How much should a small business budget for AI in its first year?

Budget in two parts. The fixed part is chat seats for the people who will use them (US$20 to US$25 a user a month on ChatGPT Business or Claude Team) plus the build and upkeep of any assistant; the variable part is what the assistant uses, billed per message, per minute or per million tokens. The number of users, the volume of conversations and the channel (website chat, WhatsApp, text or phone) decide the total more than the brand of model.

A sensible first year starts with one task, such as answering new enquiries after hours, on one channel. Count the volume, price it with the list prices below, add the build and upkeep, keep a buffer, and check the real bills after 90 days before you add seats or a second task.

Several of these prices change inside a year. Google’s Gemini 3.8 Flash doubles on 1 January 2027, Microsoft and Google show introductory offers, and Meta can update its WhatsApp rates. Re-read the providers’ pages when you renew.

Updated Nov 16, 2026

How it works

  1. Pick one task first

    Choose the job that costs you the most today when nobody does it: replies to new enquiries, booking, reminders, or office work such as sorting email. One task gives you one volume to count and one result to judge.

  2. Count the volume

    Write down how many conversations a month, how many messages in each, and how many call minutes if it is the phone. Use last month’s inbox, WhatsApp and call log, not a guess.

  3. Pick the channel

    A chat on your own website has no per-message carrier fee. WhatsApp is charged by Meta per delivered message, by the customer’s country, with the first 1,000 service replies a month free on each number. Text messages are charged per message plus US carrier fees and a registration fee. Phone calls are charged per minute, by the voice service and by the phone line.

  4. Add the model usage

    An assistant pays for the text it reads and writes, in tokens. A worked example at OpenAI’s list price for GPT-6.1 Terra (US$1 per million tokens read, US$5 per million written): a reply that reads 3,000 tokens and writes 300 costs US$0.0045, so 1,000 such replies cost US$4.50. Models that think before answering bill that thinking as written tokens.

  5. Add the channel fees

    Multiply your monthly volume by the channel’s rate: messages times Meta’s or Twilio’s rate, minutes times the voice service’s rate plus the phone line’s. Add the monthly fee for each phone number.

  6. Add seats only for real users

    Chat seats are charged every month whether they are used or not. Start with the people who will use them every week; ChatGPT Business and Claude Team both need at least two seats.

  7. Add the build and upkeep

    Someone has to connect the assistant to your calendar, inbox or CRM, write down your facts and rules, and keep them current. Count that time, in-house or a provider’s, plus the automation tool that runs the steps (n8n or Zapier, for example).

  8. Keep a buffer

    Usage-billed accounts are often prepaid: when the balance hits zero, the assistant stops. Set spending limits and low-balance alerts, and leave room for introductory prices that end during the year.

  9. Review at 90 days

    Compare the real bills with your estimate: seats used, conversations handled, cost per conversation, and which offers end next. Then decide whether to cut, keep or add a second task.

What it costs to run

List prices on each provider’s own page, read on 16 November 2026, in US dollars unless stated and before tax. They are not Betterlane prices. The build and upkeep of an assistant has no list price, so it is not in this table.

