AI voice agent for IT support: which calls it should take and which it shouldn't

Most writing about an AI voice agent for IT support starts with the technology. This article starts with the calls. A help desk phone line gets a small number of very different call types, and an AI agent is good at some of them and a liability on others. If you sort your calls first, the buying decision gets much easier.

This is a practical guide for a small IT team or MSP: how to sort your inbound calls, what to hand to an AI, what to keep for a human, and what the generic “AI receptionist” products cost when you apply them to an IT support call pattern.

Start by sorting your calls, not shopping for tools

Pull a week of voicemails, missed-call logs, or your own memory of the last ten calls. In a typical small IT shop they fall into five buckets:

1. Status calls. “Any update on my laptop ticket?” “Is the VPN issue fixed yet?”
2. New-issue intake. “My printer is offline.” “I can’t log in to email.”
3. Outage calls. Several people from one client calling about the same thing.
4. Urgent or sensitive calls. A server is down, someone suspects a breach, an executive is locked out before a meeting.
5. Noise. Sales calls, wrong numbers, vendors calling back.

Count them. For many small teams, status calls and routine intake are the majority of volume, and they are the cheapest to answer well. That is the target for automation.

What an AI voice agent can safely handle

Status calls

A status call is a lookup. The caller wants to know the state of a ticket that already exists, and the answer is sitting in your ticketing system. The only thing that makes this hard is whether the agent can see that system. An AI that doesn’t read ticket data can only say “someone will get back to you,” which is voicemail with a nicer voice. An AI that reads live ticket data can say “your ticket was assigned this morning and is in progress.” We cover how to test that claim in how AI answers support calls from ticket data.

Routine intake

Capturing a new issue is mostly structured questions: who are you, what company, what’s affected, since when, how many people. An agent can ask these consistently, every time, at 2 a.m., and write the answers into a ticket. A tired human on the fifth call of the morning skips questions. The agent doesn’t.

The standard to hold it to: the ticket it creates should be one a technician can start working without calling the person back to ask what they meant.

After-hours first response

If nobody answers after hours today, the baseline is voicemail. Any agent that captures a structured ticket and tells the caller what happens next beats voicemail, even before you consider anything clever.

What should stay with a human

Outage and security calls

When the calls are “server down” or “I clicked something and now my files are renamed,” the right outcome is a human technician quickly. The agent’s job is to recognize urgency keywords, capture the minimum details, and escalate, not to talk the caller through a diagnosis. Decide in advance which phrases trigger an immediate escalation to an on-call tech, and test them.

Anything involving identity

Password resets and account unlocks are the classic IT support call, and also the classic social-engineering target. A caller who sounds convincing on the phone is not verified. Don’t let an AI agent reset credentials on the strength of a voice. The safe pattern is intake only: the agent records the request, and a technician verifies the caller through your normal process before anything changes.

Judgment calls and relationship calls

Scope disputes, billing complaints, “I’m thinking of leaving” calls. These are for a person. A good agent identifies them and hands them off instead of defending your invoice.

A short checklist for evaluating any voice agent

Run these tests in a demo before you pay for anything:

• Call a ticket by number or by name. Does it report the real status, or deflect?
• Describe an outage. Does it escalate, and how fast does a human actually get notified?
• Ask for a password reset. It should refuse to act and log the request.
• Interrupt it and go off-script. Does it recover, or loop?
• Check what lands in the ticket. Read the ticket the call created. Is it usable?
• Ask what happens to the transcript. You want it attached to the ticket, not stranded in another tool.

For a longer buyer’s checklist, see how to choose an AI phone agent for an MSP help desk.

What generic AI receptionists cost on an IT call pattern

Many “AI phone agent” products are general-purpose receptionists built for dentists, law firms, and plumbers. They can answer a phone. Whether they can look up your tickets is a separate question we have not tested, so this section covers price only. The interesting part is how they bill, because the billing unit changes the math for an IT shop.

Smith.ai: priced per call

Per Smith.ai’s pricing page, as of 2026-10-07: Starter is $300/month with 30 calls included and $11.50 for each call over 30; Basic is $810/month for 90 calls and $10.50 per extra call; Pro is $2,100/month for 300 calls and $8.50 per extra call. Transfer destinations are capped per plan (1, 2, and 10 respectively).

Apply that to a small MSP taking about 120 inbound calls a month. On Starter that is $300 plus 90 extra calls at $11.50 ($1,035), for $1,335 in total. On Basic it is $810 plus 30 × $10.50 = $315, for $1,125. Pro is $2,100 flat. The cheapest of the three at that volume is Basic at about $1,125 a month. A per-call bill punishes exactly the thing you want, which is answering every call, including the short status checks.

GoodCall: priced per unique customer

GoodCall’s pricing page, as of 2026-10-07, bills per agent: Starter is $79/month ($66 billed annually) for 100 unique customers, then $0.50 per customer over that; Growth is $129/month for 250 unique customers; Scale is $249/month for 500. Minutes are unlimited. Starter is limited to 1 logic flow and 3 team members.

For the same shop, if we assume those 120 calls come from about 80 distinct callers (an assumption, not a measured figure), that fits inside Starter at $79. That is far cheaper than per-call billing. The catch for an IT shop is the “1 logic flow” limit on Starter: one flow has to cover status, intake, and escalation, and the plan with three flows is $129/month. Again, this is the price of answering the phone, not a statement about ticket integration, which we did not verify.

What the comparison shows

Per-call and per-customer billing are very different for the same call volume, and neither tells you what the call produces. The right comparison is cost per resolved or correctly ticketed call, and that depends on whether the agent is connected to the ticket system. A cheap receptionist that emails you a summary leaves you to retype it into a ticket.

Where ITSM fits

ITSM takes the other approach: the ticketing system and the AI phone agent are one product. Callers dial in and the AI answers using live ticket data, so status calls are lookups against the real ticket and new-issue calls become tickets directly, with no second tool to reconcile. Techs can take escalated calls in a browser softphone, with two-way SMS and email-to-ticket in the same place. For how the plumbing works, see how to set up a help desk that answers phone calls automatically.

You can try the ticketing side without a card: the free tier is 2 seats at $0 with instant signup at /register.

A rollout plan that limits risk

1. Week 1: after hours only. Replace voicemail, leave business hours with humans. Read every ticket the agent creates.
2. Week 2: overflow. Route calls to the agent only when all techs are busy or the line has rung several times.
3. Week 3 onward: status calls by default. Keep a one-key path to a human for anyone who asks.
4. Always: keep your escalation phrases and your identity rule (no resets by voice) written down and tested monthly.

The short version

An AI voice agent for IT support earns its keep on status checks, structured intake, and after-hours coverage. It should escalate outages and refuse identity actions. When you price options, ask what the billing unit is and what the call turns into afterward, because a transcript in someone else’s dashboard is not a ticket.

ITSM is a small product, and there is a free tier, so you can judge it for yourself.