Generative AI creates new content, such as text, images or code, in response to a prompt. Agentic AI works towards a goal: it plans the steps, uses tools and other software, and takes actions until the job is done or it needs help. Today's agentic systems are built around a generative model: the model understands the request and writes the words, and the agentic layer decides what to do next and does it.
The difference is easy to see in customer service. Ask a generative AI chatbot for a refund and it can explain your refund policy. Ask an agentic AI and it can find the order, check your rules and arrange the refund, or pass the case to a human agent when the rules say no. This guide explains the difference, follows one refund request both ways, and covers what should make each decision once the AI can act.
In short
- Generative AI produces content and stops; someone else acts on it. Agentic AI keeps going: it looks things up, takes actions and checks the result until the goal is met.
- They aren't rivals. An agentic system uses a generative model for language and adds tools, memory and rules around it.
- In customer service, the test is whether the AI can see the order and change it. A generative chatbot answers; an agentic AI completes the request.
- The model shouldn't make every decision. For refunds, cancellations and exchanges, written rules should decide eligibility and amounts, and the model should handle the conversation.
Agentic AI vs generative AI at a glance
| Generative AI | Agentic AI | |
|---|---|---|
| What it's for | Creating content: text, images, code, summaries | Reaching a goal that takes several steps |
| What starts it | A prompt | A goal, a request or an event, such as a new message or a delayed parcel |
| How it works | One response per prompt | A loop: plan a step, use a tool, check the result, choose the next step |
| Access to other systems | None, unless someone builds it in | Reads and writes through APIs: order system, courier, payments, helpdesk |
| Memory | The current conversation | The task's progress across steps, and often across conversations |
| Who acts on the output | The user copies, sends or approves it | The system acts, within the permissions it's given |
| How it goes wrong | It says something false or out of date | It does something wrong: the wrong action, on the wrong record, at the wrong time |
| What you check | The answer | The answer, the action and the rule behind it |
| Customer service example | Drafts a reply explaining the returns policy | Checks the order, books the return pickup and confirms it to the customer |
How agentic AI uses generative AI
The generative part of both is usually a large language model (LLM). On its own, an LLM takes text in and gives text out. It can't see your order system, and it doesn't carry a task forward once the reply is written.
An agentic system wraps that model in a loop with three additions:
- Tools: actions it's allowed to call, such as "find order", "check stock" or "create return".
- Memory of the task: what it has checked, what the customer has sent and what's still to do.
- A stopping point: the goal is met, or the case goes to a human agent.
On each turn, the model reads the situation and proposes the next step, the system runs the tool, and the result goes back to the model. The same kind of model sits in both. What changes is what surrounds it and what it's allowed to do.
That's why the line between the two keeps moving. ChatGPT is generative AI, but when it searches the web before answering, it's already behaving a little agentically. "Agentic" is a matter of degree: how many steps, how many tools, and how much the system may do without asking. If you're comparing older bots with language models rather than with agents, our guide to chatbot vs conversational AI covers that step.
One refund request, two ways
At 10 pm, a customer messages a footwear brand on WhatsApp: "The running shoes I got are too small. Can I swap them for a size 9? If you don't have a 9, I'd like a refund." The shoes were delivered five days ago, and she paid by card.
The brand's policy is simple. Unworn shoes can be returned within 15 days of delivery. Exchanges are free if the new size is in stock. Refunds go back to the original payment method, for the price paid; the delivery fee isn't refunded. A refund starts once the courier has collected the pair.
With a generative AI chatbot. It has read the brand's help-centre articles, so its first reply is friendly and correct: exchanges and returns are accepted within 15 days. It asks for her order number so the team can check size 9, and she sends it. The chatbot can't see the order, the stock or the courier, so it tells her an agent will reply. The next morning, a human agent opens Shopify, finds the order, sees there's no size 9, writes to her, waits for her answer, books the pickup and notes the refund. The AI wrote good sentences. A human agent did every step of the work, a day later.
There's a riskier version. She asks, "Will I get the delivery fee back too?" If that line of the policy isn't in what the chatbot has read, a language model may still answer, and answer confidently: "Yes, you'll get a full refund." To her, that's now your policy.
With an agentic AI. The same message sets off a sequence of checks and actions:
| Step | What the agentic AI did | What made the decision |
|---|---|---|
| Understood the request | Size 8 is too small; she wants a 9, or a refund if there's no 9 | The language model |
| Found the order | Matched her WhatsApp number to one order: delivered five days ago, size 8, paid by card | Order data in Shopify |
| Checked eligibility | Day 5 of 15, not a final-sale item; asked her to confirm the shoes are unworn, and she did | The return rules |
| Checked stock | No size 9 available | Inventory data |
| Chose the outcome | A refund, because the exchange isn't possible | The rules: the outcomes the brand allows, in order |
| Worked out the amount | The price she paid, without the delivery fee | The rules, applied to the order |
| Acted | Created the return and booked a courier pickup for Thursday | Shopify and the courier's system |
| Confirmed | Told her the pickup day, the refund amount and when the refund starts | The language model, writing the facts it was given |
The model did two jobs: it understood her, and it wrote to her. Every step that could cost money or become a promise was decided by a rule working on live data. Had it been her third return this month, or had she asked for the refund to go to a different card, the same rules would have sent the case to a human agent with the order and the checks already attached.