ItemList priceSource
ChatGPT Plus, one personUS$20 a month, billed monthly; no annual planOpenAI Help Center, "What is ChatGPT Plus?"Read on Nov 16, 2026
ChatGPT Business, Standard seatUS$25 per user per month billed monthly, or US$20 per user per month billed annually; usually 2 paid seats minimumOpenAI Help Center, "ChatGPT Business overview"Read on Nov 16, 2026
Claude Pro, one personUS$20 billed monthly, or US$17 a month with the annual subscription (US$200 billed up front)Anthropic, Claude plans and pricingRead on Nov 16, 2026
Claude Team, Standard seat (2 to 150 users)US$25 per seat per month billed monthly, or US$20 billed annuallyAnthropic, Claude plans and pricingRead on Nov 16, 2026
Google Workspace Business Starter (Gemini in Gmail and Gemini chat) and Business Standard (Gemini in Gmail, Docs, Meet and more)US$7 and US$14 per user per month with a 1-year commitment; 50% off from 30 Nov 2026 to 28 Feb 2027 for new sign-upsGoogle Workspace, Pricing (United States, prices in USD)Read on Nov 16, 2026
Microsoft 365 Copilot Business, add-on to a Microsoft 365 planUS$21.00 per user per month, paid yearly; shown at US$18.00 under an offer on 16 Nov 2026Microsoft, Microsoft Copilot for business, pricingRead on Nov 16, 2026
OpenAI API, GPT-6.1 Terra (standard processing)US$1.00 per million input tokens, US$5.00 per million output tokensOpenAI API docs, PricingRead on Nov 16, 2026
Anthropic API, Claude Sonnet 5.5US$2 per million input tokens, US$10 per million output tokensClaude Docs, PricingRead on Nov 16, 2026
Google Gemini API, Gemini 3.8 Flash (paid tier)US$0.75 per million input tokens and US$3.75 per million output tokens (thinking included) until 31 Dec 2026; US$1.50 and US$7.50 from 1 Jan 2027Google AI for Developers, Gemini Developer API pricingRead on Nov 16, 2026
WhatsApp, customer in the United States (Meta’s North America rate)US$0.0034 per delivered service or utility message, US$0.025 per marketing message; first 1,000 service messages a month free on each number; messages from customers are freeMeta for Developers, Pricing on the WhatsApp Business Platform (updated 30 Sep 2026; USD and BRL rate cards effective 1 Oct 2026)Read on Nov 16, 2026
WhatsApp, customer in BrazilUS$0.0068 per delivered service or utility message, US$0.0625 per marketing message (R$0.035 and R$0.3217 on Meta’s BRL rate card); same free allowanceMeta for Developers, Pricing on the WhatsApp Business Platform (updated 30 Sep 2026; USD and BRL rate cards effective 1 Oct 2026)Read on Nov 16, 2026
Twilio text messages, US local numberUS$0.0083 per SMS sent or received, plus carrier fees and A2P 10DLC registration feesTwilio, SMS pricing in the United StatesRead on Nov 16, 2026
Twilio phone line, US local numberUS$1.15 a month for the number; US$0.0085 a minute to receive calls, US$0.0140 a minute to make themTwilio, Voice pricing in the United StatesRead on Nov 16, 2026
Retell AI voice agents, pay as you goUS$0.07 to US$0.31 a minute, depending on the model and voice; chat agents from US$0.002 a messageRetell AI, PricingRead on Nov 16, 2026
Vapi voice agents, pay as you goUS$0.05 a minute for Vapi hosting, with the speech, model and voice providers passed through at cost; no minimumVapi, PricingRead on Nov 16, 2026
n8n Cloud, Starter plan€20 a month, billed annually, for 2,500 workflow executions (price shown in euros)n8n, Plans and PricingRead on Nov 16, 2026
ZapierFree plan US$0 with 100 tasks a month; Professional starting from US$19.99 a monthZapier, Plans & PricingRead on Nov 16, 2026

Third-party list prices, read on the date shown. They are not Betterlane prices.

What we learned building it

Our own systems: the website assistant, the phone demo line and our n8n workflows, September to November 2026

A prepaid balance at zero stops the assistant
When the model credit on our own account ran out, every model call came back refused until it was topped up. Our simpler fallbacks kept answering; anything that needed the model did not. Set a low-balance alert and an automatic top-up before you go live.
Thinking is billed, and it can use up the whole reply
In one of our n8n workflows, a model that reasons before answering spent its entire output allowance thinking and returned no text. Our check caught it and passed the message to a person. Leave room for thinking in the token limit and in the budget.
Test a phone line with a person on the other end
An automated test caller we built to ring our own demo line was flagged by the voice provider as a robocall. Budget the test minutes for a real person calling, and read the provider’s rules for outbound calls before you plan any.

When it is not worth it

  • If you get a handful of enquiries a week, a chat seat or two for the people who answer them may be all you need; an assistant’s build and upkeep can cost more than the replies it saves.
  • Do not buy seats for the whole team on day one. Seats are billed whether they are used or not; add them after the 90-day review.
  • Hold off on annual commitments until you have three months of real bills, unless the discount matters more to you than the flexibility.

Want to see the running costs on your own volume before you set the budget? The first flow is free, built in a week; if it does not do what was scoped, you pay nothing.

See the service: The free first flowLet’s talk

Questions

Is the model or the channel the bigger cost?

It depends on the channel. In the worked example above, a reply costs under half a US cent in model usage, and a WhatsApp service reply to a US customer costs US$0.0034 after the free 1,000 a month. A phone call is priced by the minute: Retell lists US$0.07 to US$0.31 a minute, before the phone line.

Do I need ChatGPT Plus or Claude Pro to run an assistant on my website or WhatsApp?

No. Those seats let a person chat. An assistant that answers your customers runs through the providers’ developer APIs, which are billed separately, per token.

Do prices differ by country?

Yes. Meta charges WhatsApp messages by the customer’s country: a service reply costs US$0.0034 in North America and US$0.0068 in Brazil. Google Workspace and ChatGPT also show different prices and currencies by region.

What should I look at after 90 days?

Seats in use, conversations or minutes handled, the cost per conversation, any balance that ran low, and which introductory prices end next. Cut what is idle before you add a second task.

Can I start on free plans?

For testing, often yes: Zapier has a free plan with 100 tasks a month, Gemini has a free tier on some models, and Meta does not charge for the first 1,000 service replies a month on each number. A live assistant usually needs paid usage, so put it in the budget from the start.

Sources

All guides