Human Assist (Resolution Steps and Suggested Replies)

Status: Public Preview — behaviour may change.

What it does

Gives representatives real-time resolution steps and suggested replies next to a live conversation in Copilot Service workspace, driven by a real-time AI agent you build with its own knowledge, tools and instructions. The rep stays in control: insert, edit or ignore each reply, and approve any tool run (a refund, an account update) before it fires.

Key facts

Needs a real-time AI agent built for Human Assist, not a repurposed self-service agent. Capabilities are switched on for reps through experience profiles.
Channels in preview: live chat, voice, SMS, Teams, Direct Line, Messaging API and WhatsApp. Email isn't supported. Voice works from the real-time transcript, not the audio.
Languages: resolution steps in English, Arabic and Spanish; suggested replies in English only. Other languages aren't blocked, but they aren't supported either.
The step checklist marks steps Completed, Skipped or Not started and warns on skipped ones. Status is AI-generated and can be wrong in long workflows.
Supervisors get a resolution-step adherence view in the Quality Assurance Agent report (omnichannel historical analytics), filterable by period, channel, queue and rep. It covers steps only; there's no analytics for suggested replies.
Reps can rate replies and steps if the tenant has Basic Copilot feedback enabled; sharing prompts and responses needs Additional Copilot feedback.

When to use / skip

Worth piloting where reps follow a documented resolution path and your knowledge base is in decent shape. Skip it if your quality process depends on deterministic, word-for-word scripted answers, or if email is where most of your volume sits.

Configuration decisions

Which knowledge sources and tools the Human Assist agent gets, and which tools need rep approval.
Which experience profiles switch it on, remembering it's all-or-nothing across voice and digital.
Whether to enable Copilot feedback at tenant level, and at which level.

Gotchas

Topics aren't supported, so no question-and-answer nodes or scripted deterministic flows. Adaptive Cards don't render; replies are plain text.
Proactive suggestions only. Reps can't ask it a question in the preview.
It uses a fixed model, whatever you pick in the agent canvas.
Profiles can't split voice from digital, and reps can't turn it off themselves.
It reads the whole transcript and doesn't drop earlier context when the customer changes topic. The context window is 128K tokens, so keep knowledge articles sensibly sized.

Consultant notes

Treat the adherence report as coaching data, not a scorecard. Step status is AI-inferred, and reps who learn they're being measured on it will tick through steps rather than resolve the issue.
The quality of the resolution steps is the quality of your knowledge. Budget time to tidy the source articles before UAT, or you'll spend the pilot arguing about the AI when the problem is a stale article.
Because access is per experience profile with no voice/digital split, pilot with a dedicated profile and a small team rather than flipping it on for a shared one.
It's a preview, so keep it out of the committed scope in a SOW and agree the fallback (the existing Copilot suggestions) up front.

Public preview as of November 2026. Revisit at GA — channels, languages and the fixed-model limit are the likeliest to move.