What AI chat actually costs for a small shop

How a chat tool is priced tells you what it was built to optimize. Here is how to read that, and the arithmetic to do before you sign anything.

You can learn more from a vendor's pricing page than from its feature list, because pricing is a statement about what the company wants more of. Per-seat pricing was built for a support team that grows. Per-conversation pricing was built to grow with your traffic, which means it grows with your best months. Per-resolution pricing was built by someone who wants to be paid for outcomes, which sounds right until you ask who decides what counts as one. Read the model before you read the number. Each of the three carries a cost the page does not print and an incentive it does not mention, and both of those outlive the figure at the top.

Per seat prices your team when your traffic is what varies

Per-seat pricing charges for each person with a login. It came out of helpdesk software, where cost really did scale with the number of agents typing. For an AI assistant that relationship is broken: it answers a thousand conversations whether you have one login or four, and the vendor's cost barely moves.

So it punishes the wrong decision. You want a second person watching transcripts on a busy week, and your weekend cover to have their own account rather than sharing yours. Both improve your service and both raise your bill for no reason connected to what the assistant did. Shops respond by sharing one login, which then makes it impossible to tell who changed an answer.

Per conversation prices your best month

Per-conversation pricing looks the fairest of the three, because it moves with usage. The problem is that usage moves with success. A newsletter lands, a product goes mildly viral, Black Friday happens, and your chat bill spikes in the same week your fulfillment costs do. The month you most need to hold cash is the month the invoice grows.

Before you accept it, ask one question: what resets a conversation? Vendors define this differently and almost never on the pricing page. Some count a session, so a visitor who comes back after lunch is a second conversation. Some reset after 24 hours, some after 30 minutes of silence, some on every page reload. A shopper comparing two sizes across three visits can be one billable conversation or four. Get the definition in writing, then check your analytics for how many of your visitors are returners. If it is most of them, the headline price is not the price.

Per resolution pays the vendor to declare things resolved

Per-resolution pricing sounds like the model that finally aligns everyone: you pay only when the assistant handles something. It is also the least auditable, because the vendor both performs the work and scores it. That is not an accusation of dishonesty. It is structural. Whoever defines resolved controls the invoice.

Insist on seeing the measurement first, in enough detail that you could recompute it. Is a conversation resolved when the visitor stops replying? Somebody who gave up looks identical to somebody who got their answer. Is it resolved when the assistant did not hand off to a human? Then every conversation where handoff failed to trigger bills as a success. Is it resolved when the visitor confirms it? That is the honest version and it is rare, because most people never answer the follow-up. If the vendor cannot show you the rule, you are not buying outcome pricing. You are buying its opinion of its own work.

The upgrade cliff is the real price

Tiered pricing has a shape you can predict. The feature a shop discovers it needs in week three tends to sit one tier above the plan it started on. It is reliably one of a short list, and it is something you will not do without once you have seen the assistant working.

Find your item before you sign. Open the comparison table, read the tier above the one you were going to buy, and ask honestly whether you would run for a year without each line in it. Whatever you flinch at is your real monthly cost, so compare vendors at that tier. A cheap starting plan you will leave in six weeks is not a cheap plan.

The costs that are not on the pricing page

The largest line in year one is often a one-off onboarding fee, quoted on a call rather than published. Ask for it in your first email. If the answer is that it depends on scope, fine, but get the number before you evaluate anything else, because it changes the ranking.

WhatsApp is usually priced as its own channel rather than included, sometimes as a plan uplift and sometimes with per-message fees passed through from the platform. If WhatsApp is where your customers already message you, price that configuration, not the website one. Then there is the cost nobody invoices: the hours your own team spends writing and correcting answer content. That work recurs whenever your policies change, and it belongs in your estimate even though it never hits a card statement. How long setup takes is a separate question; here it is one budget line.

