After-Hours Triage Automation for a Consulting Firm

The Challenge

A mid-sized professional services firm — specialists in their domain, with clients across multiple time zones — had built their reputation on responsiveness during business hours and lost ground every night when the calls kept coming and nobody was at the desks.

The problem wasn't the volume. The problem was the texture of the volume. Most inbound after-hours calls fell into one of three categories: a routine question that didn't need a human, a status request the firm could answer with information that already existed, or a genuinely urgent issue that needed a partner woken up. The firm had no way to triage these in real time, so every inbound call sat in a generic voicemail inbox until morning. Routine questions got cold. Urgent issues went unanswered. Status requests cluttered the queue and obscured the few calls that actually warranted immediate attention.

They had tried generic answering services, but those couldn't speak to the firm's domain or qualify callers with the precision that mattered. They had tried IVR menus, but clients hated them and important callers hung up. They had tried after-hours rotations, but the load was unsustainable and the team's morale was eroding.

The Approach

We built an AI-powered voice and triage layer that handles every after-hours inbound call live, answers what it can, qualifies what it can't, and escalates only what genuinely needs a human at 9pm on a Tuesday.

Voice handling. Inbound calls route to a Twilio number that connects to a state-of-the-art voice agent. The agent runs on Claude Opus 4.6 for reasoning and Vapi for orchestration, with OpenAI Realtime API as a fallback path. It has been trained on the firm's domain, the structure of common client questions, the firm's standard answers, and the explicit boundaries of what it can and cannot say authoritatively.

First-line triage. Every call is classified in real time: routine question (answered live by the agent), status inquiry (looked up against the firm's case management system and returned in the call), administrative request (logged for next-business-day follow-up), or genuine urgent issue (immediate escalation with full context handoff).

Escalation logic. When escalation is required, the system pages the appropriate on-call partner via SMS with a one-paragraph summary: who's calling, why, what's been said, what they're expecting. The partner gets context-rich notification, not a callback request. The handoff takes seconds, not minutes.

Knowledge integration. The agent has authenticated read access to the firm's knowledge base — engagement letters, common positions, standard procedures, regulatory updates — and can reference specific documents when answering questions. Where it doesn't know, it says so plainly: "That's a question I'd want a partner to answer directly. Can I get you on the calendar tomorrow morning, or is this urgent enough to escalate now?" The model is explicitly tuned to refuse rather than hallucinate.

Recording and review. Every call is recorded, transcribed, and summarized. Each morning the firm receives a digest: how many calls came in, what was handled, what was escalated, what's queued for follow-up. Partners review escalations same-day; routine calls are quality-spot-checked weekly to ensure the agent stays calibrated to the firm's standards.

Continuous improvement loop. Disagreements between the agent's classification and a partner's after-the-fact judgment feed back into a structured eval set. The agent gets sharper over time without anyone manually rewriting prompts.

What Changed

• Inbound after-hours calls answered live in real time instead of routing to voicemail
• Approximately 70% of routine inquiries resolved in the call without human involvement
• Urgent escalations reach on-call partners within 60 seconds with full call context already attached
• First-line triage moved from next-business-day to real-time — issues that would have sat overnight now get categorized and routed in seconds
• Partner after-hours load materially reduced — only genuinely urgent matters reach them, and those arrive with context rather than a callback request
• Morning queue is pre-triaged and ranked — the team starts the day with a digest and a clean priority list, not a voicemail inbox

The firm's morning routine changed. Where the day used to start with sorting voicemails, building context, and chasing callbacks, it now starts with a clean digest and a triaged queue. The calls that needed humans got humans. The calls that didn't are already resolved.

The Bigger Pattern

This is a use case that specifically rewards production-grade voice AI. A demo voice agent can hold a friendly conversation; a production voice agent has to handle the messy real-world details — accents, interruptions, unclear questions, callers who change their minds mid-sentence, network blips, and the moments where a human absolutely has to take over. Getting those edge cases right is the difference between an AI customers tolerate and an AI customers prefer.

What made this engagement possible in 2026 specifically: model latency dropped enough that real-time conversation feels natural, voice synthesis quality crossed the line where customers stop noticing it isn't a person, and the orchestration tooling matured enough that production deployments don't require a research team. Eighteen months ago, this was experimental. Now it's shippable, reliable, and economically defensible.

We built it with off-the-shelf state-of-the-art components configured against the firm's specific knowledge and operations. No model fine-tuning. No custom inference. The work was in the system design — context engineering, escalation logic, refusal calibration, observability — not in the model layer.

If you operate a professional services firm and your after-hours flow is currently a voicemail inbox, an answering service that doesn't speak your domain, or an on-call rotation that's burning out the team, this is the engagement. Let's talk.