AI Agent vs Chatbot for Customer Support: Costs, Handoffs and Rules

Key takeaways

• Intercom bills Fin at $0.99 per outcome, one per conversation even when Fin takes several actions, with a 50-outcome monthly minimum when Fin runs on another helpdesk; a conversation simply passed to your team is not billed (Intercom pricing page, checked November 7, 2026).
• Zendesk includes 5 automated resolutions per agent a month on Support Team and Suite Team and 10 on Suite Professional, then bills $1.50 committed or $2.00 pay-as-you-go; Salesforce lists Agentforce at $2 per conversation, or $500 per 100,000 Flex Credits at 20 credits per action (vendors' pricing pages, checked November 7, 2026).
• On February 14, 2024, a British Columbia tribunal held Air Canada liable for its website chatbot's wrong advice on bereavement fares and ordered it to pay C$812.02, ruling: "It makes no difference whether the information comes from a static page or a chatbot" (Moffatt v. Air Canada, 2024 BCCRT 149).
• In a Gartner survey of 3,566 customers, published August 4, 2026, 87% said an option to reach a human agent is essential when companies use GenAI for service; Gartner's September 2, 2026 release found only 27% would try a chatbot again after a negative experience.
• Article 50 of the EU AI Act has applied since August 2, 2026: people must be told they are dealing with an AI unless that is obvious. In California, SB 1001 has barred using a bot to mislead people about its artificial identity to push a sale since July 1, 2019; disclosing the bot removes liability under that law. From January 1, 2027, California's AB 1609 also requires businesses with more than $500 million in revenue to disclose support bots that could pass for a person and to offer a human on request.

A chatbot answers customer questions. An AI agent also acts on them: it checks the order, moves the booking, files the ticket and hands the case to a person when it should. Support platforms now bill AI per result or per conversation: as of November 7, 2026, Intercom charges $0.99 per Fin outcome, Zendesk $1.50 to $2.00 per automated resolution and Salesforce lists Agentforce at $2 per conversation. Whichever you pick, keep a person in reach: 87% of customers in a Gartner survey call that essential.

This guide is for founders and support leads in the US and the UAE choosing between the AI in their helpdesk and an agent built for them. It covers prices, ownership, CRM connections, handoffs, public failures, data and disclosure rules, and what to measure.

AI agent vs chatbot: what is the difference?

A scripted chatbot walks customers through buttons and preset flows. An AI chatbot understands free text and answers from your help center and policies. An AI agent also acts: it calls your order system, CRM or booking tool through their APIs, checks the result and decides whether to finish or hand off to a person.

Scripted chatbot, AI chatbot and AI agent: what each does and where it fits.

If your customers only need answers, an AI chatbot on your knowledge base is enough. An agent pays off when many requests need something done in another system: where is my order, change my delivery address, move my appointment.

AI automation vs AI agents is a similar split: automation follows fixed steps (a form comes in, a ticket is created, an email goes out), while an agent chooses its steps. Most support setups mix both: rules route and escalate, the agent answers and acts.

What does AI customer support cost in 2026?

The big helpdesks price AI per result or per conversation, and how each defines that unit shapes your bill more than the headline price.

List prices for AI support from each vendor's pricing page, checked November 7, 2026. Market figures, not our price list.

• Fin bills one outcome per conversation "even if Fin takes multiple actions" and says: "You're not charged when a conversation is simply passed to your team without an outcome."
• Zendesk bills only requests "successfully resolved by the AI agent, without any escalation to a human agent." Under the tier system Zendesk introduced on May 18, 2026, an LLM checks AI conversations: those that pass count as a Verified resolution, while a Contained resolution or an Assisted escalation "does not count against your resolution allowance."
• Agentforce prices the conversation but does not define one on the page, so ask whether handed-off conversations are billed. Salesforce says its prices are subject to change.

A worked example at list price: 10 support agents, and the AI fully resolves 1,000 conversations a month. Fin bills $990, plus any procedure handoffs and your helpdesk seats. Zendesk Suite Team includes 50 resolutions (5 × 10 agents) and bills the other 950 at $1,425 committed or $1,900 pay-as-you-go, plus $550 for the seats. Agentforce lists $2,000, plus any handed-off conversations if Salesforce bills them. None of this covers the time people spend checking answers and updating the help center; our cost guide does.

Platform AI or a custom build: who owns what you build?

Platform AI is the fastest start when your support already lives in Intercom, Zendesk or Salesforce. The vendor runs the model and updates, and its agent reads your existing help center. The trade-off is ownership: your articles stay yours, but the procedures and tuning you build live inside the vendor's product, and the per-resolution bill grows with volume.

