Comparison
Refik and Intercom, read on the same day.
One of these carries two priced lines on the same page: a seat for every person who answers, and a separate charge each time its AI agent resolves something. The other carries one line — a conversation price for a single shop’s storefront, with nothing added on top on the yearly plan. That is the whole comparison, and the rows below are the evidence for it, each one dated, each one from a page you can open yourself.
Method
Two sides, two dates, and an empty cell that means something.
The right-hand side of every row was read out of Intercom’s own pages on 10 October 2026 — the pricing page and the home page, nothing else, no analyst report and no review site. The quotes are theirs, word for word, including the odd-looking ones: where a price sits in one column of their pricing table and a second price sits in the next, we have kept both fragments exactly as they appear rather than smoothing them into a sentence they did not write. The left-hand side was read out of our own source code on 4 October 2026, and the plan figures come from the pricing page in this same directory, which fails its build if a price in the code changes without the page changing with it.
A third page was requested that day and did not arrive. The feature page for their AI agent answered with a redirect to a different domain, so every sentence served there belongs to the page that answered rather than the one we asked for. None of it is in the table. One more sentence about Shopify was read on that redirect target and is also missing on purpose: it was a quote from one of their customers rather than a sentence of their own, and this page prints quotes from nobody’s customers — we have none of our own to print, and borrowing theirs would be the same dishonesty wearing a different coat.
An empty cell is a third state and it is on purpose. It does not mean “they do not have it”. It means we did not read it on that day, so printing a dash there would turn a gap in our reading into a claim about their product. Where a cell is empty, the reason is written underneath, and one of the empty cells is ours.
Three things are missing from this page by design. There is no verdict row and no score: a seat price, a price per resolved question and a price per conversation are three different units, and dividing one by another would produce a number that looks like a measurement and is not one. There is nothing here about how well either assistant answers a shopper, because we did not measure that on either side. And there are no traffic or search figures for either side, because the tool that would have read them was not connected on the day the rest of this was read.
The rows
What each side charges for, and where it runs.
Read the first row before any of the others. It is the reason the rest of the table looks the way it does: on their pricing page two meters run at once, a seat for each person answering and an outcome for each question the AI closes, while ours runs one meter that counts shoppers who started talking. A shop with one owner and a busy storefront and a shop with eight agents and a quiet storefront come out on opposite sides of that line, and no amount of arithmetic on this page will move them back.
Intercom read from intercom.com/pricing and intercom.com; Refik read from this repository and the pricing page.
| What the row measures | Refik | Intercom |
|---|---|---|
| What you are billed for | ||
| The billing unit | Conversations in a billing period. Three bands: 300, 1,000 and 3,000. The price is for the shop, not for the people who read what it collected. | Two units side by side on one page: a seat, and a resolved question. “Intercom helpdesk $29 per seat/mo” sits beside “Fin AI Agent $0.99 per outcome”. |
| Entry tier, list price in USD | Starter, $39 a month or $390 a year, for the whole shop. | “Intercom helpdesk $29 per seat/mo”, charged per seat, with the AI line priced apart from it. |
| Upper tiers, list price in USD | $99 and $249 a month; $990 and $2,490 a year. | “Intercom helpdesk $85 per seat/mo” and “Intercom helpdesk $132 per seat/mo”. Both tiers carry the same “Fin AI Agent $0.99 per outcome” line. |
| Charges on top of the subscription | On monthly billing, $0.15, $0.12 or $0.10 for a conversation above the band, capped at one band’s worth. On yearly billing, none. | The AI line is itself a usage line: it is quoted per outcome rather than per month, so it moves with how many questions get resolved rather than with the calendar. |
| Seats included in the plan | “Includes 20 free Lite seats” on the first tier read, and “Includes 50 free Lite seats” on the tier above it. | |
| Free trial | 14 days on every plan, with a conversation allowance scaled to the plan. | “14-day cardless free trial”. |
| What you can buy, and where it runs | ||
| Buying the assistant on its own | There is no second product. One subscription, one assistant, and no part of it is sold apart from the rest. | The AI agent is offered without the help desk underneath it: “Intercom helpdesk Not included” appears beside “Fin AI Agent $0.99 per outcome”. |
| Platforms it connects to | Shopify only. The catalogue, the policies and the order lookups all come from one store’s Admin API, and there is no second platform in the code. | “Connect Intercom and Fin to the tools you already use, like Slack, Shopify, and Github, so your team has full context in one workspace.” |
| Channels a shopper can use | The web storefront, and only that. WhatsApp, Instagram, Messenger, email and SMS are not built, and that is written on the integrations page rather than left to be discovered. | |
| Handing a question to a person | A handover opens and the shopper is told a person will pick it up. The server can file that as a ticket in exactly one help desk; Intercom is not that one, and the settings screen for choosing a target is not built. | |
| How each side describes itself | A shopping assistant on the storefront: it recommends, it looks up an order, and it hands a hard question to a person. | “Intercom is the only helpdesk with a natively integrated AI Agent, Fin”. |
- Seats in the plan, our cell is empty. We did not read our own console seat limits on 4 October. What we did read is a per-shop price and a conversation band, so any number of seats we printed here would be invented rather than measured. Their cell is filled because their own pricing page states it.