That's the difference that matters to a support team. The generative chatbot answered; the agentic AI completed the request, which is what customer service automation should mean. The same pattern covers cancellations, exchanges, address changes and damaged items, the everyday work of returns management.
Flowcall's AI Agent handles requests like this the agentic way. Returns, refunds and exchanges are ready-made workflows it follows step by step, on WhatsApp, Instagram, Messenger, Live Chat and email. With Flowcall's Shopify integration, it checks the customer's order, delivery and payment and completes the refund or exchange when your rules pass. When they don't, it creates a ticket with the conversation, a summary, the order and what's left to do, and routes it to the right team.
The model talks, the rules decide
Agentic AI is usually described by its autonomy. In customer service, autonomy isn't the goal. A customer asking for a refund doesn't want the AI to be creative; she wants the answer your policy gives, every time. So the useful question is which part of the system makes each decision.
| Decision in a refund or exchange | Who should make it | Why |
|---|---|---|
| What the customer wants, in her words and language | The language model | Understanding free text is what generative AI does well |
| Which order it's about | A match on phone number, email or order number | A guess here can show one customer's order to another |
| Whether the item qualifies (window, condition, product) | A written rule | The same answer every time, and one you can change in one place |
| Which outcome to offer (exchange, refund, store credit) | A written rule, choosing from outcomes you allow | The AI shouldn't offer something your policy doesn't |
| The amount | A calculation from the order | Money is never a creative task |
| How to say it | The language model, in your brand's voice | Tone, clarity and the customer's language |
| Goodwill, suspected abuse, a very angry customer | A human agent | Judgement your rules can't write down |
Vendors call the rules layer different things: guardrails, business rules, procedures, deterministic decisions. The test is the same for all of them. When a customer pushes ("my friend got a full refund"), does the answer come from the rule, or from the model's wish to be helpful? Writing the rules down is most of the work: a clear refund policy for customers, and a customer service standard operating procedure (SOP) for each request type that says how your team applies it. The last row matters as much as the others: the cases the rules send on need a clear owner, which is the job of escalation management.
We built Flowcall's AI Agent this way. It runs on the SOP Engine: the AI follows your SOP step by step while talking naturally, your rules make each decision, and order, delivery and payment details come from your systems. Eligibility checks decide what the customer qualifies for, and the AI picks from the outcomes you allow instead of improvising values. Like code, a workflow that's set up correctly gives the same result on every run, so the AI doesn't invent a refund amount or a return window, and a fix stays fixed.
Where each goes wrong, and what prevents it
Generative AI goes wrong with words: an invented policy, an out-of-date answer, a promise nobody can keep, all delivered in the same confident tone as a correct answer. Customers take those words as yours. In February 2024, a Canadian tribunal held Air Canada responsible for its website chatbot, which had told a customer he could apply for a bereavement fare retroactively, contradicting the airline's own policy page. The tribunal rejected the argument that the chatbot was responsible for its own statements: the airline is responsible for everything on its website, chatbot included (Moffatt v. Air Canada, 2024 BCCRT 149).
What prevents it: answers grounded in your current policies and order data, a knowledge base that someone keeps up to date, and a handoff when the AI doesn't know instead of a guess.
Agentic AI goes wrong with actions: a refund on an order that didn't qualify, a cancellation after the parcel shipped, a change on the wrong customer's order, a loop that never finishes, or a "done" when the action failed. Put simply, generative AI can say something wrong; agentic AI can also do something wrong.
What prevents it:
- Actions only through rules. The AI can take only the actions you list, and only after the required checks pass.
- Limits that route to a human agent. Refunds above an amount, repeat claims and anything your policy calls an exception.
- Test conversations before launch, and again after every change to a rule, a prompt or an integration.
- A record of every action and its result, so you can see what happened on any conversation and why.
- Review of what the AI did, not only what it said. A polite conversation can still hide a wrong refund, so customer service quality assurance has to cover actions too.
We built Flowcall's AI Agent to be tested before it talks to a customer. You run simulated customer conversations against a workflow before it goes live and rerun them after every change. For any real conversation, you can see which workflow ran, what data it used and which action it tried. And Flowcall's AI Auto QA checks what the AI Agent said and did: wrong, incomplete or made-up answers, loops, missed or late handoffs, and promises or actions it wasn't allowed to make.