A custom agent, built by your developers or an AI chatbot development company, keeps the parts in your accounts: code and prompts in your repository, the model account and logs in your company's name, and the test set as a file you can run against any model. You choose the model and the data region, and you own the upkeep: when models change, someone re-runs the tests.

A custom build tends to pay off when the platform doesn't support your helpdesk or channels such as WhatsApp and Telegram, when the agent must act across orders, billing and a CRM, when data must stay in a set region, or when per-resolution fees outgrow the cost of running your own.

Start with a platform when volume is modest and your answers live in one help center. Either way, put ownership in the contract: our 12 questions for an AI automation agency include the handover list, and our n8n, Make and Zapier vs custom code comparison covers what to build on.

How does an AI agent connect to your CRM, helpdesk and order system?

Through the same APIs your other tools use. The agent reads and writes where your team already works and usually needs five kinds of access:

1. Know the customer. Find the contact and past tickets by email or phone, and confirm identity before sharing account details.
2. Find the answer. Search your help center, policies and past tickets (RAG), and keep the source with each answer so a person can check it.
3. Act within limits. Call the order, booking or billing system for a short list of actions: status, an address change before dispatch, a cancellation inside the allowed window. Anything else needs a person's approval.
4. Write it back. Save the transcript, a summary, tags and every action to the ticket, so reporting stays in your helpdesk.
5. Hand off. Assign the ticket to the right queue with that summary, and alert a person when a run fails.

Give each API key the least access its job needs: a key that reads orders and changes addresses should not issue refunds. Card and ID numbers stay out of prompts and logs. If a system has no API, use an export or a webhook, or automate a different step first.

The same agent can work in Telegram, as a bot or a Mini App (what a Mini App costs), or in WhatsApp, which in the UAE has its own rules and fees.

Case study · EncoreWebsite, CRM and AI call notes for a Dubai agency3products shipped by one team

Encore, a Dubai PR and events agency, is not a support desk, but the CRM we built for it has an AI assistant that answers questions about the work and opens or closes tasks, and each new task messages its assignee in Telegram. What carries over: AI that reads and writes in a CRM, and alerts that reach people where they work.

What can go wrong? Four public failures

Each points to a safeguard to build in before launch.

Air Canada: the bot's answer is your answer. After a grandparent's death, a customer asked the chatbot on Air Canada's website about bereavement fares. It said they could apply within 90 days of the ticket being issued, but the airline did not allow retroactive claims. In Moffatt v. Air Canada, decided February 14, 2024, the tribunal ruled that "It makes no difference whether the information comes from a static page or a chatbot" and ordered the airline to pay C$812.02. The safeguard: policy answers come from the current policy text, with the source, or go to a person.

DPD: an update broke the bot. In January 2024, a customer got the parcel firm's chatbot to swear at him and write a poem about how terrible the company was. DPD said "An error occurred after a system update" and that the AI element was disabled, the BBC reported; one post about it was viewed 800,000 times in 24 hours. The safeguard: re-run your test set after every change to the model, prompt or vendor settings, including attempts to make the bot misbehave.

Cursor: the bot invented a policy. In April 2025, users of the AI code editor were logged out when they switched devices. A support email from "Sam" called it "expected behavior" under a new login policy. There was no such policy and no human behind the email, Fortune reported; users reported cancelling, and a cofounder acknowledged an "incorrect response from a front-line AI support bot." The safeguard: label AI replies, and when the sources don't cover a question, the agent says so and hands off.

Klarna: cost was not the whole goal. In February 2024, Klarna said its AI assistant had 2.3 million conversations in its first month, two-thirds of its service chats, the equivalent of 700 full-time agents, with a 25% drop in repeat inquiries. In May 2025 it began hiring human agents again. In February 2026 its CEO said: "If AI can do customer service, it means it's going to be the cheap customer service." There has been "too much focus on cost," he added, CX Dive reported. The safeguard: let the AI take volume, and keep people for the cases where the relationship matters.

What should the AI do when it doesn't know?

Stop and pass the case to a person with everything it has collected. Customers accept AI on those terms: in Gartner's survey of 3,566 customers, 50% said their interactions are easier when companies use GenAI, and 87% said an option to reach a human agent is essential. Gartner says GenAI is more helpful when it attempts "resolution only when confidence is high, with a clear path to human support."

Write the handoff triggers down before launch:

• No source. The answer isn't in your help center, policies or the customer's data.
• The customer asks. Any request for a person goes to a person, at once.
• Sensitive topic. Complaints, refunds above a set amount, legal threats, health, safety, anything about a minor.
• Failed action. An API call returns an error or something unexpected.
• A loop. The customer rephrases the same question twice.

