A chatbot maintenance retainer should pay for accountable operation: reviewing real conversations, correcting source gaps, testing customer actions, reporting verified outcomes, and managing approved changes. Charging every month for an untouched embed is not a durable service.
Customer questions, websites, policies, staff, product catalogs, and calendars change. Even a strong launch configuration can become stale. The retainer defines who notices change, who approves it, and how the system is tested afterward.
What a monthly retainer should include
| Workstream | Monthly work | Evidence for the client |
|---|---|---|
| Conversation quality | Review a defined sample of unanswered, escalated, low-rated, and high-value sessions | Findings, examples, and approved actions |
| Knowledge health | Update approved pages or documents and resolve conflicting information | Change log with source owner |
| Search quality | Check failed searches, broken links, product cards, and stale public results | Test queries and corrected destinations |
| Conversion workflows | Test leads, bookings, dispatch alerts, and confirmations | Timestamped test records and recipients |
| Experience | Check mobile, dark mode, accessibility, close behavior, consent, and major browsers | Regression checklist |
| Operations | Review permissions, connectors, notification owners, and known incidents | Access review and incident notes |
| Reporting | Compare agreed metrics with baseline and explain material changes | Concise report with decisions, not vanity totals |
Use a risk-based review cadence
Not every client needs the same frequency. A stable brochure site may need a monthly review; a changing store, seasonal service, or high-volume support workflow may need weekly checks. High-risk workflows may require approvals, monitoring, and controls beyond a standard agency retainer.
Weekly or automated checks
- Notification delivery for critical lead, booking, or escalation routes.
- Availability and timezone behavior for active booking workflows.
- Broken installation, domain, authentication, or connector warnings.
- Sudden increases in errors, unresolved questions, or failed customer actions.
Monthly review
- Representative conversation sample and negative feedback themes.
- Top intents, search misses, source gaps, and repeated escalations.
- Lead and booking completion based on the agreed definitions.
- Approved content changes and regression tests.
- Named actions, owners, and due dates for the next period.
Quarterly governance review
- Business goal, scope, prohibited topics, and human escalation boundaries.
- User access, roles, connector permissions, credentials, and former staff.
- Privacy notice, retention expectations, consent, and data minimization.
- Whether the workflow still creates enough verified value to continue or expand.
The NIST AI Resource Center and the NIST Generative AI Profile are useful references for risk-aware monitoring and evaluation. Apply controls proportionately to the client's use case.
Report metrics the client can interpret
| Metric | Define it before reporting | Avoid |
|---|---|---|
| Conversations | What starts a session and which environments count | Presenting raw opens as customer outcomes |
| Answered requests | How the sample or classification determines a useful answer | Calling every AI reply a resolution |
| Qualified leads | Required fields and qualification rule | Counting test or incomplete submissions |
| Bookings | Completed confirmation and excluded cancellations/tests | Counting calendar views as appointments |
| Human takeover | What triggers it and whether it was handled | Treating all escalation as failure |
| Satisfaction | Collection method, response count, and period | Claiming broad satisfaction from a tiny sample |
Set boundaries before the first invoice
- Number of assistants, sites, languages, and workflows covered.
- Conversation sample size and reporting frequency.
- Included content updates, configuration changes, and meetings.
- Support channel, support hours, response target, and emergency definition.
- Excluded development, integrations, migrations, copywriting, and compliance advice.
- Usage charges, third-party costs, taxes, and approval for overages.
- Change-request process for new journeys, connectors, voice numbers, or client entities.
Price these obligations with the white-label chatbot pricing framework. Put the boundaries into the proposal and statement of work, then confirm the operating owners during client onboarding.
A simple monthly meeting agenda
- Review incidents, broken actions, and urgent customer-impacting issues first.
- Explain what customers asked and where the assistant lacked sufficient information.
- Review leads, bookings, human handoffs, and satisfaction using agreed definitions.
- Approve knowledge and workflow changes with an owner for each source.
- Choose no more than a few prioritized improvements and define their tests.
- Record decisions, exclusions, and any work that requires a separate estimate.
When to recommend expansion
Expand only when the current workflow is stable and evidence supports the next use case. A client might add voice after the website assistant handles the same intents reliably, or add a connector after the team has documented permissions and approval requirements. New complexity should not be used to hide an unresolved foundation.
Related agency resources
- How to start an AI chatbot agency
- AI chat for agencies and client sites
- Chirps documentation
- Chirps plans and usage
This article is general operational guidance. It does not create a service-level agreement and is not legal, security, accounting, or compliance advice.