← All posts
AI toolsPricingSmall business

The hidden cost of 'free' AI tools for a small business

6 November 2026 · 9 min read

Most AI tools aimed at small businesses now start free. That's not a trick in itself — but "free" almost always means "free up to a line you haven't crossed yet". This post covers where the meter actually sits on the tools you're most likely to be offered, how to work out what you'd pay at your real volumes, and the questions worth asking before a free tool becomes part of how customers reach you.

Free tiers aren't charity, they're a sampling strategy

There's nothing dishonest about a free tier. A company gives away the small usage because small usage costs them almost nothing and because habit is valuable. The business model is that some fraction of users grow past the line, and by then switching is a nuisance.

The problem for a small business isn't the eventual bill. It's that the bill arrives at the moment the tool has become load-bearing — when your customers already message you there, your bookings already flow through it, your number already points at it. You're not choosing between free and paid. You're choosing between paying and unpicking something that works.

So the useful question isn't "is it free?" It's "what is being counted, and what happens on the day I go past the allowance?"

The four things that actually get metered

Nearly every AI tool bills on one of four units. Each behaves differently as your business grows, and each has a different failure mode.

UnitTypical ofGrows withWhat catches people out
TokensText and chat tools, messaging agentsLength of conversations, not just the numberContext is re-sent on every turn, so a long thread costs more than it looks
MinutesVoice agents, phone answeringCall length × call volumeA chatty five-minute call costs five times a quick one
Messages or conversationsChat widgets, WhatsApp-style toolsNumber of customer threadsDefinitions vary — a "conversation" may be a 24-hour window, not one exchange
SeatsAnything with a dashboardHeadcountYou add an admin or a second engineer and the price steps up

Knowing the unit tells you how your bill behaves when you have a good month. Per-seat pricing is flat and predictable. Per-minute pricing rises with exactly the thing you want more of: customers ringing you.

Example one: free until you're big enough to charge

Meta's WhatsApp Business Agent is a good, clean example because the pricing change is already public. It answers questions, books appointments and qualifies leads inside WhatsApp, and at low volume it is genuinely free. From August 2026, Meta begins charging larger users roughly $2 per million tokens.

That's a low rate. Text is cheap to run. But look at the shape of the deal rather than the number: the tool is free precisely until your usage is big enough that Meta wants paying, and by that point WhatsApp is probably one of the main ways your customers contact you. You can't easily tell them to stop.

Do the maths: what $2 per million tokens means

Assumptions you should replace with your own numbers. Say an average customer thread runs twelve messages back and forth. Because the model re-reads the conversation and your instructions on every turn, assume that thread burns around 20,000 tokens in total.

At 500 threads a month, that's 10 million tokens — about $20, or roughly £16 at the time of writing. At 2,000 threads a month, around £64.

For most small firms, token cost is not the thing to be frightened of. The thing to watch is dependency: when the pricing changes again, or the free allowance moves, your reply channel moves with it.

We've written separately about what that tool covers and where it stops in Meta's free WhatsApp AI agent. The short version: judge it on fit, not on the word free.

Example two: per-minute pricing and the demo call

Voice is where the arithmetic turns. Voice AI is usually priced per minute rather than per token — xAI's Grok Voice, for instance, is $0.05 a minute, roughly 4p. In a demo that feels like nothing. A two-minute test call costs 8p. You think: how could this ever be expensive?

Because demo calls are short and real calls aren't. A caller who wants a price, describes a leaking pipe, checks you cover their postcode and then asks about Saturdays is on the phone for five minutes. A new-patient enquiry at a clinic can run longer. Minutes multiply by two numbers at once — how many people ring, and how long each one talks.

Do the maths: 4p a minute at real volumes

A one-van trade taking 40 calls a week at an average of three minutes: 120 minutes a week, about 520 a month. At 4p, roughly £21 a month. Cheap.

A busy salon or clinic taking 120 calls a week at an average of five minutes: 600 minutes a week, about 2,600 a month. Roughly £104 a month — and that's the model's share only.

Add the telephony leg, the phone number, any transcription or SMS, and the hours you spend building and fixing the thing. The per-minute rate is the headline, not the total.