What the assistant costs to run underneath

Under any subscription there is a metered cost the vendor pays. Every reply sends a model the relevant product entries, the knowledge you wrote, the conversation so far and the customer's question, and gets words back. That traffic is measured in tokens and billed per token by the model provider, so cost is not flat per conversation. One question about shipping is cheap. A visitor who pastes a long list of SKUs and goes twenty turns deep is not, and a handful of very long conversations can account for most of a day's spend.

Every vendor handles this, and there are only two ways to handle it. Go silent: at the cap the assistant stops answering, and a customer typing into your widget gets nothing back, or a canned line asking for their email. Or spend less per reply: keep answering, on a cheaper model. Going silent is the simpler of the two to build, so treat it as the default answer until a vendor tells you otherwise. It is also the one your customer notices.

Starly does it the second way, with two limits. A per-chat daily token ceiling stops one runaway conversation consuming the day and leaving everyone else without an assistant. A monthly budget, when exhausted, downgrades the model rather than switching the assistant off: replies get shorter and less nuanced, and they keep arriving. Whatever you buy, ask which of the two behaviors it implements at a limit. No pricing page answers that, and it decides what a customer sees on your worst day.

A degraded answer beats a dead chat window

This is a judgment call, so state it as one. A cheaper model gives a flatter answer. It follows your written policies less gracefully and sometimes misses nuance a better model would have caught. That is a genuine cost, and on a high-consideration product where the assistant is doing real selling you should care about it.

Weigh it against the alternative in front of a real customer. A chat window that takes a message and never responds reads as a broken store, and you cannot explain that to somebody who has already closed the tab. A worse answer is recoverable: the customer still gets the shipping cutoff, still gets the size, and can still ask again. Silence is not recoverable. When you evaluate a tool, ask to see the exact text a customer receives once a limit is hit, and then ask yourself whether you would be content for that sentence to be the last thing a shopper reads before concluding your store is broken.

Where Starly sits, and what is wrong with it

Starly is one flat price. Every feature included, no tiers, no usage meter on the bill, first week free. That is the whole pricing model, stated once so you can hold it against the sections above rather than take it on trust.

The trade-off is real. A flat price means a shop with forty conversations a month pays the same as a shop with four thousand, so the small shop subsidizes the busy one. That is only acceptable because the number is small enough not to be a planning decision. If flat pricing were $400 a month the subsidy would matter and you would be right to want metering. At a small flat figure the value of never having to model your usage curve is worth more than the few dollars of theoretical unfairness. If your volume is genuinely tiny, the right conclusion is not a different pricing model. It is the last section of this page.

What you are actually signing, past the monthly number

Two contract details move more money than the gap between the plans you are comparing, and neither appears on a pricing page. The first is the term. A monthly rolling price and the same figure billed annually are not the same product: one is a subscription and the other is a bet on the next twelve months, usually placed in the first week, which is the week you know least about whether any of this works. Annual discounts are real and worth taking eventually. Taking one before you have run a month is paying for certainty you do not yet have.

The second is what happens on the day you stop paying, and it deserves three questions asked in email rather than on a call. Does the assistant stop answering immediately or at the end of the period you already paid for, because a widget that goes dark mid-week on your storefront is a worse outcome than an awkward invoice. Can you export the knowledge you wrote and the transcripts you accumulated, and for how long after cancellation do they stay retrievable. And what notice is owed to you before a price change, given that a tool wired into your theme and your phone number is not something you re-shop over a weekend.

None of that is exotic and most vendors answer it plainly when asked. The reason to ask in writing is that the answers are worth having eleven months from now, when the person who told you on the call has moved to another company and the plan you signed up to has been renamed twice.

Do the arithmetic yourself

Take what you would pay in a month, on the tier you will actually end up on, plus one twelfth of any setup fee. Divide it by the conversations you genuinely had last month, from your current chat tool, your inbox, or your own count of the questions that reach you. That is your cost per conversation. Set it against what one conversation is worth: average order value, times margin, times the share of chats you expect to end in an order.

