AI agents: what they are and how to put one to work

An AI agent is software that receives a goal, reads the company’s data and context, decides the next step, takes action in the systems it has permission to use and asks a person for approval when the rule requires it. A chatbot answers within a script, a copilot suggests and waits for someone to act, and traditional automation repeats a fixed sequence without deciding anything. The agent combines reading, deciding and acting, so it needs a design that leadership signs off on: goal, context, permissions, rules, human approval, owner and measurement. To put the first one to work, pick a high-volume task, done on data the CRM already holds and with mistakes that are cheap to fix, and give it an owner with a first and last name.

At Iasbeck&Co we implement CRM and AI agents in B2B operations, and the question I hear most from sales, marketing and service leaders is what they are buying when a vendor says “agent”. In October I ran a live workshop showing agents inside a real CRM. The uses by area are in AI in sales, marketing and CS; here I lay out the concept and the design method we use in our projects.

What are AI agents?

An AI agent works in a loop. It receives a goal, such as “follow up on deals in proposal sent with no activity for 15 days”, finds those deals in the CRM, reads each customer’s history, decides what to write and takes the permitted action, such as saving a draft or sending the message. When the action goes past a limit someone set, it stops and asks for approval. Underneath, the agent runs on a market language model, and the platform it runs on determines which data and actions it can reach.

Four verbs sum up the definition: read, decide, act and ask for approval. If acting is missing, you have an assistant that writes well and hands the work back to a person. If deciding is missing, you have an automation that follows the path someone drew. Whenever a tool presents itself as an agent, ask what it reads, what it does on its own and at what point it calls a person.

That question protects the budget. In a Gartner press release from June 2025, the firm estimated that only about 130 of the thousands of vendors presenting themselves as agentic AI are real, and called “agent washing” the practice of rebranding assistants, chatbots and robotic process automation as agents without substantial agentic capabilities. In the demo, ask to see the agent decide and act on data similar to yours.

What is the difference between a chatbot, a copilot and an AI agent?

The four terms get mixed up in vendor presentations, and each one calls for a different decision from leadership. The table uses a single scenario, the customer who asked for a proposal and went quiet.

	Chatbot	Copilot or assistant	Traditional automation	AI agent
What it does	Answers questions within a script or answer base	Suggests text, summaries or analysis when someone asks	Runs a fixed sequence when a trigger fires	Reads the context, decides the next step and acts in the system
Who decides the next step	The script	The person who asked	Whoever designed the flow	The agent, within the rules and approvals
With the customer who went quiet	Replies if the customer writes again	Writes a follow-up when the rep asks	Sends the same email to every deal idle for 15 days	Reads each deal’s history, writes a message for each one and asks for approval before sending
When it makes sense	Repeated customer questions	Individual analysis and writing	Predictable tasks with clear rules	Repeated tasks that require judgment on each case

In the same release, Gartner recommends using agents when decisions are needed, automation for routine workflows and assistants for simple retrieval. I apply that filter: routing leads by region or creating the onboarding task when the contract is signed fits in a regular automation, cheaper and more predictable. The agent comes in when each case needs reading, such as classifying a ticket from what the customer wrote, and the two technologies work together: in the workshop, the automation flow triggered the triage agent on every new ticket.

Anatomy of an AI agent: the parts leadership defines

Building an agent inside a CRM has become quick: you describe the task in a conversation and the tool assembles the agent. The work that requires someone who knows the operation lies in answering the questions below before switching anything on, and that is the stage where I see most agents fail. In the workshop I showed an agent structured into inputs, method, rules and output format; in our projects we add approval, owner and measurement.

Part	Question leadership answers	Example in the ticket triage agent
Goal and trigger	What does the agent deliver and when does it run?	Classify every new ticket as soon as it is opened
Context	What company knowledge does it consult?	Product knowledge base and the definition of each category
Data and permissions	What records does it read and where does it write?	Reads the conversation and the company’s tickets; writes four fields
Rules and format	What does it follow and how does it return the result?	Only the categories on the list, with the customer’s sentence as justification
Human approval	What does it do alone, what waits for approval and what does it never do?	Classifies alone; never replies to the customer on a critical ticket
Owner	Who answers for the result and adjusts the instructions?	Support coordination
Logging and measurement	Where is the trail and how do you know it works?	Classifications corrected and time to first response

