Field notes · Voice & phone automation

Piloting an AI voice agent on an agency call line, scoped under the micro-purchase threshold

How a 311 line or a program's call queue can try an AI voice agent in weeks, with a scope small enough to buy directly, and what the pilot has to prove before anyone commits to more.

7 min readVonar AI engineering

Most agency call lines have the same shape. A handful of question types make up most of the volume: hours and locations, the status of a request, how to report something, what documents to bring. Behind them sits a queue of calls that genuinely need a person, and those calls wait because the simple ones are in front of them.

An AI voice agent is a good fit for the first group and a bad fit for the second. A pilot exists to find out where that line sits for your line, with your callers, on your phone system, before you buy anything at scale.

Why the micro-purchase threshold is the right size

A pilot that fits under the micro-purchase threshold can be bought directly by a cardholder, without a solicitation, in the time it takes to agree a scope. The current dollar value is set in FAR 2.101 and changes; your contracting office will know today's number. What matters is the discipline it imposes: the pilot has to be small, defined, and finished.

That discipline is useful on its own. A pilot that needs a six-month procurement will be designed to justify six months. A pilot that has to fit in a few weeks will be designed to answer one question: does this reduce the queue without making callers worse off?

What a scoped pilot looks like

The pilot we recommend has five parts, and every one of them produces a written record.

  • Call review. A sample of recent calls, reviewed and tagged by intent. This is where the list of question types comes from, and the share of volume each one carries. It takes days, not weeks, and the tagged list is a deliverable you keep.
  • A narrow agent. The agent answers only the top intents, in plain language approved by the program, and transfers everything else to a person with a short summary of what the caller already said. It does not try to be clever at the edges.
  • A real line, a small slice. The agent takes a defined share of calls (after hours, or one overflow path, or one published number) so it meets real callers while a person is always reachable.
  • Measurement. Containment (calls resolved without a transfer), transfer rate, abandonment, time to a person for the calls that need one, and a caller satisfaction prompt at the end. Measured against the same weeks last year or the untouched share of the line.
  • A written result. What the agent handled, what it got wrong, what it cost to run, and a recommendation: expand, change, or stop.

What it has to run on

Agencies rarely get to choose their phone system for a pilot, so the agent has to meet the one in place: a SIP trunk, a carrier forwarding rule, or a queue in an existing contact-center platform. The pilot should be designed so nothing about the current system changes except one routing rule, and so that rule can be turned off in a minute.

Where the agent runs matters as much as what it does. Call audio and transcripts are records. The pilot should say where they are stored, for how long, who can read them, and that they stay inside the client's boundary or a government region of a cloud provider. Recording notices, TTY and relay paths, and language coverage are part of the scope, not an afterthought.

What the pilot is allowed to prove

Be honest about the two outcomes. A good pilot can show that a third of calls never needed to wait for a person, and that the people who did need one got there faster. It can also show that callers on this line ask things that need judgement, and that an agent should only take the after-hours message. Both are worth knowing, and the second one is worth knowing for the price of a pilot rather than a program.

If the result is expand, the pilot's own records become the requirement for the next step: the intents, the scripts, the transfer rules, the measurements. That is the point of doing it small first.

Working with us

We build voice agents that resolve calls and route work, deployed inside the client boundary with an audit trail of every call. A pilot on the shape above is a few weeks of work with a fixed price. If you would like to talk through your line, tell us what you're working on and an engineer will reply within one business day.