After-Hours Intake Chatbot: Governance First, Technology Second

The situation

A mid-sized organisation providing health-adjacent services — including mental health referrals and crisis support links — received a significant volume of after-hours queries via its website. Most queries were about services and eligibility; a smaller number involved people in acute distress who needed to reach an on-call staff member urgently.

Out-of-hours staff coverage was expensive and impractical at scale. The organisation explored an AI chatbot to handle after-hours inquiries.

What they did

Rather than finding a chatbot first and then thinking about governance, the board set its risk appetite before any vendor conversations began.

The non-negotiable rules: the chatbot would answer frequently asked questions and provide information about services, but would not give clinical advice, would not advise on medication or symptoms, and would always make a clear human escalation path available. Every session would begin with a mandatory notice: "You are talking to an AI." Urgent cases — flagged by keywords and by users identifying themselves as in crisis — would be escalated immediately to the on-call staff member.

Vendor selection and data due diligence were conducted by management, with findings reported to the board before deployment.

What it cost

Approximately $3,000 per year for the vendor platform and integration with the on-call scheduling system.

What the board did

The board set the risk appetite before the build, not after. This meant the vendor brief was built around the board's risk parameters from the start, rather than management presenting a ready-built solution for retrospective approval.

The board also required a monthly incident log — any session resulting in an escalation or a user complaint would be reviewed at the next board meeting. This gave the board ongoing visibility into how the tool was actually performing in the field.

The outcome

After-hours query volume handled by the chatbot reduced the load on on-call staff by approximately 40 per cent. Genuine urgent escalations now reach on-call staff faster, because the system flags them proactively rather than waiting for a staff member to check a general inbox.

No serious incidents have occurred. The monthly incident log has captured two edge cases, both addressed with chatbot script updates.

Lessons for directors

• Setting the risk appetite before the build — not after — changes everything downstream. Vendors build to your brief; if your brief does not include governance requirements, they will not volunteer them.
• A mandatory "You are talking to an AI" notice at session start is a minimum standard for any consumer-facing AI tool. Do not deploy without it.
• A monthly incident log reviewed at board level is a lightweight mechanism that provides meaningful oversight — approximately ten minutes of board time per meeting.