AI Coaching for Support Agents: Feedback From Real Conversations

By Linh Ngo · Updated Oct 11, 2026

AI coaching for support agents is the practice of using AI to read or listen to real customer conversations, pick out the moments that show a skill gap or a strength, and hand those moments to a manager as concrete material for a one-on-one. The AI does the finding. The manager does the coaching. That split is the whole idea, and teams that blur it usually end up with either a pile of scores nobody acts on or a system agents resent.

Most support managers coach from a handful of conversations they happened to sample. A team of 30 agents handling thousands of contacts a week can't be understood from five tickets each. AI changes the sampling problem: it can look at every conversation and surface the three that matter for this agent this week. This guide covers how that works, who does what, how to set it up in steps, and how to keep the program on the right side of the line between development and surveillance. It's part of the AI for Customer Service collection; for the wider picture, start with AI in customer service.

Key Facts

• In a draft working paper on a European financial services call center, AI-enhanced coaching cut average call handling time by about 7 percent, with larger gains for lower-tenured agents (Fouarge et al., Netspar working paper, January 2025).
• That study covered 187,839 calls handled by 147 agents during a staggered 2023 rollout (Netspar working paper).
• In a separate study of 5,179 support agents, an AI assistant raised issues resolved per hour by 14% on average and by 34% for novice and low-skilled workers, with minimal impact on experienced workers (Brynjolfsson, Li and Raymond, NBER Working Paper 31161).
• 51% of US workers say they know their employer monitors them with technology, and 56% of those monitored report typically feeling tense or stressed at work (American Psychological Association, 2023 Work in America survey).
• 45% of monitored workers say their work environment hurts their mental health, versus 29% of workers who aren't monitored. APA stresses these are correlations with no confirmed root cause (APA).
• The EU AI Act prohibits AI systems that infer emotions of people in the workplace, except for medical or safety reasons (EU AI Act, Article 5).

What AI Coaching Actually Does

Picture the traditional version. A team lead pulls five tickets per agent at the end of the month, reads them, and writes up feedback. The sample is tiny, the feedback arrives weeks after the conversation, and the agent can barely remember the contact. Often the five tickets are the ones that were easiest to find, not the most useful.

AI coaching replaces the finding step, not the conversation. In ACE terms it's the Meeting Intelligence pattern applied to support: ingest the conversation (text from chat and email, or transcribed audio from calls), analyze it against what you care about, and generate something a manager can use, such as a short list of flagged moments with the relevant lines highlighted.

Typical things it surfaces:

• Coachable moments. A conversation where the customer was confused for several turns, where the agent skipped a required step, or where a tough situation was handled unusually well.
• Patterns across conversations. An agent who consistently under-explains refund timelines, or one whose replies get shorter and cooler late in the shift.
• Strengths worth copying. The same scan that finds gaps finds the agent whose explanations of a billing error keep landing. That becomes a shared example for the team.

The research on call selection supports the approach. A 2023 paper on a system called AI Coach Assist describes using language models to decide which calls are coachable based on the quality assurance questions managers set, so supervisors aren't limited to random samples (Coach Assist, arXiv). It's a system paper without reported field results, so treat it as a description of the method, not proof of impact.

Where does it differ from AI agent assist? Agent assist works during the conversation, suggesting replies and surfacing articles in real time. Coaching works after it, turning finished conversations into development. They feed each other: coaching tells you which topics agents need more help on, and assist is where you can put that help.

What the Evidence Says

Be careful with big claims here. The evidence is encouraging but young.

The most relevant study is a working paper from researchers studying a large financial services provider in Western Europe. The firm rolled out an AI tool that analyzes the spoken content of calls to support its coaching process, and it did so in stages, which lets researchers compare agents who'd received it with those who hadn't. They report that AI-enhanced coaching reduced call handling time by about 7 percent, that the mechanism was better communication style, and that the gains were substantially stronger for lower-tenured agents (Netspar working paper). Two cautions apply. It's a draft, and the treated agents were considerably less experienced than the control group, which is the kind of imbalance you'd want a later version to address.

A broader result points the same direction. The NBER study of generative AI assistance found the biggest benefits for newer workers and concluded the tool seemed to spread the practices of more able workers (NBER Working Paper 31161). That study is about real-time assistance, not coaching. But the shared pattern is useful for planning: expect AI-supported development to help newer agents most, and don't promise veterans a dramatic jump.

Notice also what these studies don't show. They measure speed and communication quality, not whether customers were happier or whether agents stayed longer. Put those on your own scoreboard.

Who Does What: AI Finds the Moments, Humans Coach

The cleanest way to run this is a hard division of labor.

The last row matters most. AI flags are a starting point for a conversation, not a verdict. An algorithm sees that a reply was long; it doesn't know the customer was elderly and needed things spelled out. A manager who knows the context and asks first, "walk me through what happened here", gets better information than any score, and the agent leaves feeling heard.

Zendesk's QA documentation shows how a mainstream product frames this. Its admin guide describes autoscoring that evaluates and scores all customer interactions automatically, and then separately describes creating coaching sessions where you "bundle specific conversations and pinned notes", plus tracking whether agents have reviewed the feedback they received (Zendesk Help Center). The scoring is automatic. The session is a human event. For the scoring side in depth, see AI quality assurance for support.

Real Examples by Function

The method is the same everywhere. The moments differ.

Tier 1 chat support. The AI flags conversations where the customer asked the same question twice. The manager and agent read two of them together and find a pattern: the agent answers the question asked, not the one underneath. Goal for next week: confirm understanding before answering.