The person gets the transcript, a summary and what the agent already checked, so the customer doesn't repeat anything. Gartner's follow-up release says the same: provide "a visible path to a human agent and pass along the information and context already collected." After hours, the agent says when a person will reply and opens the ticket.

Instagram AI AgentClient under NDA

Case study · Under NDAHundreds of Instagram accounts, one AI agent100+accounts run from one Telegram chat

Our closest published case: for a client under NDA, this agent reads comments and DMs across 100+ Instagram accounts and hands anything sensitive to a person. If a run breaks or a DM needs judgment, a person gets a Telegram alert with the step, the account and the reason. It is not a helpdesk bot; what carries over is the handoff and alert pattern.

Is AI customer support safe? Where does the data go?

It is as safe as the access you give it and the terms you agree in writing: which model provider processes conversations and where, what is stored and for how long, and whether the provider may train on your data. Our guide to choosing an AI automation agency has the one-page list to ask for.

The laws, briefly:

• UAE. The federal PDPL applies to companies in the UAE and to companies abroad that process personal data of people in the UAE; companies based in free zones with their own data protection laws, such as DIFC and ADGM, follow those laws instead. Article 18: "The Controller shall include the human element in reviewing automated processing decisions at the request of the Data Subject." Our WhatsApp guide covers consent and transfers.
• EU. If you offer goods or services to people in the EU, the GDPR applies even without an office there (Article 3). The AI vendor processes data under a written contract with you (Article 28), transfers outside the EU must meet Chapter V, and people have the right not to be subject to solely automated decisions with legal or similarly significant effects (Article 22).
• US. State privacy laws apply, each with its own thresholds. California's CCPA gives consumers the right to know what you collect about them, to delete and correct it, and to opt out of its sale or sharing. Chat transcripts can hold personal information, so make them easy to find and delete.

Do you have to tell customers they are talking to a bot?

Increasingly, yes, and it costs nothing to do it everywhere.

• California. Since July 1, 2019, SB 1001 has made it unlawful to use a bot online to mislead a person in California about its artificial identity in order to push a sale. Disclosing that it is a bot, clearly and conspicuously, removes liability under that law.
• California, from 2027. AB 1609, signed on September 28, 2026, takes effect on January 1, 2027 and covers businesses with more than $500 million in annual revenue. They may not present a customer service chatbot as a human, and must say clearly that it is not human when a reasonable person could be misled. During business hours, when a customer asks for a person, they must make a good-faith effort to connect them within 15 minutes or offer an appointment within one business day.
• Maine. A 2025 law bars using an AI chatbot in trade with a consumer in a way that may mislead a reasonable consumer into believing they are talking to a human, unless they are told "in a clear and conspicuous manner" that they are not. Violations fall under Maine's Unfair Trade Practices Act.
• Utah has its own AI disclosure rules; our AI voice agent guide covers them.
• EU. Article 50 of the AI Act has applied since August 2, 2026, per the Commission's implementation timeline. AI systems that interact with people must be built to tell them they are dealing with an AI, unless that is obvious, "at the latest at the time of the first interaction."

Our rule in every market: the agent says it is an AI in its first message, and the customer can ask for a person at any point. Cursor's "Sam" shows what an unlabeled bot costs when it gets something wrong.

How do you launch AI support and measure it?

Start small, test, and compare against a baseline. Gartner found that only 27% of customers would try a chatbot again after a negative experience and advises leaders to "prioritize reliability over reach": begin with "targeted, thoroughly tested deployments" and expand "only after performance is reliable."

Before launch:

• Baseline. Pull the last few months from your helpdesk: volume by topic, first response time, time to resolution, satisfaction score, repeat contacts and the share of requests that arrive after hours.
• Test set. Turn a few dozen past requests, with their right outcomes, into a test the agent must pass before launch. Include angry customers, edge cases and attempts to pull it off topic.

After launch, every week:

• Verified resolution rate. A conversation that ended without a handoff is not always solved; check a sample by hand.
• Repeat contacts on the same issue within a few days. Klarna reported a 25% drop in repeat inquiries.
• Handoff rate and reasons, and the time from handoff to a person's first reply.
• Satisfaction for AI and human conversations, side by side.
• Wrong answers found in the weekly sample, each fixed in the content or the rules and added to the test set.
• Cost per resolved conversation, fees included, against the baseline.

Checklist: before you launch AI customer support

• Scope. The request types the AI takes on day one; everything else goes to a person.
• Platform or custom. List prices at your volume, and how handoffs and unresolved conversations are billed.
• Handoff rules. Triggers, what the person receives, what happens after hours.
• Access. One least-privilege key per system; no card or ID numbers in prompts or logs.
• Data terms. Provider, region, retention and training, in writing before the build.
• Disclosure. The agent says it is an AI in its first message.
• Test set. Passed before launch and re-run after every change.
• Metrics. A baseline before launch, a weekly review after.
• Ownership. Code, prompts, test set and logs in your accounts, or an export path from the platform.

