AI coaching for support agents is software that reads an agent’s scored conversations, finds the patterns in them and drafts the feedback a manager would otherwise prepare by hand. Done well, artificial intelligence coaching takes over the preparation, not the coaching: the manager still checks every point and has the conversation.
In short
- AI coaching in support comes in several forms; this guide covers coaching drafted after the conversation from QA scores.
- It is good at the preparation work: reading every conversation, finding patterns and linking each point to evidence.
- It falls short on context outside the transcript, on trust and motivation, on process problems and on anything that decides an agent’s job.
- A manager should review every AI-drafted card with a short checklist before an agent sees it.
- Before you buy, test the tool on the questions in the evaluation checklist below.
If you are here for the product, our page on agent coaching shows how Kaizo drafts coaching from every rating an agent receives. This guide is about the method, including the parts AI should not do.
What is AI coaching for support agents?
AI coaching for support agents is the use of AI to prepare, deliver or reinforce feedback on how agents handle customer conversations. In a support team it usually means one of four things, and they work very differently.
| Type | When it acts | What it does | What it needs from you |
|---|---|---|---|
| Real-time agent assist | During the conversation | Suggests replies, next steps or articles while the agent types or talks | A clean knowledge base and agents who trust the suggestions |
| Coaching drafted from QA scores | After the conversation | Reads scored conversations, finds patterns, drafts feedback for a manager to approve | A scorecard that reflects your standard, and scoring you have validated |
| AI role-play and simulation | Before or between shifts | Plays a customer so the agent can practice a hard conversation | Realistic scenarios taken from your own contacts |
| General-purpose AI coach | Whenever the agent asks | Answers questions about skills, careers or wellbeing | Clear limits on what it may advise |
Most of this guide is about the second type, because that is where QA data turns into coaching. Kaizo works in that mode: it scores conversations after they end and drafts coaching from the results; it does not listen in or coach during a live conversation. For the live mode, see our explainer on real-time agent assist.
What the AI does well in coaching
AI is good at the preparation: it reads every scored conversation, finds the pattern behind an agent’s results, picks representative examples and drafts the feedback with links to the evidence. That is the part of coaching managers most often skip for lack of time.
Without help, a team lead preparing a one-to-one reads a handful of tickets, remembers a few more and writes notes from memory. The feedback then rests on whatever the manager happened to see. AI coaching changes three things:
- Coverage. The draft is built from all of the agent’s scored conversations, not the three a manager had time to open. Feedback drawn from the full record is harder to dismiss as unlucky selection.
- Evidence. Each point links to the conversations behind it, so the agent can read the moment being discussed instead of arguing about impressions.
- Time. On Kaizo’s coaching cards, the card arrives written with the evidence attached, and the manager reviews and personalizes it. Kaizo reports a 90% reduction in coaching prep time on its coaching page.
There is also outside evidence that AI help can lift support agents, with a caveat.
A study of 5,179 customer support agents by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, published by NBER, found that an AI assistant increased issues resolved per hour by 14% on average and by 34% for novice and low-skilled workers, with minimal impact on experienced and highly skilled workers.
That tool gave suggestions during conversations, so it is evidence for real-time assist, not for coaching drafted after the fact.
How coaching drafted from QA scores works
Coaching drafted from QA scores runs in five steps: score, find the pattern, draft the card, have a manager approve it, and follow up on the same criterion. The steps below show what each involves.
Step 1: Score every conversation against your scorecard
Coaching can only be as specific as the scores under it. If a sample covers a few conversations per agent per month, there is not enough data for a pattern. When every conversation is scored against the same criteria, each agent has a full record to draw from.
Step 2: Find the pattern, not the worst ticket
The AI looks for criteria an agent misses repeatedly, and for strengths worth naming. One bad conversation is an incident; the same missed step across twenty conversations is a coaching topic.
Step 3: Draft the card with evidence
A good draft names one strength, one thing to work on and one specific practice, each with links to the conversations behind it. Generic advice such as “show more empathy” is a sign the draft needs work.
Step 4: The manager reviews, edits and approves
This is the step that makes it coaching. In Kaizo, a manager reviews, edits and approves every card before an agent sees it, and adds the personal note. The AI does not talk to agents directly.
Step 5: Follow up on the same criterion
After the coaching conversation, watch the criterion that was coached over the next few weeks.
Agents can also comment on a card to add context or disagree, and because each point links to a conversation, the disagreement is settled by reading it together. Our guide on turning QA data into coaching covers how to measure the change.
Where does AI coaching fall short?
AI coaching falls short wherever the answer is not in the transcript: an outage, a broken process, an agent’s circumstances, or the trust between agent and manager. It also cannot make decisions about an agent’s job, and its advice is only as good as the scorecard behind it.
Here are the limits we see most often, and what to do about each.
| Where it falls short | Why it happens | What to do instead |
|---|---|---|
| Context outside the conversation | The AI sees the transcript, not the outage, the missing tool or the policy change that morning | The manager checks for system causes before approving; see being scored for things outside your control |
| Process problems disguised as skill gaps | Twenty agents giving the same wrong answer looks like twenty coaching topics | Look across agents first; a shared failure goes to whoever owns the macro or policy |
| Trust and motivation | Feedback lands when the agent trusts the person giving it, and that comes from the relationship | Keep the coaching conversation human, and let the agent push back |
| A weak or unvalidated scorecard | The AI repeats whatever the scorecard rewards, including vague criteria | Fix the scorecard and validate the scoring before using it for coaching |
| Too many points at once | A full record shows every miss, and a draft can list them all | Coach one behavior at a time; park the rest |
| Generic advice | Some drafts restate the criterion instead of pointing to the moment | Reject drafts that do not cite a specific conversation |
| Experienced agents | The NBER study found little gain for experienced agents | Use the data to confirm strengths and give them harder cases, not routine tips |
| Decisions about someone’s job | Scores and drafts are inputs, not verdicts | Keep pay, promotion and performance decisions with people, with a dispute route |
The scorecard problem
AI coaching inherits every weakness of the scoring under it. If your empathy criterion is vague, the AI will coach a vague version of empathy at scale. Before you rely on AI-drafted coaching, run the scoring against conversations your team has already rated. Our protocol on how to validate AI QA scoring takes a week.
