AI training · AI agents
AI agent training: working alongside agents
An agent does not answer a question, it does a piece of work. So the team needs different skills from the ones a chat assistant asks for: describing the task and its limits, checking the result, handing a conversation over in good time, and maintaining the knowledge base the agent draws on.
Definition
What is an AI agent?
An AI agent is a program that is given a goal and then carries out the sequence of steps needed to reach it: it understands the request, takes data from the company’s systems, performs an action and reports on it. Unlike a chat assistant, which returns text to a person, an agent works inside the process itself: it books the appointment, updates the record, answers the customer. So its error does not stay on a screen but reaches a customer or a system immediately, and every other difference on this page follows from that one.
AI agent training is a practical programme for the team that will work with such an agent every day. It covers four things: how to describe the task and the limits, how to check the output, when a person should take the conversation over, and how to maintain the sources the agent draws on. It is a separate part of the AI training for companies programme, for teams where AI is no longer just a help with text. The Lithuanian edition is at AI agentų mokymai, and the terms recurring throughout are explained in the AI glossary.
Context
Agents are multiplying faster than the skill of supervising them
Four numbers that explain why deploying an agent is not only a technical project.
Microsoft’s Work Trend Index survey was conducted by an independent research firm among 20,000 full-time employed or self-employed knowledge workers who use AI at work, across 10 markets, between 18 February and 7 April 2026. Dell’Acqua and colleagues ran a pre-registered field experiment with 758 consultants.
The difference
An AI agent and a chat assistant: what actually separates them
The distinction looks like terminology until the first error. Then it turns out to be a distinction about accountability.
| Criterion | Chat assistant | AI agent |
|---|---|---|
| What starts the action | A person writes a request and gets an answer | The agent is given a goal and carries out the sequence itself |
| Access to systems | Usually none: it works only with what is in the conversation | Calendar, customer database, CRM, telephony or email |
| What the output is | Text a person still reviews and then uses | An action that happened: an appointment booked, a record updated, a reply sent |
| Where an error goes | Stays in the chat window until a person uses it | Reaches the customer or the company system immediately |
| What it asks of the team | The skill of framing a request and checking a text | A written goal, limits, handover rules and regular review |
| The nature of supervision | Every line is reviewed before use | A sample of conversations, all exceptions and all handovers are reviewed |
| Where failure is noticed | Immediately, because a person reads the answer | Later, from a customer’s question or from a weekly report |
In practice the line is not sharp: the same system can behave as a chat assistant on a website and as something closer to an agent elsewhere. What matters is not the label but the question of whether the system has the right to change anything by itself. Where it does, both the training content and the supervision routine change. How an agent connects to a calendar, a CRM or a phone system is described on integrations, and the system this training is usually built around is the AI teammate.
Skills
What the team has to know to prepare and supervise an agent
Six skills. None of them is technical, and none of them appears by itself.
01
State a goal, not a single request
A chat assistant needs a question. An agent needs a described outcome: when the task counts as done, which steps to take, what to ask the customer for, and what never to do without confirmation. It is closer to a work instruction for a new colleague than to a query in a search box.
02
Describe the limits and the exceptions
Most of an agent’s quality is not what it can do but what it is not allowed to do. Limits are stated concretely: do not change a price, do not promise a date, do not give a medical or legal assessment, do not offer a discount. A team that does not write these down discovers them later in a customer complaint.
03
See where the agent gets its information
An agent answers from the company’s knowledge base, its documents or its system data. When an answer goes stale, the model is not at fault, the source is. The team has to be able to find that record and fix it themselves, without waiting for a supplier to do it.
04
Read a transcript like an auditor
An agent’s work is reviewed against specific questions rather than a general impression: was the request understood, was the right information taken, did the action happen, did the customer get what they asked for. Teams usually do not have this skill, because nobody previously had to assess a sample of a colleague’s conversations.
05
Tell an agent error from a process error
When an agent fails to book an appointment, the cause is often not in the agent: there were no free slots in the calendar, the service is named differently in the system, one rule in an internal document contradicts another. The training gets a team into the habit of asking first where the chain broke.
06
Notice a convincing wrong answer
The most dangerous output is not an obvious error but a fluent, tidy, wrong one. We work deliberately with such examples, because recognising them takes practice: correct language and a confident tone read to a person as a signal of reliability all by themselves.
