Case study
Talking News Federation: an AI Help Desk Where a Person Approves Every Reply
Last updated
The short answer
New Odyssey built the AI help desk the Talking News Federation uses for member support. Each inbound email becomes a ticket, is classified by intent and routed, and has a reply drafted from the organization's own knowledge before anyone opens it. A person approves every reply. Suspected spam is quarantined for review, and every AI decision is written to a tamper-evident audit trail. It is in production.
The situation
The Talking News Federation is a volunteer-run membership organization. Its member support arrives by email and is answered by its own people, who are not technical staff.
The help desk was built to take the sorting and the first draft off those people without taking the decision away from them, and to show afterwards exactly what the AI did.
What we built
| Part | What it does |
|---|---|
| Email to ticket | Every inbound email becomes a ticket, with replies kept in one thread |
| Routing | The AI classifies the intent of each message and routes it |
| Drafted replies | A reply is drafted from the organization's own knowledge before an agent opens the ticket |
| Human approval | Agents approve every reply before it is sent |
| Spam handling | Suspected spam is quarantined and reviewed by a person, not deleted |
| Member records | Contact lookup and activity logging in the organization's CiviCRM |
| Audit trail | A hash-chained, tamper-evident log of AI decisions that can be verified on demand |
| Data protection | Subject access search and export, erasure and scheduled retention |
| Access | Agent, reviewer and admin roles, with optional multi-factor authentication |
How it was built
- The AI prepares and a person decides. Nothing reaches a member without an agent approving it.
- It fits the tools already in use: the help desk reads from and writes to the existing CiviCRM.
- It was designed for non-technical users, so the AI's suggestions sit inside the normal ticket screen.
- It runs on AWS serverless services, so the cost follows the volume of mail.
What this shows
The same pattern applies to back-office work in any company: the agent does the sorting, the lookup and the first draft, the person keeps the decision, and the record shows what happened. It is how we build custom agents and how Pegasus agents work.
Frequently Asked Questions
- Is the help desk live?
- Yes. It is in production.
- Does the AI send replies on its own?
- No. It drafts them. An agent approves every reply before it is sent.
- Is this a Pegasus agent?
- No. It is a custom build for one organization. Pegasus agents are packaged for common workflows and are built on the same principle of human approval and a full audit trail.
- Are there published results?
- Not yet. The help desk records first-response time, resolution time and how often drafts are used, but no figures have been published.
Next step
Have a workflow like this?
Tell us about it and we will say plainly whether an agent fits.