The relationship problem
Coaching changes behavior when the agent believes the coach wants them to succeed. No draft can create that. The risk with AI coaching is that managers send the card and skip the conversation, which turns coaching back into monitoring. Our guide to coaching from QA without micromanaging covers how to keep the tone right.
The regulation question
In the EU, AI systems intended to “monitor and evaluate the performance and behaviour” of workers are listed as high-risk in Annex III, point 4(b), of the EU AI Act.
The AI Omnibus, Regulation (EU) 2026/1744, moved the date from which those high-risk obligations apply to 2 December 2027, a date the Council of the EU also confirms.
Whether a given coaching tool falls under point 4(b) depends on what it is intended for and how you use it, so ask your legal team, and keep a person accountable for every decision about an agent.
A manager’s checklist before approving a drafted card
Before approving, check that each point cites a real conversation, that the cause is the agent and not the process, that the card asks for one change, and that the tone fits this person. Six questions take a few minutes per card.
| # | Check | If the answer is no |
|---|---|---|
| 1 | Does every point link to at least two conversations that show it? | Remove the point or find better examples |
| 2 | Is this the agent’s behavior, not a macro, tool or policy problem? | Send it to the owner of the process instead |
| 3 | Is there one main thing to work on? | Cut the card down to the most important behavior |
| 4 | Does it name a real strength, with evidence? | Add one; a card with only misses reads as a reprimand |
| 5 | Is the practice step something the agent can do next shift? | Rewrite it as a specific action |
| 6 | Would you say this to the agent in these words? | Edit the tone and add your personal note |
Here is an example of a draft and the manager’s edit. The card and agent are illustrative, not a customer case.
AI draft. Strength: clear next steps in refund conversations. Work on: confirming the issue is resolved before closing. In 14 of 41 chats this week the conversation ended without the customer confirming the fix. Practice: ask “Has that solved it for you?” before the closing line.
Manager’s edit. Kept the strength and the practice. Checked the 14 chats: 6 were order-tracking contacts where the tracking page was down on Tuesday, so the customer could not confirm anything. Changed the count to 8 of 35 and added: “Tuesday was not on you, the tracking page was down. The other eight are worth a look together on Thursday.”
That edit is the job. The AI found the pattern in minutes; the manager knew about Tuesday.
Questions to ask before you buy a coaching tool
Judge a coaching tool that uses AI on where its feedback comes from, whether a person approves it, how it handles disagreement and what it does with process problems. Ask each vendor the same questions and ask to see the answers on sample data in a live demo.
- Source. Is the feedback drawn from all of an agent’s scored conversations, or from a sample?
- Evidence. Does every point link to the conversations behind it?
- Approval. Does a manager review and approve each card before the agent sees it?
- Disagreement. Can agents reply to a card, and where does that reply go?
- Scoring accuracy. Can you test the scoring against your own past ratings before it drives coaching?
- Process versus skill. Does the tool separate shared failures, such as a broken macro, from individual skill gaps?
- History. Are coaching history, scores and one-to-one notes kept together, so a new manager starts with context?
- Language. Does it draft coaching in the languages your agents work in?
- Channels. Does it cover the channels your team works, and on one scorecard?
- Governance. Can you show who approved each piece of feedback and when?
In Kaizo, coaching history, QA scores, one-to-one notes and tasks sit in one place, and cards are available in every major support language. For running the conversation itself, our coaching one-to-one template gives a structure.
How to start with AI-drafted coaching in one team
Start with one team and one behavior: validate the scoring for that criterion, draft cards for it only, and review every card with the six checks.
Compare that criterion before and after a month, and keep the first month narrow on purpose. One criterion shows whether drafts are specific, whether managers actually edit them and whether agents respond. Widen it to the full scorecard once the review habit is in place.
See AI-drafted coaching cards on sample data. Kaizo drafts coaching from every rating an agent receives, linked to the conversations behind it, and a manager approves each card before it is shared. Book a demo.
Frequently asked questions
What is an AI coach in a call center?
An AI coach in a call center is software that analyzes agents’ conversations and produces feedback, either live during the call or chat, or afterwards from QA scores. The useful versions link every point to specific conversations and leave the final feedback and the coaching conversation to a manager.
Will AI replace coaches in customer support?
AI is unlikely to replace coaches, because the part that changes behavior is the relationship and the judgement about context, which AI does not have. What it replaces is the preparation: reading conversations, finding examples and writing a first draft. Managers then have more time for the conversation itself.
Does AI coaching actually improve agent performance?
There is good evidence that AI help can lift support agents, especially newer ones: an NBER working paper found a 14% average productivity gain from a real-time AI assistant, and 34% for novices. Evidence on coaching drafted from QA scores is thinner, so measure the coached criterion yourself before and after.
Should agents know their coaching was drafted by AI?
Yes. Tell agents how their feedback is produced, which conversations it draws on and that a manager reviews every card. Agents trust feedback more when they can see the evidence and reply to it, and being open about the process avoids the sense of being watched by a machine.
Is AI coaching the same as real-time agent assist?
No. Real-time agent assist suggests what to say or do while the conversation is happening. AI coaching from QA scores looks back at finished conversations, finds patterns and prepares feedback for a manager. Some teams use both; they solve different problems and need different checks.