Why the skill of writing a good request is not enough
In Microsoft’s 2026 Work Trend Index, the workers it classes as Frontier Professionals, 3,233 of the 20,000 AI users surveyed, are considerably more likely than the rest to say their manager sets quality standards for AI work: 83 percent against 57 percent. Similar gaps appear on whether the manager openly uses AI (85 against 64) and creates space for experimentation (84 against 61). That is a question of arrangements rather than of tools, and it is settled by agreeing how output gets checked, not by granting access.
Suitability
Where an AI agent fits in a process, and where it does not
The honest answer is rarely “everywhere” or “nowhere”. Usually a company has two or three processes where an agent gives a clear benefit, and ten where it only adds another step.
Where an agent fits
- The process repeats and has a clear outcome: taking an enquiry, booking an appointment, sending a reminder, checking a status.
- There are many cases and they resemble each other, so one rule covers a large share of situations.
- The company already has a written procedure or a knowledge base to work from.
- The result can be checked objectively: either the record is in the system or it is not, either the slot is taken or it is free.
- An error is corrected quickly and cheaply, because either the customer or the system itself notices it.
Where an agent is better left out
- The process is undocumented and every member of staff does it their own way. An agent in that environment freezes one arbitrary version of it.
- The decision rests on unwritten context: the customer’s history, a previous agreement, something a colleague happens to know.
- Rare exceptions make up the bulk of the work. The agent then covers a minority of cases while supervision costs more than it saves.
- The outcome requires legal or financial accountability with a named person’s signature.
- The data cannot leave the internal perimeter and the process cannot be done without it.
The invisible boundary we teach separately
In a pre-registered field experiment with 758 consultants, AI helped participants complete 12.2 percent more tasks and complete them 25.1 percent more quickly, with more than 40 percent higher quality than a control group. Yet on a task chosen deliberately to sit outside the model’s capability, the same participants using AI were 19 percentage points less likely to produce a correct solution than those working without it. The boundary between those two regions is invisible from the outside, so recognising it is a separate objective of the training rather than something that can be assumed.
Quality
How to judge an AI agent’s output
A general impression will not do. An agent almost always sounds tidy, so “looks fine” is precisely the judgement that lets an error through. The training uses a five-step review the team can run itself afterwards.
Was the request understood correctly
Read the customer’s first words and the agent’s first reaction. If they already diverge, the rest of the conversation is a tidy answer to the wrong question. That error is usually fixed in the task description rather than in the model.
Was the information taken from the right source
Check whether the agent relied on a current record or on general knowledge. If the answer is right but did not come from the company source, next time it may be wrong and the cause will go unnoticed.
Did the action actually happen
An agent can say an appointment is booked when there is no record in the system. So a review always covers the system side, not only the text of the conversation. This step is what separates reviewing an agent from listening to an ordinary call.
Did the customer get what they asked for
Judge the outcome through the customer’s eyes: were they left with an unanswered question, did they have to repeat themselves, did they have to make a second call. A conversation can be formally correct and useless at the same time.
Was the boundary set correctly
The last question is not about the agent but about the arrangement: should this case have stayed with the agent at all. If not, the handover rule is what gets fixed, not the phrasing. After the first two weeks of review, this is usually the layer that changes.
The second question teams ask straight after their first review is how many conversations to check. The practical routine: everything for the first two weeks, then a daily sample, and once the work settles a weekly sample plus two groups that are always checked. The first group is every case the agent handed to a person, because that is where you see whether the boundary is right. The second is every case where a customer came back about the same thing, because that is the commonest sign that a first conversation looked successful and solved nothing. Those two groups are a small share of all conversations, so reviewing them is realistic even for a small team.
Who is responsible for the output
DigComp 3.0, the fifth edition of the European Digital Competence Framework published by the European Commission’s Joint Research Centre in 2025, states it directly in competence statement CS1.2.10: “Recognise that AI systems may produce output which is inaccurate, even if it may seem plausible, and that the human using the AI system is responsible for checking the quality and validity of information and content generated.” That is the practical basis for the whole review routine: responsibility does not move to the system along with the task.
Handover
When an agent should hand a conversation to a person
The handover rule is the most important line in the whole description of an agent. It states where automatic work ends and human accountability begins. Without it an agent tries to settle things it should not, and the customer notices before the company does.
