A custom AI assistant costs a small business from nothing extra to several thousand dollars up front. Set up inside a chat plan you have, such as a ChatGPT or Claude project or a Gemini Gem, it costs $0-$25 a seat a month plus a few hours. A hosted chatbot builder runs about $40-$500 a month. A bespoke build adds one-off development.
One date to know: Gemini Gems become skills from November 2026, so anything you set up as a Gem carries over as a skill. The price that surprises owners isn't the setup. It's the upkeep. An assistant is only as good as the documents behind it, so someone has to update prices, policies and opening hours as they change, and check its answers every month. Budget two to four hours a month of that person's time whichever route you choose, and decide who that person is before you build anything.
The three routes and what each is for
"Custom AI assistant" covers three quite different things. Work out which you actually need, because the costs differ by a factor of a hundred.
- A configured assistant inside a chat plan. A ChatGPT project, a Claude project or a Gemini Gem with your instructions and reference files. Staff use it inside the chat app. Best for internal help: drafting in house style, answering questions from your handbook, preparing quotes from a price list.
- A hosted chatbot builder. A service such as Chatbase that trains a bot on your website and documents and puts it on your site or messaging channels. Best for customer questions: opening hours, delivery areas, order status, simple enquiries.
- A bespoke build on an AI model's API. A developer connects a model from Anthropic, OpenAI or Google to your own systems. Best when the assistant must look things up in, or write to, your booking system, stock system or CRM in ways off-the-shelf tools can't.
Don't build any of these as a custom GPT. OpenAI is retiring custom GPTs across ChatGPT plans; they stop running on 11 December 2026. For shared business context, use projects or Gems. Our tutorial comparing custom GPTs, Claude Projects and Gemini Gems covers the migration and the differences.
Itemised costs for each route
List prices in USD as of September 2026. Internal time is valued at an illustrative $30 an hour; use your own figure.
| Cost item | Inside a chat plan | Hosted chatbot builder | Bespoke build |
|---|---|---|---|
| Software, monthly | $0 if you already have ChatGPT Business, Claude Team or Google Workspace; otherwise $20-$25 a seat (two-seat minimum on team plans) | Chatbase, for example: $40 (700 message credits), $150 (4,000), $500 (15,000) on monthly billing; about 20% less annually | API tokens, often $5-$150 a month at small-business volumes (see the sums below), plus hosting |
| Setup, one-off | 3-8 hours of internal time ($90-$240) | 1-3 days of internal time, or a freelancer | Developer time: typically weeks, not days |
| Knowledge preparation | 2-6 hours tidying documents | 4-12 hours tidying and writing FAQs | Same, plus structuring data for lookups |
| Testing | 2 hours | 4-8 hours, including awkward questions | Days, including failure cases |
| Upkeep, monthly | 1-2 hours | 2-4 hours reviewing conversations and updating content | 2-4 hours, plus developer time for fixes and model changes |
| Extras | None usually | Extra credits ($40 per 1,000 on Chatbase), branding removal ($99 a month on Chatbase) | Monitoring, security review, a data processing agreement with the model provider |
The chatbot builder figures come from Chatbase's pricing page as an example of the category; other builders price differently, often by messages or conversations. Check the unit carefully. A "message credit" can be one reply or several, depending on the model the bot uses.
Route one in practice: a staff assistant for about $60 a month
Take an illustrative funeral director with three office staff who keep asking the same questions: what the current price for a particular coffin range is, what the procedure is for a repatriation, how to word a notice for the newspaper. The business already pays for Claude Team: three Standard seats at $20 each on annual billing, $60 a month. The assistant costs nothing extra in software.
Setup: the office manager spends about five hours gathering the price list, the procedures manual, notice templates and the house-style note into one project, and writing instructions. Here's an illustrative version of those instructions and a test answer:
Project: Office assistant
Instructions:
You help the office team of a family funeral director.
Answer only from the attached price list, procedures
manual and templates. Quote the document and section you
used. If the answer isn't in the files, say "Not in our
documents - check with the manager". Never quote a price
that isn't in the current price list. Use plain, gentle
British English for anything that may be read by a family.