Goal, trigger and context

An agent with a vague goal, such as “help the sales team”, burns credits and produces text nobody uses. A good goal fits in one sentence with a verb, an object and a criterion, and the trigger says when the agent works: a request, a fixed schedule or an event in the CRM. Context is what someone wrote for the agent to consult, such as what the company sells, to whom, the criteria for each stage and the tone of messages, and in the workshop we called this work phase zero of the implementation. Recent platforms allow separate knowledge repositories, so the triage agent consults only the product base and the negotiation agent only the sales playbook, which reduces the chance of it mixing information and making up answers.

Permissions and rules

Each agent should read what the task requires and write in as few places as possible: the triage agent writes four ticket fields and sends no message to anyone. The rules say what it follows, such as using only the options in a closed list and marking “not mentioned” when information is missing. Objective rules make review easier, because the reviewer knows what to look for.

Approval, logging and measurement

With each leader I design three lists: what the agent does on its own, what it does and waits for approval, and what it never does. Responsibility stays with whoever approved, because no customer accepts “the AI did it” as an explanation, so every action leaves a trail in the record with what the agent read, what it decided and who approved. Measurement combines the agent’s quality, such as drafts approved without edits, with the effect on the operation, such as response time.

AI agents in the CRM: examples by area

The examples below come from the workshop and work with native or custom agents inside the CRM. From one company to another, what changes is the context and the approval threshold.

Sales: proposal follow-up and pipeline report

A common question is whether an agent can be tied to a pipeline stage and used to follow up on sent proposals. It can: the deal passes the agreed period with no activity, and the agent reads the history, writes a message that picks up where the conversation stopped and leaves the draft for the rep to approve. I also showed a sales results agent that returns revenue, conversion by stage and an action plan with deals at risk. That agent only reads, and its risk lies in the data: if the team pushes deals forward to improve the report, the agent analyzes the fiction as if it were the real operation.

Marketing: nurturing with credit control

The nurturing agent rewrites, on each send, an email in the automated journey based on the contact’s data and engagement. Before activating it, the analyst compares the original version with the personalized one. The team sets the percentage of sends that go through the agent, which caps credit spend and keeps a group receiving the standard email for comparison.

Service and CS: ticket triage

The triage agent runs on every new ticket, reads the conversation, the company’s other tickets and the knowledge base, and writes category, priority, sentiment and a summary. In the workshop, it classified a ticket as a product problem, urgent and critical, quoting the sentence in which the customer said they could not issue any invoice. The full design of this use is in AI in customer service.

Agent	Trigger	Reads	Does	Who approves	Metric
Proposal follow-up	Deal idle for X days	Deal history	Message draft	Rep; manager at decision stage	Approved without edits and deals revived
Pipeline report	Manager request	Deals and activities	Report and action plan	Manager validates before use	Gaps against the manager’s review
Personalized nurturing	Journey send	Contact data and engagement	Rewritten email	Analyst approves the template	Response against the standard group
Ticket triage	New ticket	Conversation and knowledge base	Category, priority, sentiment	Does it alone	Classifications corrected

How do you choose the company’s first AI agent?

Leadership decides whether to trust a second agent by looking at the results of the first. For the first one, I look for a task with high volume, so the gain shows up within weeks; raw material that already exists in the CRM, such as emails, transcripts and tickets; mistakes that are cheap to fix before they reach the customer; and a metric the area already tracks. Ticket triage and call summaries usually meet all four criteria. Follow-ups sent to customers in negotiation come later, once the team has learned to review what the machine writes.

Who owns the agent?

Every agent needs a person who answers for the result, adjusts the instructions and decides to switch it off. The owner is usually the leader of the process the agent works in, with technical support from whoever administers the CRM or from the revenue operations team. With IT as the sole owner, the agent ends up correctly configured and out of step with the process; with the leader and no technical support, nobody touches the configuration after the first month.

What does it cost and how do you measure it?

CRM platforms usually charge agent actions in credits, in an amount that varies with the complexity of the request. Before switching it on, multiply the expected monthly volume by the average consumption per action, measured in a test with real cases, and set a monthly cap per agent and who may switch on a new agent; the return against that cost is covered in AI ROI. To measure, compare the area’s metric in the four weeks before the agent with the four weeks after. If the correction rate does not drop after two instruction adjustments, look for the problem in the context or the data before switching tools.