For most small shops that collapses into one question. Put your own numbers in: if your average order is $60 and your margin is 40%, each order contributes about $24, so a subscription in the low tens of dollars breaks even at roughly one extra order a month and is ahead at two. Then decide whether you believe an assistant answering the questions that currently go unanswered produces those two orders. Do the same sum for a per-conversation plan at your busiest month rather than your average one, because that is the month that will hurt. If you are unsure, that is what a free week is for, and it tells you more than any comparison table.

Price is the weakest tiebreaker you have

Price is a weak tiebreaker between chat tools, and this page has spent most of its length on it. At small-shop volumes the difference between two plans is money you can absorb inside a month. The difference between an assistant that quotes your real shipping policy and one that invents a plausible-sounding one is not money you can absorb, and it does not appear on any invoice, which is precisely why it loses to the number on the pricing page in most buying decisions. So use pricing to eliminate the vendors whose incentives point the wrong way, and then decide on accuracy. Put the same short list of money questions to every tool on your shortlist — the standing set on the page about invented answers is the one to use — with your own policies loaded in, and read the replies side by side.

Three conditions no plan fixes

Three things make all of this arithmetic beside the point: a question volume low enough that you are faster than any assistant, an inbox that is almost entirely order chasing, and policies that exist nowhere in writing. Each is argued properly on the readiness page. What belongs here is only that none of them is a pricing problem. There is no plan cheap enough to fix them and no plan expensive enough either, so a low headline number is not a reason to skip the check and buy on the grounds that it barely costs anything.

The threshold worth watching is not visitors, it is questions you did not answer. When messages arrive overnight and at the weekend that you only get to on Monday, and the same three questions keep coming back, that is when paying for an assistant starts returning more than it costs. Until then, keep your money.

Common questions

How much does AI chat cost for a small business?

It depends entirely on what the vendor meters. Per-seat plans charge for each team member with a login, per-conversation plans scale with your traffic, and per-resolution plans charge for outcomes the vendor scores. Starly is one flat price with every feature included and the first week free. Whatever you compare, price the tier you will actually end up on rather than the entry tier, and add any one-off onboarding fee, because that is often the largest line in the first year.

What counts as one conversation for billing?

There is no standard definition, which is why it is worth asking before you sign. Some vendors count a browser session, so a shopper who returns after lunch is billed as a second conversation. Others reset after 24 hours, after 30 minutes of silence, or on every page reload. A single shopper comparing sizes across three visits can be one billable conversation or four depending on the rule. Ask for the definition in writing, then check what share of your own visitors are returners before you estimate a monthly bill.

Are there setup fees for AI chat, and how big are they?

Some vendors charge a one-off onboarding or implementation fee that is quoted on a call rather than published, and it is often the largest single line in your first year. Ask for the number in your first email, before you compare features. Beyond that fee, budget for two costs that never appear on an invoice: WhatsApp is usually priced as a separate channel rather than included, and your own team spends real hours writing and correcting the answer content whenever your policies change.

What happens when an AI chat tool hits its usage limit?

There are two possible behaviors and vendors rarely publish which one they implement. The tool either goes silent, meaning a customer typing into your widget gets no reply or a canned request for their email, or it keeps answering on a cheaper model that gives shorter, less nuanced replies. Starly does the second: a per-chat daily token ceiling stops one long conversation consuming the day, and an exhausted monthly budget downgrades the model instead of switching the assistant off. Ask any vendor this question directly.

Does the price still make sense at very low volume?

Usually not, and the volume thresholds behind that belong to the readiness decision rather than to pricing. The pricing-specific point is that at low volume no plan is cheap enough to change the answer, because the cost deciding it is the hours you spend writing and correcting content, and those hours do not scale down with the invoice the way a per-conversation line does. Run the arithmetic anyway before you conclude anything: take the monthly price of the tier you will actually end up on, add a twelfth of any onboarding fee, divide by last month's real conversation count, and hold the result against your margin on a single order.