, and written clearly enough that a customer could act on them? Remove contradictions and outdated procedures, and assign owners with review dates. If the knowledge base is a collection of long documents nobody maintains, expect the agent to reflect that. Cleaning knowledge is usually the highest-value readiness task.
Data: what the agent can see and act on
Agents need accurate customer and case data to personalize answers and take actions. Check the quality of the records they will read: duplicates, missing fields, and stale statuses lead to wrong answers. Define precisely which data the agent may access and which actions it may take, such as looking up an order, updating an address, or creating a case. Each permission should be deliberate and as narrow as possible.
Data 360 for unified customer data →Guardrails, security and trust
Decide what the agent must never do: give legal or medical advice, promise refunds beyond policy, reveal other customers' information, or discuss topics outside its scope. Configure instructions and restrictions accordingly, and review how personal information is handled and logged. Involve security, legal, and privacy teams early, since their questions are easier to answer in design than after launch.
Testing, handoff and monitoring
Test with realistic and adversarial conversations before customers see the agent: ordinary questions, confusing ones, off-topic requests, and attempts to push it beyond its limits. Design a smooth handoff to a person that passes the conversation history, so customers do not repeat themselves. After launch, review transcripts regularly, track resolution and handoff rates, and collect feedback to refine instructions and knowledge.
Implementation and integration services →Decisions to settle before you build
- Scope. Define the topics the agent handles and the ones it declines.
- Escalation rules. Specify what triggers handoff and who receives it.
- Ownership. Name a business owner responsible for the agent's quality and a technical owner for its configuration.
- Measures. Choose the success measures and the review cadence before launch.
A realistic first 90 days
- Days 1 to 30. Select the use case, clean the relevant knowledge, review data quality, and define scope and guardrails.
- Days 31 to 60. Configure and test the agent with real and adversarial conversations, and build the handoff.
- Days 61 to 90. Launch to a limited audience, review transcripts weekly, and expand when quality measures are met.
Pitfalls to avoid
- Launching on messy knowledge. The agent will repeat whatever is in the articles, including errors.
- Broad permissions. Give the agent only the access it needs for the use case.
- No human handoff. Customers who cannot reach a person quickly lose trust in the whole service.
- Treating launch as the finish. Agents need continuous review and improvement based on real conversations.