What are the risks of putting AI agents into operation?

The first risk is the agent working on bad data. The duplicated company gets two follow-ups on the same morning, and the contact with an outdated job title gets nurturing written for the role they left. When a person did the work, the error showed up one customer at a time; with the agent, it shows up in every record it touched in that hour. Before switching it on, clean up the fields that agent will read and leave the rest of the database for after the pilot.

The second risk is a sprawl of agents with no owner. When anyone can create an agent in minutes, within a few months the company piles up dozens of them, with similar instructions, different contexts and credits adding up on the same bill, and customers get messages with inconsistent tone and information. The answer is an inventory with the owner, what each agent reads and does, who approves and the month’s consumption, reviewed by leadership every month. The rules about who may create an agent and with what data belong in the company’s AI use policy.

AI agents and team training: what HR and L&D need to consider

Agents change the work of the people beside them: the rep reviews drafts and the support analyst corrects classifications. In the workshop, when someone asked whether the people who build agents need training, the answer was yes, and I added the line I repeat in our projects: the tool became simple, and simple is different from easy. HR recognizes the need through signals in the operation: agents with nobody answering for them, people who approve drafts without reading, credit spend that surprises finance and leaders who cannot say what their area’s agents do. Leaders take part to decide goal, approval, budget and owner; practitioners practice reviewing and correcting on cases from their own area; CRM administrators learn to configure agents, repositories and credit limits. In the following weeks, evaluate the rate of approval without edits and the number of agents with an owner in the inventory.

Roadmap to put your first AI agent to work

This is the roadmap we use with an area’s leadership to take an agent from paper to operation in four weeks.

Week 1: choose the task and fill in the agent sheet. Apply the four criteria and settle on a single agent. Fill in a sheet with the parts of the anatomy and record the metric for the previous four weeks.

Week 2: prepare the context and the data. Write the content the agent consults, such as ticket categories, stage criteria and message tone, in a repository scoped to it. Clean up only the fields it will read.

Week 3: test with real cases and calibrate the cost. Run the agent in test mode on a few dozen recent cases and compare with what the team did. Adjust the instructions, measure the average credit consumption and set the monthly cap and the initial coverage percentage.

Week 4: switch it on with approval and measure. Put the agent into the flow of a small team, with actions that reach the customer on the approval list. At the end of the week, the owner presents approval without edits, corrections and consumption, and leadership decides whether to expand, adjust or switch it off.

What stalls this roadmap is rarely the configuration. It stalls when leadership has not defined who approves what and when the team does not know how to review what the agent produces. That is the work of the in-company applied AI training, built on the company’s process and data, with groups for leadership, for practitioners and for CRM administrators, which delivers the prompt library for each area and the agent architecture, with what each agent reads, what it does on its own and who approves the rest.

Frequently asked questions

What is an AI agent, in a few words?

It is software that receives a goal, reads company data, decides the next step and takes action in systems such as the CRM, asking a person for approval when the rule requires it. It works within the permissions the company has defined.

What is the difference between an AI agent and a chatbot?

A chatbot talks to whoever asks and follows a script or an answer base. An agent acts in the company’s systems, decides what to do in each case and can be triggered by an event in the CRM, such as a new ticket.

Will AI agents replace people on the team?

Agents take on repeated tasks, such as triage, first drafts and reports, and people’s work comes to include reviewing, correcting and deciding. In the projects I follow, the time freed up goes to conversations with customers and to maintaining the context the agents consult.

How much does an AI agent in the CRM cost?

It depends on the platform and the volume. On platforms that charge by credit, the math is monthly volume multiplied by average consumption measured in a test, with a monthly cap per agent set by leadership.

Who should own an AI agent?

The leader of the process the agent works in, with technical support from whoever administers the CRM or from the RevOps team. The owner answers for the result, adjusts the instructions and decides whether to expand or switch off the agent.

How does HR know the team needs training on AI agents?

The signals show up in the operation: agents with no known owner, approvals without reading and credit spend beyond forecast. Training involves leaders, the people who work with the agents and those who configure them, and the evaluation tracks corrections in the following weeks.