What is AI-powered agent assist?

Take a customer who writes to a software company's support desk: "I can't track my employee's hours during business trips." The sentence is an example in the working paper behind a study of these tools, and the paper notes that the right answer depends on which version of the product the customer has. An agent who doesn't know that has to stop and search.

An assist tool reads the message as it arrives. On a phone call, Five9's agent assist product transcribes the conversation in real time, which gives the tool text to work from.

Next the tool works out what the customer is asking. Amazon Connect's documentation says its assistant detects customer intent during calls, chats, tasks and emails. It then gives the agent "real-time generative responses along with links to relevant documents and articles." So the agent sees a drafted answer beside its source and can check one against the other before sending anything. The agent can also ask a question directly, in plain language or keywords.

Some products add structure around the answer. Five9 shows guidance cards and checklists that recommend the steps to follow. Genesys Agent Copilot gives the agent next best actions during the conversation. When it ends, Copilot provides a summary, the reason for contact and the resolution, and it suggests wrap-up codes.

The suggestions come from material your company connects and settings your company chooses. Amazon lists knowledge bases and step-by-step guides among the things an administrator can integrate. Google's Agent Assist runs each conversation under a conversation profile, a group of settings that determines which suggestions the human agent receives. Someone on your side decides what the tool may suggest and what it draws on.

A virtual agent and post-call analytics sit beside agent assist, and the three are easy to confuse under the label "AI." A virtual agent talks to the customer in place of a person, and post-call analytics studies the conversation after it ends. Agent assist works between them, during the part of the conversation a human handles.

Google's documentation lays out the order. A customer may first reach a virtual agent, which tries to resolve the request by itself. If it can't, the conversation is handed to a human, and at that stage "Agent Assist supplies real-time document and response suggestions to the human agent."

Timing is the practical difference. Xima, the platform behind Atlantech's contact center service, documents AutoQA evaluations that grade a call against your own standards. Its guide says they "are processed after a call has concluded and are generally available in a minute or less." That serves a supervisor who coaches after the call, where agent assist serves the agent during it. Xima's guide also describes an AI agent that chats with customers under a nickname and greeting you choose, which is the virtual-agent neighbor. When a quote says "AI," ask which moment of the conversation the tool works in.

A peer-reviewed study of one deployment measured the effect. Erik Brynjolfsson, Danielle Li and Lindsey Raymond followed the rollout of an AI assistant to 5,172 customer-support agents and published the results in 2025 in The Quarterly Journal of Economics. Agents with the assistant resolved 15% more issues per hour on average. Less experienced agents improved both the speed and the quality of their work, while the most experienced agents saw "small gains in speed and small declines in quality."

According to the project's working paper, the agents did chat-based technical support for a Fortune 500 software company, and their chats averaged 40 minutes. The tool suggested responses and linked to the company's internal documentation. Agents were free to ignore it, and on average they followed 38% of its recommendations.

So the study measures long, written, technical conversations. It doesn't measure phone calls, short billing questions or sales. Its gains were concentrated among less experienced agents, and that is the part to hold up against your own team.

Five9's page adds a customer quote, in which TruConnect says the tool reduced its average handle time by 30 seconds.

On the Genesys pricing page, Agent Copilot features are billed in AI Experience tokens with Cloud CX 1, CX 2 and CX 3. They are included in Cloud CX 4, which costs $240 per user per month, billed annually. Every organization gets 250 tokens a month on named-user licensing, and the page doesn't publish what an extra token costs.

Amazon Connect bills differently. Its pricing page lists agent assist and post-contact summaries among the capabilities of a service billed at $0.038 per voice minute plus telephony charges, and it says there are no tokens.

Suppose 25 agents are on Cloud CX 3 at $155 per user per month and the company wants Copilot included. Cloud CX 4 costs $85 more per user per month, which comes to $25,500 a year, and it includes more than Copilot. The study's gains were concentrated among less experienced agents in written technical support. Set the $25,500 against how many of your 25 are new and how much of their work looks like that.

Suggestions draw on the material you connect, so stale material can produce stale suggestions. A suggestion can also be wrong when its source is current. Someone has to keep the material up to date and watch what the tool does with it, and that work continues for as long as the tool is in use.

A tool that works from a transcript depends on the transcript's accuracy. Test transcription on your own recordings before you rely on it, because accents, product names and line noise all affect what the software hears.

The last requirement is an agent who reads before sending. Agents in the study followed 38% of the recommendations they were shown, and the working paper notes that a model can make incorrect ones. The same care applies to the summary at the end, because it often becomes the record other teams read later.