Billing and collections. Billing is where tone matters most and mistakes cost the most. The AI flags calls where the customer's language shifted from neutral to frustrated right after a policy explanation. The coaching conversation is about how to explain a late fee, not about the fee. The agent who explains it best becomes the model.

Technical support. The flags are about process: skipped diagnostic steps, premature escalations, tickets closed without confirmation. Managers use them to improve the troubleshooting guide as well as the agent, since three agents skipping the same step usually means the step is badly documented.

Email and written channels. AI is strongest here because text is easy to analyze. Common coaching topics are reply length, unclear next steps, and missing acknowledgment. See AI for email support for how this channel works.

New-hire onboarding. In week three, instead of waiting for a monthly review, the manager gets a short list from the new agent's own conversations: two things done well, two to practice. Given the research above on newer workers, this is where coaching is most likely to pay off.

How to Implement AI Coaching, Step by Step

1. Start with the agents, not the tool. Before you buy or switch anything on, tell the team what you're planning and why. Explain what will be analyzed, who sees the output, and what it won't be used for. Do this first, because rumor fills any silence.
2. Define what good looks like. Write a short scorecard: five to eight behaviors that matter, such as understanding the question, accuracy, clear next steps, appropriate tone. Involve experienced agents in drafting it. Rubrics invented by management alone tend to measure what's easy to count.
3. Check your data. Conversation analysis needs complete, correctly attributed transcripts. Fix tagging, channel coverage and consent problems before you start. Preparing support data for AI covers the groundwork.
4. Run it in shadow mode. For a few weeks, let the AI score conversations while managers also review a sample by hand. Compare. If the AI and your best reviewers disagree often on a behavior, fix the rubric or drop that behavior.
5. Pick a cadence and a cap. One weekly or biweekly coaching conversation per agent, with no more than two or three moments to discuss. A list of twenty flagged items is a to-do list, not coaching.
6. Give agents the same view. Agents should see their own flagged moments and scores before the conversation, and be able to add context or disagree. This is the single biggest lever for trust.
7. Turn moments into one commitment. End each session with a single, specific thing to practice, written down by the agent. Check it next time.
8. Share the good stuff. Once a month, share an anonymized or consented example of an excellent conversation with the whole team. If AI coaching only ever shows people their errors, it feels like surveillance. Showing wins fixes that.
9. Review the system quarterly. Look at flag accuracy, agent feedback, and outcomes (handle time, resolution, CSAT, attrition). Retire flags that don't help.

Avoiding the Surveillance Feel

This is where programs live or die. The same technology can feel like a coach or like a hall monitor, and agents decide which within about a week.

The numbers explain why it deserves attention. In the APA's 2023 survey, 56% of workers who experience monitoring reported typically feeling tense or stressed at work, and 45% said their work environment harmed their mental health, compared with 29% of workers who aren't monitored. Customer and client service roles reported monitoring awareness at 55%, higher than the 51% overall (APA). APA is explicit that these are correlations and the root cause isn't confirmed, so don't quote them as proof that monitoring causes stress. Do treat them as a warning about the environment you're entering.

Practical safeguards:

• Be transparent. Document what's analyzed, what scores mean, who can see what, and how long data is kept. Put it in writing and update it when the setup changes.
• Give notice and follow the law. Requirements on recording, consent and employee data differ by country and sometimes by state. Check with legal counsel and any works council before launch. Don't assume a customer-facing recording notice also covers internal performance analysis.
• Avoid emotion-inference scoring of your own staff in the EU. The EU AI Act bars AI systems that infer emotions of people in the workplace, outside medical or safety uses (EU AI Act, Article 5). If a vendor feature rates agent "mood" or "stress" from voice or video, get a legal read before turning it on, wherever your staff sit.
• Use data for development first. Say plainly, and mean it, that flagged moments are conversation starters. If you also intend to use scores in performance reviews, say so openly, set thresholds with human review, and give agents a route to contest.
• Keep humans in decisions. No pay, promotion or disciplinary decision should rest on an AI score alone.
• Let agents see their own data first. Surprise is what makes feedback feel like surveillance.
• Collect less. Analyze what your scorecard needs. Don't keep every possible metric just because the tool can produce it.

Pitfalls to Avoid

Scoring everything and coaching nothing. AI makes scores cheap, and teams drown in dashboards. If a flag never leads to a conversation, stop generating it.

Coaching to the metric. If handle time is weighted heavily, agents will rush customers. The 7 percent reduction in handling time from the Netspar paper is only good if quality holds, and the authors report it came through better communication, not shortcuts. Balance speed with resolution and satisfaction.

Treating the AI score as truth. Models misread sarcasm, regional phrasing, multilingual conversations and unusual product situations. This is a particular risk in multilingual support. Always let the agent explain before judging.

Ignoring what agents say. If experienced agents say a flag is wrong, investigate. They're usually your best calibration data.

Using it only on struggling agents. Run the same program for the whole team, including top performers. Selective use looks like punishment.

Forgetting the roles around it. Coaching time is a manager's time. If you add AI coaching without freeing up team leads, it won't happen. See how AI changes support team roles for how to plan that capacity.

A Simple Way to Measure Whether It's Working

Pick four measures before you start and look at them every quarter: the quality score on your scorecard, one customer outcome (CSAT or resolution), one speed measure, and one people measure, such as agent confidence in a short pulse survey or attrition. Compare agents who've been coached with the system against a baseline period, and watch newer agents separately, since that's where the research suggests the biggest effect. If speed improves while the pulse survey drops, you've built a surveillance tool. Fix that before scaling.