- Channels, their cell is empty. Their channel list was not among the two pages read that day. Ours is written out because a missing channel is the kind of thing a shop finds out after it has paid.
- Handing a question to a person, their cell is empty. Their routing and agent-handover documentation was not among the pages read, and the one feature page we asked for redirected elsewhere. Ours is written out, shortfall included, because a shop already running this help desk needs to know the state of that wire today rather than after installing.
- Billing currencies are on neither side. Our pricing page prints USD, and what a merchant is actually charged in runs through Shopify’s billing, which we did not read on 4 October. No currency list was read on their side either. An empty row on both sides is worth less than no row, so there is no row.
- Neither side’s answer quality is here. Not theirs, not ours. Measuring it would need the same shoppers asking the same questions of both, on a real catalogue, and that has not been done.
The difference that matters
Two meters running at once, against one.
A seat price asks how many people you need on the other end. It is the right unit when the work is a queue of tickets that humans pick up, and it is why a help desk charges that way: the cost follows the team. A price per outcome asks something different — how many questions the machine finished — and it is an honest unit for a product sold as a worker rather than a tool, because the bill arrives when the work lands. A conversation price asks a third thing: how many shoppers started talking at all. It is the right unit when nobody is meant to pick anything up, when the assistant answers and a person only appears for the questions it refuses to guess at.
We are not going to convert one of those into another on this page, and the refusal is deliberate rather than coy. Their page quotes $0.99 for a resolved question; ours quotes $0.15, $0.12 or $0.10 for a conversation above the band. Those two numbers are counted by different rules, written on different pages, and we did not read either side’s counting rules — not what makes an outcome an outcome on their side, and not whether their meter and ours would agree about a shopper who asks two things and buys nothing. A table that divided one by the other would be arithmetic performed on top of an unread definition, which is the most convincing way to be wrong.
There is a second difference and it is structural rather than numeric. Their own sentence puts their AI agent next to Slack, Shopify and Github in one workspace, which is the shape of a product that sits above every tool a team already uses. Ours sits inside one of them. It has no workspace, no second platform in the code, and no context beyond one store’s catalogue, stock and policies — and that narrowness is the only reason it can answer a stock question about a variant without being told anything first.
That is also why the two are not really alternatives for the same job. A shop that answers email, phone and social messages from several people has a help desk problem, and a storefront assistant does not solve it. A shop whose owner is the only person answering, and whose pain is the same nine questions arriving before someone decides to buy, has a storefront problem — and buying seats for it means paying for a queue nobody is standing in.
There is a case where both are true at once, and it is common: a shop with a help desk already in place that wants the storefront to stop sending it the easy questions. For that case the handover matters more than either price, and here is the honest state of ours. The server can open a ticket in exactly one help desk, this is not that one, and the settings screen for choosing a target is not built — so a shop on Intercom cannot wire the handover up from our panel today. The integrations page lists what is connected and what is not, in the same words.
Who each one fits
Where we would tell you to buy the other one.
A help desk fits you if support arrives through more than one door — email, phone, a social inbox — and more than one person answers it. If you need a shared queue, assignment, service-level reporting across channels, or anything that lives after the sale for weeks, that is a help desk’s job and not ours. Our channel list is one channel long and we are not going to pretend otherwise.
Buy the other one as well if what you want is one assistant with the whole team’s context in a single place. Their own page describes that shape and it is a real one: a worker that reads the chat tool, the store and the code repository and answers from all three. We do not do that and we are not building towards it. Our assistant reads one shop and refuses to answer from anything it has not read.
We fit you if the questions that cost you money arrive before the order exists: which of these two is right for me, is this in stock in my size, where is the order I placed on Friday, can I send this back. Those are storefront questions, they arrive on a product page, and answering them needs the catalogue and the policies of one shop rather than a ticket queue. A per-shop price fits that work too, because the number of people in your team has nothing to do with how many shoppers open a product page.
Neither of us fits you if what you actually need is a person. Both products put a human at the end of the path; ours does it by opening a handover and telling the shopper a person will pick it up, and it is told never to say a refund is approved. If there is nobody to pick it up, no assistant on either side of this table fixes that.