An AI Assistant for Customer Support: Your Helpdesk's AI Add-On, a Chatbot Platform, or a Custom Assistant on Your Own Data

An AI assistant for customer support answers your clients' questions from your help centre and can take simple actions for them, such as looking up an order. It can also draft replies for your support team to check. There are three ways to get one. Most helpdesks, the software your team uses to handle tickets and chats, now sell AI features of their own. A separate AI agent platform can connect to the helpdesk you keep. A firm can also build a custom assistant on your own data and systems. Vendors charge per resolution, per conversation, per action or per seat. The right route depends on what the assistant must do and where your data may go.

What do we get from our helpdesk's own AI features?

Helpdesk vendors sell two kinds of AI. An AI agent answers clients directly, in chat or by email. A copilot helps your support team inside the helpdesk, for example by drafting replies.

Zendesk's pricing page says: “Zendesk AI agents are included in every Suite and Support plan, with pricing based on the successful outcomes they deliver.” Intercom's says: “All plans include access to Intercom and Fin AI Agent.” Fin is Intercom's AI agent. All sources in this article were opened on 5 November 2026.

This route reuses what you already run. The AI agent answers from the help centre articles in the same tool, and when it cannot help, it passes the conversation to your team in the same inbox.

In return, you have less choice: the vendor usually picks the model that writes the answers, a large language model (LLM) of the kind behind ChatGPT. Intercom's article on model training says “customers aren't able to select which AI models Fin uses to provide the Services.” Your instructions for the AI agent also live in the vendor's tool. This route fits when most answers are already in your help content and your compliance team accepts where the vendor processes data.

When does a separate AI agent platform make sense?

Some AI agent and chatbot platforms are sold on their own and connect to your existing helpdesk. Intercom offers Fin to teams that keep another helpdesk: “Use Fin with your current helpdesk including Salesforce and more.” Zendesk's pricing page offers AI agents from Forethought, which it says work “whether you use Zendesk or another service provider”.

A platform fits when you want to keep your helpdesk but its own AI agent fails your test. It also fits when you run more than one helpdesk and want one assistant across them.

The setup gets more complex: one more vendor processes your conversations under its own terms, and the handover to your team crosses two systems. Test the handover before launch. A client whose history is lost on the way has to explain everything again.

When is a custom assistant on our own data worth building?

A custom assistant is software a firm builds for you around a language model. For each question, it searches your own documents and passes the matching passages to the model, a method called retrieval-augmented generation (RAG). It reaches your other systems through their APIs, the interfaces one program uses to talk to another.

Building is worth it when at least one of these is true:

• Answers depend on systems no vendor connects to, such as your trading platform or your own back office
• Actions follow rules that a vendor's setup screens cannot express, such as limits that depend on a client's account type
• Your rules say client conversations must stay on servers you control
• At your volume, per-resolution fees cost more than the people and servers to run your own system

Building takes longer, and after launch you are responsible for running the system. The article on who runs an LLM system without a machine learning (ML) team explains who can do that work. Fine-tuning, which means training a model on your own tickets, is usually not needed, as the article on fine-tuning or RAG shows.

How do vendors charge for AI in customer support?

Vendors use five billing units. Prices below are in US dollars, from each vendor's public pricing page on 5 November 2026:

• Per resolution: Zendesk's Suite Team plan costs $55 per support agent per month billed yearly, or $69 billed monthly, and includes 5 automated resolutions per support agent per month. Suite Professional costs $115 billed yearly or $149 billed monthly, and includes 10. Further resolutions cost $1.50 each with a committed volume, or $2.00 pay-as-you-go
• Per outcome: Intercom charges $0.99 per Fin outcome on top of its seats, the monthly licences for each member of your team. A billable outcome includes a resolution, when “No further help is requested after Fin's last answer”. It also includes a Procedure handoff, when “Fin completes a Procedure you've configured to hand off to a human.” A Procedure is Intercom's name for a workflow you set up for Fin. With another helpdesk, Fin needs no Intercom seats, and the minimum is 50 outcomes a month
• Per conversation or per action: Salesforce lists Agentforce, its AI agent, at $2 per conversation. Its other option is Flex Credits, sold at $500 per 100,000. Each action, such as updating a record, uses 20 credits, so an action costs 10 cents
• Per seat: copilots for your support team are priced per support agent. Zendesk lists Copilot at $50 per support agent per month billed yearly, for Professional plans and above. Intercom lists Copilot at $29 per support agent per month billed yearly, or $35 billed monthly

Compare the definitions before the prices. Under Intercom's pricing, a handover you have set up as a Procedure is billed like a resolution. Zendesk bills only requests resolved “without any escalation to a human agent”. Ask each vendor to price one month of your real volume, by type of conversation, with the minimums included.

What changes when the assistant takes actions?

To answer from your help centre, the assistant needs good articles and little else, since it only reads what you have published. Actions are different. An order lookup reads data from your systems, and a refund or an address change updates it. Each action needs:

• A connection to the system, through an API the system offers or one someone builds
• A check of who the client is before the assistant reads or changes their account, such as asking them to sign in
• An account for the assistant in that system, with rights to the actions on its list and nothing more
• A person's approval for anything that moves money or changes a client's record, enforced in the system that runs the action

Zendesk lists AI agent integrations that “Connect to external systems to read, write, parse data, and trigger actions.” Intercom says Fin “Takes action on external systems”. The rights and approval rules in your own systems still decide what the assistant can do. The article on AI agents and human approval, linked above, explains which actions need a person.

Where are conversations and help content processed, and does the vendor train on them?

Every route sends client conversations and help content to a language model. Check where that data is stored and processed, and whether anyone trains a model on it.

