By KYN AI Advisory Team — AI implementation specialists, Singapore
Intercom Fin is an AI support agent that works inside Intercom's helpdesk and messenger. A WhatsApp-first support agent starts from the channel your customers already use and works from your own records. If your customers write to you on web chat and email, and you use Intercom, Fin is the shorter path. If most of them write on WhatsApp, as in much of Southeast Asia, build the agent around WhatsApp and connect the helpdesk second.
What does Fin do?
Intercom describes Fin as an AI agent that answers customer questions using your help content, and that can hand over to a human in the Intercom inbox. See Intercom's Fin page and pricing page for current features and terms. We do not quote the pricing here because the model changes.
Intercom also publishes guidance on evaluating AI agents. Its questions are worth reading whichever product you choose. Our own list is in how to evaluate an AI agent vendor: 15 questions.
What is a WhatsApp-first agent?
It is a support agent whose main channel is WhatsApp. It reads the incoming message, finds the customer in your records, answers from approved sources, and hands the conversation to a person when it should. It uses the WhatsApp Business Platform, which differs from the WhatsApp Business app. See WhatsApp Business app vs platform and the cost note in WhatsApp Business API cost in Singapore.
Where do customers actually write?
Start there. Count your enquiries from the last 90 days by channel: WhatsApp, email, web chat, phone, social messages. If WhatsApp is the majority, a helpdesk built around web chat is the wrong centre.
WhatsApp matters in the markets KYN serves. A customer in Jakarta or Kuala Lumpur often expects a reply in the same thread they used last time. That thread also carries photos, voice notes and order details. A good agent reads all of it, and a good handover passes the whole thread to the person.
How do they compare?
| Question | Intercom Fin | WhatsApp-first custom agent | |---|---|---| | Home channel | Intercom messenger, email, other connected channels | WhatsApp | | Knowledge source | Help content and connected sources | Your records, documents and rules | | Customer lookup | Intercom contact data | Your database: orders, quotes, accounts | | Handover | To the Intercom inbox | To a named person, with the thread and the context | | Language | Check supported languages | Built for your languages, including Bahasa | | Setup | Short | Weeks to a few months | | Cost shape | Per Intercom terms | Build fee, then monthly retainer | | Record | Intercom | Your cloud account |
What should the agent answer, and what should it not?
The same rule applies to both. Split questions into three groups.
- Answer alone. Opening hours, order status, document requirements, how-to steps. The answer exists in an approved source and cannot hurt the customer if slightly off.
- Answer and flag. Delivery estimates, refund eligibility, anything with a date or a number. The agent answers from a record and flags it for review.
- Hand over. Complaints, legal threats, anything about money owed, anything the customer repeats twice. A named person takes it.
Anything about money, filings or deletions should wait for a person. See approval gates for high-risk agent actions.
How do you test a support agent before launch?
- Pull 200 real past conversations.
- Have a support lead mark the correct outcome for each: answer, answer and flag, or hand over.
- Run the agent over them without sending anything.
- Count wrong answers, missed handovers and invented facts. A missed handover is worse than a wrong answer.
- Repeat after every change to the content or the model.
Run the same test on Fin and on a custom agent. It is the only comparison that uses your questions. Read prompt injection in business AI agents as well. A customer message can contain instructions aimed at the agent, and both options need a defence.
What does a test result look like?
The figures below are illustrative. They show the shape of the score sheet, not a benchmark.
| Outcome on 200 past conversations | Count | Reading | |---|---|---| | Correct answer, sent alone | 110 | Safe to automate | | Correct answer, flagged for review | 40 | Keep a person in the loop | | Correct handover | 35 | Working as designed | | Missed handover | 5 | Fix the rule before launch | | Invented fact | 10 | Tighten the approved sources |
Fifteen failures out of 200 is 7.5 per cent. The split matters more than the total. Five missed handovers are a rules problem, so add a trigger such as a repeated message or the word "refund". Ten invented facts are a content problem, so remove the agent's access to unapproved sources. Re-run the same 200 conversations after each fix and keep the scores in one sheet.
Edge cases to include: a voice note with no text, a photo of a damaged item, a message in mixed English and Bahasa, and a customer writing on behalf of someone else.
Which one should you pick?
- Pick Fin if your volume is on web chat and email, you already use Intercom, and your help content is in good shape.
- Pick a WhatsApp-first agent if most customers write on WhatsApp, you need answers from your own records, or you need the thread and the record in your own account.
- Run both if you serve both audiences. Route by channel, and write all outcomes to one log.
Our guide to custom WhatsApp systems vs off-the-shelf tools covers the build side in more detail.
What does the handover look like?
A handover is where support agents succeed or fail. The agent should pass the customer's name, their account or order, the full thread including images and voice notes, what it already told them and why it is handing over. The person should see that on one screen and reply in the same thread. Measure the time from handover to first human reply, and the share of handovers where the customer had to repeat themselves. Both numbers reveal the quality of the design more clearly than any answer-accuracy figure. Set a target for each before launch, and review them weekly for the first month.
At the end of the pilot, write one page: the median time from handover to first human reply, the share of threads where the customer had to repeat themselves, and the missed handovers from your 200-conversation test. Name the result that would change the channel decision, and re-read it at month six.
Where KYN fits
KYN builds WhatsApp-based systems for SMEs in Singapore, Hong Kong, Indonesia and Malaysia. A manufacturer we work with receives a 06:30 morning report on WhatsApp for each department, at about S$0.05 per report. Systems run in a cloud account in the client's name, with named accounts, row-level security and a log of every action with its actor. Messages that carry money, filings or deletions wait for a named person.
The first 14 days are a review of your stack and interviews with your team. A full build ships in under 90 days. See operations systems and integrations.
Frequently asked questions
Can Intercom Fin work on WhatsApp?
Intercom lists WhatsApp among its channels, so check the current terms and limits on its site. The question to test is how well it answers from your own order and account records, not whether the channel connects.
Is a WhatsApp-first agent cheaper than Fin?
It depends on volume and pricing. Fin is priced by Intercom, and a custom agent is a build fee and a monthly retainer. Estimate both on your monthly conversation count.
Can a custom support agent hand over to Intercom?
Yes. A custom agent can create a conversation or ticket in Intercom through its API, so your team keeps one inbox while the agent handles WhatsApp.
Checked 30 October 2026.
Curious whether Marketing agents fits your business? Talk to KYN on WhatsApp — no forms, just a conversation.