That's the real trap with per-minute pricing: it's not hidden so much as misjudged. The rate is published and honest. What's missing is your own average call length, which you probably don't know until you measure it. We went through this in more detail in what 5c a minute actually covers, and the wider market picture is in AI voice agent platforms in 2026.

The costs that never appear on any pricing page

Metering is the visible part. Three other costs show up on free tools and rarely get counted.

Your time

A no-code builder can stand up a voice agent in an afternoon. Getting it to handle a caller with a strong accent, a noisy kitchen behind them, and a habit of interrupting is a different job — see why voice AI trips up on accents and noise. If a weekend of your time is worth £500 of billable work, the free tool just cost £500 before it answered a call. An honest DIY weekend sets out what that actually involves.

Rework on output you can't use

Free writing tools produce drafts fast. The complaint that comes up again and again is that the draft is generic enough to need rewriting, which means you've paid in editing time rather than subscription fees. We collected those complaints in what people actually complain about.

Reputation you can't buy back

Some free tools spend your reputation instead of your money. Free cold-email volume is the classic: it costs nothing per send and can cost you a domain. Why the AI outreach backlash is happening covers the deliverability side. The same logic applies to anything speaking to customers in your name — a bad interaction is a real cost, it just isn't itemised.

What happens on the day you cross the line

This is the single most important thing to find out, and it's usually buried. When your usage passes the free allowance, one of four things happens:

  • It stops. The agent goes quiet, callers hear nothing, messages don't get answered. Cheapest outcome financially, worst commercially — and you may not notice for a day.
  • It degrades. You're pushed onto a slower or smaller model. Nobody tells the customer. Answers get vaguer and you find out from a complaint.
  • It starts billing automatically. No confirmation step, no cap. Fine until a spam wave or a single very long call pushes a bill somewhere you didn't plan for.
  • It asks you first. Rarest and best. The tool pauses, emails you, and waits for a decision.

Ask the question directly before you adopt, and ask whether you can set a hard spend cap. If the answer is "you can monitor usage in the dashboard", that means no cap — that means you're the cap.

Where free is exactly the right answer

Not every free tool is a trap, and a paid managed service isn't always the better buy. Free is genuinely the right choice when:

  • Your volume is low and likely to stay low. If you take fifteen calls a week, per-minute pricing will probably never bite, and a fixed monthly fee may be worse value.
  • The tool isn't load-bearing. A free transcription tool or image resizer can vanish tomorrow and nothing breaks.
  • You're testing an idea. Free tiers are a fair way to find out whether customers will message a bot at all before you commit money to it.
  • You enjoy the building. Some owners genuinely like configuring this stuff, and that time isn't a cost to them.
  • It's a preview, not a dependency. Our own AI website builder works that way — you see a live preview free and decide afterwards whether to publish. Seeing the thing before you pay is a fair use of free.

The flip side: if a tool answers your phone, books your diary, or is the channel customers use to reach you, treat "free" as a pricing stage rather than a feature. Price it as though you're already paying, and decide whether it's worth that. For the fixed-fee comparison, AI receptionist cost UK lays out what a managed line costs against pay-as-you-go minutes, and answering service pricing models explains why the model matters more than the headline number.

Eight questions to ask before you adopt

Copy these into an email to the vendor. The speed and clarity of the reply tells you almost as much as the answers.

  1. What exactly is metered — tokens, minutes, messages, conversations or seats? Define a "conversation".
  2. What's the free allowance in those same units, and over what period?
  3. What happens automatically when I cross it: stop, degrade, bill, or ask me?
  4. Can I set a hard spend cap, and what does the service do when it's hit?
  5. Is the published rate the whole cost, or are telephony, numbers, transcription and SMS billed separately?
  6. What's included in support, and is support available on the free tier at all?
  7. If I leave, what can I export — call recordings, transcripts, contacts, booking history — and in what format?
  8. How much notice do you give before changing prices, and has the free tier changed before?

Then do one more thing, and it's the one most owners skip: measure your own volumes before you sign anything. Pull last month's call log and work out how many calls you took and how long the average one lasted. That single number turns every per-minute price in the market from a guess into arithmetic — and it's the number that decides whether 4p a minute is a bargain or the expensive option.