What does the first 30 days look like when deploying a new AI customer support agent?
Most teams can plan the first month as three blocks of work.- Days 1 to 7: knowledge and shadow mode. Connect your help content, approve the Articles the agent will answer from, and connect the agent to your helpdesk or widget with an Internal Comment rule, so your team reads every reply before a customer does.
- Days 8 to 14: actions and go-live. Connect the systems the agent needs (billing, CRM, orders), build Intent Rules for your highest-volume requests, and move main traffic to Direct Reply once spot checks pass. Fini is live in 14 days.
- Days 15 to 30: full volume. Add the remaining channels, work the Review queue, and track resolution rate and escalation reasons in Analytics. Fini is fully autonomous in 30 days.
The three milestones
Click through the milestones to see what is live, what your team does, and the exit check for each.
Day 1: Integration complete
The goal on day 1 is a working agent that answers common questions from your real content, running on a real channel, with your team able to see every reply before customers rely on it. No code is required.1
Create the agent
On Home, select or create the agent under Your bots and name it by function (
Support-AI, Billing-Agent). See Quickstart, Step 1 and Home.2
Connect your knowledge
In Knowledge → Sources, add your help center sitemap, individual URLs, files, or a connected store (Notion, Google Drive, Confluence, Zendesk Help Center). Set the Language training preference to match your content; use Auto-detect for multilingual knowledge bases.
3
Generate and approve Articles
Run bulk Magic Articles generation over your Sources, then work through the drafts, duplicates and conflicts in Review. Organize approved content into folders in Articles and assign folders to the agent. This is the standard first-day workflow described in the Knowledge overview.
4
Deploy in shadow mode
Connect the surface your customers already use: a helpdesk integration such as Zendesk or Intercom, or the Widget. For regulated industries, start with an Internal Comment rule in Reply Rules so the agent posts internal notes your team can read before anything reaches a customer.
5
Test in Inbox
Send a direct question, a paraphrased question and an out-of-scope question through the channel. Open each reply in Inbox and read the AI Steps trace. Fix gaps in knowledge, tone in Main Guidelines, and missing context with Attributes.
Day 14: Agentic workflows
Between day 1 and day 14 the agent moves from answering questions to resolving them: it connects to your backend systems (billing, CRM, internal tools) and starts taking real actions, such as processing refunds, updating accounts and pulling customer data.1
Connect customer context
Define User Attributes under API Setup to load each customer’s plan, status, billing date or orders. Enable Use in Rulebooks on the attributes your workflows branch on.
2
Connect Actions
Define Actions (shown as External Actions) for the operations the agent should perform: cancel a subscription, issue a refund, look up a transaction, update an address. An Action only runs when a Rulebook Tool node invokes it.
3
Build Intent Rules for your top workflows
In Rulebook → Intent Rules, use Generate with AI to describe a workflow or upload a workflow diagram, then refine the tree. Start with the two or three intents with the most volume. See Intent Rules and the Cancellation flow walkthrough, or the Use cases tab for refunds, disputes and account changes.
4
Set escalation and safety controls
Add business-specific escalation triggers to the Planning Prompt (Controlling when the agent escalates), configure Guardrails such as Banned terms, Confidential attributes and URL allowlist, and keep sensitive intents on Internal Comment or No Reply in Reply Rules. If you use the widget with a helpdesk, create the widget escalation Business Rule.
5
Build your first Test Suite cases
Open Test Suite and click Create test case to turn real conversations (or new ones you create in Inbox) into test cases, each with criteria for what should pass. Use Criteria groups for checks that many cases share, such as escalation behavior, and run them. Runs use recorded or mock Action responses, so check live end-to-end Actions in a sandbox environment or a controlled pilot. See Test Suite.
6
Switch to direct reply
When your Test Suite cases pass and Inbox spot checks meet your bar, move the agent’s main traffic from Internal Comment to Direct Reply in Reply Rules. Keep the stricter rules for sensitive intents in place.
Day 30: Full autonomy
By day 30 the agent runs at full volume and maintains itself: self-learning is active, knowledge gaps are detected automatically, and voice, chat and email are unified on the same agent. Your team’s job shifts from building to reviewing.1
Work the Review queue
Fini’s background AI drafts fixes for knowledge gaps and conflicts it finds in live conversations. Approve, edit or reject them in Review. These drafts never publish without a human.
2
Expand to every channel
Add the remaining surfaces on the same agent: native Email, Voice from Voice in the agent sidebar, the Help Center, or more helpdesk channels. Use Channel Prompts for per-channel differences. For voice, you can port an existing number, forward calls, or connect over SIP.
3
Gate every change with Test Suite
Run your test cases or collections on every meaningful change (a prompt edit, a knowledge update, a rule change) and before each release. Turn new failures into test cases from Inbox (the flask icon) as you find them.
4
Track resolution in Analytics
Run the weekly health check in Analytics: the KPI cards, the Resolution rate trend, and the Knowledge performance and Intent rule breakdowns. Read Escalation reasons to see what to fix, and use Knowledge performance to find high-volume, low-resolution slices and fix them. See How to measure resolution rate.
What can move the dates
The milestones assume your content and systems are ready to connect. In practice these are the things that most often shift a date:- Knowledge quality. A help center with outdated or contradictory pages produces more Review work on day 1. See How Fini handles messy or conflicting docs.
- API access. Actions need a reachable endpoint and credentials for each system. If an API isn’t available yet, the workflow can run as an escalation with full context until it is.
- Approval cycles. Security and compliance review on your side (DPA, BAA, questionnaires) usually runs in parallel. Start it early with the Security overview.
- Voice setup. Voice needs a phone number connected to Fini: you can port an existing number, forward calls, or connect over SIP. Plan the routing change with your telephony team.
The 90-Day Money-Back Guarantee measures performance over the first 90 days of a deployment, so the day 30 milestone sits well inside that window. Eligibility applies; see Money-back guarantee and the terms.
Related
Quickstart
The day 1 steps in detail, from a fresh workspace to a deployed agent.
How Fini works
What happens to each conversation once the agent is live.
Fini on top of your helpdesk
How Fini connects without a migration.
Test Suite
Regression checks that gate every change after go-live.