How We Built a 24/7 Voice Support Agent for a B2B SaaS Platform

Nikhai Jaysen · September 13, 2026

A B2B SaaS platform was fielding 70+ support calls per week — most of them the same 8 questions. We built a voice AI agent to handle tier-1 calls end-to-end. Here's the build, the design decisions, and what changed in the first week.

The Support Queue That Didn't Need Humans

A B2B SaaS platform — subscription analytics, around 200 business customers — was fielding between 60 and 80 inbound support calls per week. Their support team was two people. After three months of call logs, both people knew that roughly 70 percent of those calls were the same eight questions: how to connect a specific integration, why a report showed different numbers than expected, how to export data in a particular format, how to add or remove a team member.

These weren't difficult questions. Every one of them had a documented answer in the help centre. But customers called because documentation scanning is slower than asking someone, and the support team answered because that was the job. The problem was what the calls were displacing: escalations, account health reviews, the calls where a customer was one bad experience away from churning. Those were getting pushed to the back of the queue or handled at end of day when the team was already depleted.

What they needed wasn't more headcount. It was a filter — something that could handle the repeatable calls end-to-end, without a human on the line, and hand off cleanly when the call needed one.

What We Built and Why

The design decision we made early: this voice AI agent would not try to handle everything. It would handle the top eight call types with full resolution — walking the caller through the exact steps in real time — and escalate immediately for anything outside that scope. A clean hand-off, not a dead end.

The agent opens by identifying the account. If the caller's phone number is on file, the agent greets them by name and confirms the account. If not, it asks for an account ID — one step, spoken, no menus. From there it asks a single open question: "What can I help you with today?" The response routes the call into the appropriate flow.

For integration questions, the agent pulls live account data before speaking. It already knows which integrations are active and which are pending, so instead of reading generic documentation, it says: "I can see your HubSpot integration is pending authorisation. I'll walk you through the connection now — you'll need to be in your HubSpot account as we go." The caller is guided step by step, with the agent waiting at each prompt and confirming progress before moving on.

For anything outside the eight flows — a billing dispute, a data discrepancy that required investigation, a feature question the agent couldn't answer with certainty — the agent says: "I'm going to connect you to a team member who can look at this properly. One moment." Transfer happens within two seconds. No hold music loop, no re-explaining the problem. The transcript of the call so far goes to the human agent before the call connects.

What the Build Actually Required

The voice AI infrastructure was not the hard part. The hard part was everything that had to happen before a line of agent code was written.

We listened to three weeks of recorded calls — with permission — and built an accurate taxonomy of call types, including the edge cases and the ways customers described the same problem differently. "My report looks wrong" and "the numbers don't match" are the same issue, but the conversation that follows is different depending on which integration is the source. The agent had to recognise both phrasings and route correctly.

Script writing for voice AI is different from writing for chatbots. On a call, silence is a response. We wrote scripts for the success path, the error path, and the confusion path — what the agent says when a caller goes quiet, when they say something unexpected, when they confirm incorrectly and need to backtrack. These scripts went through eight rounds of revision before they sounded like something a person would actually want to hear at 7pm on a Tuesday.

The CRM integration was non-negotiable. Every call — resolved by the agent or escalated — created a support ticket with the full transcript, the identified issue type, and the outcome. The team could see exactly what the agent was handling and how, which also gave them data to expand the agent's scope over time.

What the First Week Looked Like

The support team fielded 17 calls in the first full week the agent was live. Their previous weekly average was 74. The 17 were all genuine escalations — account issues requiring investigation, a billing question, one caller who was upset and specifically asked to speak to a person. The agent identified each of these correctly and transferred without the caller having to repeat themselves.

Calls started being answered within one ring around the clock. Customers who had been leaving voicemails late in the evening — a common pattern the team had accepted as normal — were now getting resolved on those calls. One customer called on a Sunday afternoon to walk through an integration they had been stuck on for two weeks. The agent resolved it in eight minutes. The customer sent a message to their account manager the following morning.

The support team's description of the change: the first week where they felt they were doing actual support work rather than reading documentation out loud.

The Part That Takes Longest

Every week your support team spends answering the same eight questions is a week they aren't spending on the calls that matter for retention. The voice AI agent doesn't get tired, doesn't need a script in front of it, and doesn't make a caller wait until Monday. But it also doesn't build itself — the call taxonomy, the script revision cycles, the account data integration — that work has to be done before the agent goes live, and it has to be done properly.

We've built this for SaaS platforms ranging from early stage to several thousand customers. The setup is the same. The call mix is always different. If your team is spending more than half their support calls on questions with known answers, get in touch — we'll map your specific call types and show you exactly what the tier-1 filter looks like for your platform.

Frequently Asked Questions

Can a voice AI agent handle SaaS support calls end-to-end?

Yes, for a defined scope. This agent fully resolves the top eight call types — integration setup, report discrepancies, data exports, adding team members — by pulling live account data and walking the caller through the steps. Anything outside that scope transfers to a human within two seconds, with the transcript sent ahead.

How much support call volume did the voice agent remove?

Weekly calls reaching the two-person support team fell from an average of 74 to 17 in the first full week. All 17 were genuine escalations — account issues needing investigation, a billing question, one caller who asked for a person — and each was transferred without the caller repeating themselves.

What is the hardest part of building a voice support agent?

Not the voice infrastructure. It is the call taxonomy and the scripts. We listened to three weeks of recorded calls to map how customers describe the same problem differently, then wrote scripts for the success path, the error path and the confusion path. Those went through eight rounds of revision before launch.

Does the voice agent log calls anywhere?

Every call — resolved or escalated — creates a support ticket with the full transcript, the identified issue type and the outcome. That gives the team visibility into what the agent handles and the data to expand its scope over time.

Related Service

Voice AI Agents

Handle inbound calls, qualify prospects, and follow up with leads — using natural-sounding AI voice agents that never miss a call.