In the training the team writes its own boundaries, because they depend on the business. Even so, five rules are worth considering for any team: customer dissatisfaction or a repeat contact about the same thing, the same request misunderstood twice, a request outside the described scope, any decision involving money or a contract, and every case in which the agent would have to guess. A sixth is often added in health, legal and financial work: any question whose answer could be understood as a professional assessment.
The other half matters just as much: the handover has to be smooth. The customer should not repeat what they have already said, and the member of staff should get a short summary of the conversation before picking up. Technically that is a matter of how the systems are wired together, but organisationally it needs two more things: a person who is actually reachable at that moment, and an agreement about how quickly they answer. Without those, the technical capability goes unused.
The commonest mistake in handover rules
The boundary is set too high, because the wish is for the agent to resolve as much as possible. The result is the opposite: a customer spends two minutes with a system that cannot help them, and only then reaches a person, in a worse mood than at the start of the conversation. An early handover costs one member of staff a minute; a late one costs the relationship.
Operations
What changes in a company once an agent does real work
This is the part buyers usually appreciate too late. An agent is not one more licence alongside the existing ones: it changes four things in the daily life of a company, and each of them needs a human decision.
A named person appears, instead of shared responsibility
An agent needs an owner: a member of staff who reviews a sample of conversations each week, maintains the knowledge base and decides whether a rule should change. Without a name in a job description, this function quietly disappears during the first quarter.
Part of the work becomes supervision
Somebody who used to answer calls now handles the cases the agent did not resolve, and checks how it resolved the others. That is different work, and it has to be taught separately, because the accuracy demanded of it is higher than in ordinary service work.
Company knowledge has to be written down
An agent cannot ask the colleague at the next desk. Everything it needs to know has to be in a document. So deploying an agent almost always forces internal information to be tidied up, and that benefit remains even if the agent’s scope is later narrowed.
Measurement becomes weekly
Instead of an annual survey there are weekly figures: how many enquiries the agent completed without a person, how many it handed over, how often it handed over too late. Those numbers change the conversation with management, because they can be compared with last month.
Those changes explain why BCG puts about 70 percent of AI implementation challenges with people and process and only about 10 percent with the algorithms themselves. Microsoft’s 2026 data points the same way: organisational factors such as culture, manager support and talent practices account for more than twice the reported AI impact of individual mindset and behaviour, 67 percent against 32. The practical conclusion is plain: a deployment plan needs a section about people, not only about integrations. How those decisions are taken at management level is on AI training for executives, and the rules that govern them belong in the internal AI use policy.
Programme
How AI agent training runs
The session is run at keyboards with your process and your material. If no agent is running yet, we work with the process you are planning to hand over to one.
| Part | What we do | What you leave with |
|---|---|---|
| Anatomy of an agent | We take one real conversation apart into steps: what was understood, what was retrieved, what was done, where it was handed over | A shared vocabulary in the team for talking about an agent’s errors concretely |
| Describing the task and the limits | The team writes its own process goal, permitted actions and prohibitions | A first list of limits that can go straight to the deployment team |
| Review exercises | We assess a sample of conversations against the five-step routine and work through the disagreements | One consistent way of judging, so two people rate the same conversation the same way |
| Handover rules | We set when a conversation goes to a person and who picks it up at that moment | A written handover procedure with named people and response times |
| Maintaining the knowledge base | We find a stale record, correct it and check the result | The skill of fixing the source yourself, without waiting in a supplier queue |
| Measurement and a return session | We agree which measures you watch weekly and when we review the result | A short list of measures and a date for the second session |
The programme is agreed after a conversation, because teams arrive from very different positions: one already runs an internal AI co-pilot and needs a supervision routine, another is still considering its first process and gains more from a suitability assessment. The data-handling boundaries that apply to every module are on the security page. Function-specific programmes exist for accounting teams and legal teams. Scope and price are agreed per engagement.
FAQ
Frequently asked questions.
Founder & CEO, AInora
Building AI digital administrators that replace front-desk overhead for service businesses across Europe. Previously built voice AI systems for dental clinics, hotels, and restaurants.
View all articlesStart with one process.
An hour on a call to look at which of your processes suits an agent, which does not, and what the team will need in order to supervise it. No obligation.