Generative AI still has a job: helping human agents
Agentic AI doesn't make generative AI redundant in a support team. Generative AI is the right tool wherever a human agent reviews the output before anything reaches the customer or changes a record:
- drafting replies for human agents to edit and send;
- summarising a long ticket or a call before a handover;
- filling in the fields an agent completes when closing a ticket;
- translating a customer's message, or a reply, into another language;
- drafting help-centre articles and saved replies from your policies;
- reading conversations for quality reviews and recurring complaints.
The risk is lower because a human agent is the last step. Our guide to generative AI for customer service covers these uses in more depth.
In Flowcall's AI Helpdesk, this is the Zero-Tab Ticket: everything a human agent needs to answer and act on a case is on the ticket. It shows the customer's orders, delivery and payments from your connected systems, Agent Assist drafts the reply and fills the closing fields, and Pre-Filled Actions fill the refund or replacement form from the conversation. The agent checks the form and clicks once, instead of opening Shopify and copying the details across.
Five questions that show whether a support AI is agentic
The word "agentic" is also used for products that only answer. When you compare AI chatbots for customer service, five questions separate the two in a demo:
- Can it see a real order right now? Ask about a test order's status, items and payment. A generative-only bot answers from policy pages.
- Can it change something? Ask it to cancel an unshipped test order or change the address, then check the store admin yourself.
- Who decides the refund: the model or a rule? Ask what stops it refunding an order outside the return window, or above a set amount, and ask to see where that rule is written.
- What happens when the rule says no? Ask to see the ticket a human agent receives. It should carry the order, the checks the AI ran and the reason it handed over.
- Can you see what it did, and test changes before customers do? Ask for the record of one conversation, step by step, and how a change to a rule is tested before it goes live.
A product that fails the first two is generative AI with a chat window. That can still be useful, but it won't take order, return and refund work off your team.
Which one does your support team need?
| Choose | If |
|---|---|
| A generative AI chatbot | Most questions have answers in your policies and FAQs, order changes are rare, and your order data can't be reached through an API |
| Agentic AI for customers | A large share of messages are order changes, returns, exchanges and refunds, your orders sit in Shopify or another system with an API, and you can write your rules down |
| Generative AI for your agents | Human agents handle most conversations, and you want faster replies and less typing |
| Both | You handle a high volume of order work: agentic AI completes the routine requests, and generative AI helps human agents with the rest |
If your queue is mostly orders, returns and refunds, you're in the last row: the requests that fill an ecommerce customer service queue are actions, not questions. Start with one request type that is frequent, follows a rule you can write down and sits in a system the AI can reach, then add more after testing. Our guide to agentic AI for customer service covers where to start and how to roll it out, and the wider picture of AI customer service covers the rest of the stack.
How Flowcall combines them
Flowcall uses generative AI for language and rules for decisions. The AI Agent understands customers on WhatsApp, Instagram, Messenger, Live Chat and email, checks their orders, applies your rules and completes the refund, exchange or cancellation. Anything the rules send on goes to the AI Helpdesk with the work done so far. The same AI Agent helps shoppers before they buy, finding and comparing products from your catalogue. In the AI Helpdesk, generative AI drafts replies and fills forms for human agents, and AI Auto QA reviews what the AI and your team said and did.
To see Flowcall's AI Agent check an order, apply return rules and complete a refund on WhatsApp, book a demo. Qualifying brands get a free proof of concept on the service problem they most need solved, built with their team and tested before anything goes live.
Frequently Asked Questions
Is ChatGPT agentic AI or generative AI?
ChatGPT is generative AI: a chatbot built on large language models that writes answers to prompts. When it searches the web or uses other tools to finish a task, it behaves agentically, which is why the line between the two keeps moving. For customer service, the label matters less than whether the AI can reach your order systems and follow your rules. Our guide to ChatGPT for customer service covers what it can and can't do for a support team.
What are examples of agentic AI?
In customer service: cancelling an order that hasn't shipped, changing a delivery address, booking a return pickup, creating an exchange, starting a refund within set limits and confirming a cash-on-delivery order on WhatsApp. Outside support, examples include coding assistants that write and test code, IT systems that reset access and investigate alerts, and travel tools that book within a budget.
Is agentic AI the same as an AI agent?
Nearly. An AI agent is a single system that uses tools to complete a task. Agentic AI describes the approach, and a system may combine several agents. Our guide to agentic AI vs AI agents separates the two terms. Vendors use both loosely, so judge an AI agent by two questions instead of the label: what can it see, and what is it allowed to change?
Does agentic AI hallucinate?
The language model inside it can, so the design decides whether that matters. When the model handles the conversation and written rules make the decisions on live data, the refund amount and the return window come from your systems and your policy, not from the model. Test each workflow on simulated conversations before it goes live, and review what the AI did as well as what it said.
Will agentic AI replace human agents?
Not for a consumer brand with real exceptions. Agentic AI takes over the requests your rules can settle, which changes what human agents spend their day on: exceptions, judgement calls and upset customers. We look at the wider question in will AI replace customer service.