We build AI agents that answer customers, act in your CRM and helpdesk and hand off to a person by clear rules. See our AI automation service or book a call with a partner.

Questions we get asked

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

A chatbot answers: a scripted one through buttons and preset flows, an AI chatbot in free text from your help center and policies. An AI agent also acts: it looks up orders, changes bookings and updates tickets through your systems' APIs, then finishes or hands off to a person. If your customers only need answers, an AI chatbot is enough. AI automation is different again: it follows fixed steps, while an agent chooses its own.

How much does AI customer support cost in 2026?

At list price on November 7, 2026, Intercom charges $0.99 per Fin outcome, Zendesk $1.50 to $2.00 per automated resolution beyond the allowance included with agent seats, and Salesforce $2 per Agentforce conversation. These are market figures from the vendors' pages, not our price list. A custom agent has no per-resolution fee: you pay for the build, model usage, hosting and the time people spend reviewing answers.

Can an AI support agent work with our CRM and helpdesk?

Yes, through the same APIs your other tools use. A typical agent looks up the customer and past tickets, searches your help center, takes a short list of actions such as an order status check or an address change, writes the transcript and a summary back to the ticket and assigns it to the right queue. Give each key the least access it needs, and keep card and ID numbers out of prompts and logs. Without an API, an export or webhook often works.

What happens when the AI doesn't know the answer?

It should say so and hand the case to a person with the transcript, a summary and what it already checked, so the customer doesn't repeat anything. Set the triggers before launch: no source for the answer, a request for a person, a sensitive topic, a failed action or a customer repeating the question. After hours, it says when a person will reply. In Gartner's 2026 survey, 87% of customers called a human option essential.

Is AI customer support safe, and where is the data stored?

It is as safe as the access and terms you set. Agree in writing which model provider processes conversations, in which region, what is stored, for how long, and whether the provider may train on your data. In the UAE, the PDPL lets customers ask for a person to review automated decisions. The GDPR requires a written processor contract with the AI vendor, and California's CCPA gives consumers the right to know, delete and correct what you collect.

Who owns an AI support agent: us, the platform or the developer?

On a platform such as Fin, Zendesk or Agentforce, your help center articles stay yours, but the procedures and tuning you build live inside the vendor's product, so check how you can export conversations. With a custom build, put ownership in the contract: code and prompts in your repository, the model account and logs in your company's name, and the test set as a file you can run against any model.

Sources

1. Fin pricing · Intercom, 2026
2. Zendesk pricing · Zendesk, 2026
3. About automated resolution tiers · Zendesk, 2026
4. Agentforce pricing · Salesforce, 2026
5. Moffatt v. Air Canada, 2024 BCCRT 149 · Civil Resolution Tribunal of British Columbia, 2024
6. DPD error caused chatbot to swear at customer · BBC News, 2024
7. A customer support AI went rogue—and it's a warning for every company considering replacing workers with automation · Fortune, 2025
8. Klarna AI assistant handles two-thirds of customer service chats in its first month · Klarna via PR Newswire, 2024
9. Klarna pursues 'Uber' style customer service model · Customer Experience Dive, 2026
10. Gartner Survey Finds 87% of Customers Say Companies Using GenAI for Customer Service Must Provide Access to a Human Agent · Gartner, 2026
11. Gartner Survey Finds Only 27% of Customers Would Try a Chatbot Again After a Negative Experience · Gartner, 2026
12. SB 1001: Bots: disclosure · CalMatters Digital Democracy, 2018
13. AB 1609: Customer service chatbots · CalMatters Digital Democracy, 2026
14. Title 10, §1500-DD: Required disclosure of use of artificial intelligence chatbot to engage in trade and commerce · Maine Legislature, 2025
15. Article 50: Transparency obligations for providers and deployers of certain AI systems · European Commission, AI Act Service Desk, 2026
16. Timeline for the implementation of the EU AI Act · European Commission, AI Act Service Desk, 2026
17. Federal Decree by Law No. (45) of 2021 Concerning the Protection of Personal Data · UAE Legislation, 2021
18. Regulation (EU) 2016/679 (General Data Protection Regulation) · EUR-Lex, 2016
19. California Consumer Privacy Act (CCPA) · State of California Department of Justice, 2026

Share this article

About the author

Viacheslav Kasatkin

Co-founder, Head of Business Strategy

Sets direction on fintech, AI and business strategy.

Meet the leadership