Zendesk's AI Trust page lists regional data hosting in the US, the UK, Ireland, Germany, Japan and Australia. For the third-party LLMs it uses, the page says: “Customer data is never used for model training”. Its page on generative AI adds that Zendesk “is one of the first companies to access OpenAI's European endpoints”, so ask which model provider and region serve your account. Zendesk's own models “may be trained on customer data to the extent instructed by the customer.”

Intercom offers data hosting in the EU or Australia, besides the US, to new workspaces on its Advanced and Expert plans under a contract. Its product terms say “Customer Data will be sent to such Third Party AI Providers”, meaning the model companies it uses. The terms add that Intercom “contractually restricts” them from training on it.

Intercom does train its own Fin models. Its help article of 8 September 2026 says they are trained on data with personal identifiers removed, taken “from production Fin conversation logs”. That data can include “any knowledge content/articles retrieved by Fin”. You can request an opt-out in the workspace settings, which Intercom processes within 30 days. Workspaces hosted in the EU or Australia are excluded automatically.

If your regulator or your own policy says client conversations may not be processed outside servers you control, a vendor-hosted helpdesk AI does not meet that rule. The assistant then needs a model on your own infrastructure. The article on a private ChatGPT on your own servers explains what that takes.

How do we measure resolution honestly?

You answer for what the assistant tells clients. In a small claims decision of February 2024, British Columbia's Civil Resolution Tribunal ordered Air Canada to pay a passenger who had relied on its website chatbot. The chatbot suggested a lower bereavement fare could be claimed after buying a ticket at the regular fare, which the airline did not allow. The tribunal wrote: “It should be obvious to Air Canada that it is responsible for all the information on its website. It makes no difference whether the information comes from a static page or a chatbot.”

A vendor's resolution count is what it bills you for. Intercom's, quoted above, rests on the client not asking for more help. Zendesk looks at conversations where the client did not ask for clarification or for a human. It bills one only if an LLM reading the text confirms the request “was satisfactorily resolved”. In messaging, Zendesk ends a conversation two hours after the first message by default. A client who gives up and phones you the next day may still count as resolved.

Track your own measures for each type of question:

• Repeat contacts: the same client coming back about the same issue within a set period, such as seven days, on any channel
• A weekly sample of conversations counted as resolved, read by your own staff against your help content
• Client ratings of conversations the assistant handled, next to the ratings your team gets
• Handovers, with the reason for each and whether your team had to ask the client to start again

Before launch, test the assistant on your own past tickets, with pass marks agreed in advance. The article on testing an AI assistant, linked above, sets out how.

How should the assistant hand a conversation to a person?

Make the way to a person easy to find, and pass the whole conversation along, so your team sees what the client was told. Write down which cases always go to a person, such as complaints and disputed payments. Outside your team's working hours, the assistant should say when someone will reply.

Do we have to tell clients they are talking to AI?

In the EU, Article 50 of the AI Act, the EU's law on artificial intelligence, covers AI systems that interact directly with people. They must be designed so that people “are informed that they are interacting with an AI system”, unless that is obvious. The information must come “at the latest at the time of the first interaction or exposure”. The European Commission's AI Act Service Desk lists 2 August 2026 as the date Article 50 started to apply.

With a vendor, check how its AI agent introduces itself. Zendesk says its AI agents “are identified to end-users directly in Zendesk's messaging interface”, while Intercom's product terms require its customers to make sure people are informed. With a custom build, agree the wording with your lawyers.

What do we own, and can we switch vendors later?

Keep these on storage your company controls, since each carries over to the next vendor or build:

• Your help content, the articles the assistant answers from
• Your prompts, the written instructions the assistant follows, such as procedures and escalation rules
• Your test cases, with the expected answers your staff wrote
• Conversation logs, including the sources the assistant used and the actions it took

In Intercom's product terms, your input includes help centre content and “any training content for an AI Product”, such as procedures and escalation rules. Input and output “are Customer Data under the Agreement”. Check the same clause in every contract.

Conversation history is the hardest part to move. Intercom's help article on moving to another hosting region says: “Historical conversation data cannot be migrated to the new workspace. It can only be exported for record-keeping”. Export logs on a schedule while you are a customer.

With a custom build, the contract should also give you the code and name any components the builder keeps.

What should we ask a helpdesk or AI platform vendor?

• What exactly counts as a billable resolution or outcome, and is a handover ever billed?
• Which model providers process our conversations, and in which region?
• Does anyone train a model on our conversations or help content, and how do we opt out?
• Which actions can the assistant take in our systems, and where are approvals enforced?
• Can we export all conversations and logs, with the sources behind each answer, at any time?
• When you change the model, how much notice do we get to rerun our test cases?

What should we ask a firm that would build it?

• Which of our systems will the assistant read from, and which will it change?
• Where will the model run, and what data leaves our servers?
• How will you test it on our past tickets, and who on our side signs off?
• How does the handover to our team work, and how does the assistant tell clients it is AI?
• Will we own the code, prompts, test cases and logs at the end?

Where does amBrain fit?

amBrain publishes this blog. On its own product, it says: “Our own product, shyash.ai, runs on AI agents that research, write, check and measure from one shared memory of company facts. A person approves everything that goes out.”

This article is not a case study and describes no client work. It does not claim that amBrain has built a support assistant for any client. The line above is about amBrain's own product, and the examples are general.

The companies and public bodies named here are cited from published pages. The article claims no relationship between amBrain and any of them and gives no amBrain prices or timelines.

amBrain has been building software since 2019. On how it works with clients, it says: “Three formats: full delivery, a dedicated team, or engineers embedded in your team.” Its line on ownership is “The client keeps full ownership of the product and the code, except our reusable components.” Ask for the list of those components by name before you sign.

This article is not legal advice either. Ask your lawyers how the AI Act and privacy law apply to your assistant. The questions above apply to any firm you consider.