Test question: "What's our charge for a Saturday service?"
Illustrative answer:
"The price list (section 4, Additional charges) lists a
weekend service supplement of $350. The procedures manual
(section 2.3) says Saturday services need the manager's
approval before booking."What you'd check: that section 4 really says $350 and is the current version. The instruction to cite a section makes that check take seconds. First-year cost: five hours of setup plus an hour a month of upkeep is about 17 hours, roughly $510 of internal time, on top of seats the business already pays for.
For a Google Workspace business, the same thing is a Gemini Gem, which can be shared with colleagues like a Drive file. See setting up Claude Projects as a shared team assistant for the fuller setup.
Route two in practice: a customer assistant for a florist
An illustrative florist gets around 400 website chats and messages a month: delivery areas, same-day cut-off times, whether a particular bouquet is available, care instructions. It wants a bot on its website that answers these and hands anything else to staff.
On a builder at the $40 tier, 700 message credits a month. If each chat averages three bot replies and each reply uses one credit, 400 chats use 1,200 credits, which is over the tier. The choices:
Option A: $40 plan + 500 extra credits
(1,000-credit top-ups at $40) = $80 a month
Option B: $150 plan with 4,000 credits = $150 a month
(lots of headroom)
Option C: annual billing on the $150 plan,
about 20% less = about $120 a monthCheck how many credits your chosen model uses per reply before choosing; a more capable model may use several. Setup takes two or three days, most of it writing a clear FAQ from the questions staff already answer. Our tutorial on building the FAQ your AI chatbot needs covers that step, and the first week's testing matters more than the plan tier.
Route three in practice: when a bespoke build earns its cost
A bespoke assistant makes sense when the answer lives in your own systems and changes by the hour: stock levels, booking slots, order status. Picture an independent bookshop that wants customers to ask "do you have this in stock, and can you hold it for me?" and have the assistant check the stock system and place a hold.
No builder does that out of the box with this bookshop's stock system, so it needs a developer. The costs split into two parts.
One-off development
Developer rates vary widely, so treat this as arithmetic, not a quote. If a freelancer charges an illustrative $500 a day and the job takes 12 days (connecting to the stock system, building the hold function, the chat interface, testing and handover), that's $6,000. A more complex job, or an agency with project management, could be two or three times that. Get three quotes against the same written brief; our tutorial on briefing a developer on a custom AI workflow shows how to write one.
Monthly running costs
The model's API is billed per token (roughly, per chunk of text; a million tokens is about 750,000 words). Here's the sum for 1,500 customer questions a month, each using about 4,000 tokens of input (instructions, the question, and stock data) and 400 tokens of output:
Input: 1,500 x 4,000 = 6,000,000 tokens a month
Output: 1,500 x 400 = 600,000 tokens a month
Claude Sonnet 5 ($2 in / $10 out per million):
6 x $2 + 0.6 x $10 = $12 + $6 = $18 a month
Claude Haiku 4.5 ($1 in / $5 out):
6 x $1 + 0.6 x $5 = $6 + $3 = $9 a month
gpt-6-luna ($0.10 in / $0.50 out):
6 x $0.10 + 0.6 x $0.50 = $0.90 a monthTokens are cheap at this scale. Hosting the app is usually a similar small monthly sum. The real monthly cost is the developer's time for fixes and changes. Agree a support arrangement, say two hours a month on call (about $125 at the illustrative day rate), before the build starts, and read what AI maintenance costs after go-live for the items people forget.
The token mistake that multiplies the bill
Token costs stay small only if each request carries a sensible amount of text. An illustrative mistake: to make the bookshop's assistant "know everything", a developer includes the shop's full 200-page catalogue, about 150,000 tokens, in every request instead of looking up the relevant titles first.
Input: 1,500 questions x 150,000 tokens = 225,000,000 tokens
Claude Sonnet 5 at $2 per million input: 225 x $2 = $450 a month
Compared with the targeted version: $18 a monthSame assistant, same questions, 25 times the cost, and often slower answers too. The fix is retrieval: search the catalogue first and send only the few entries that match the question. Ask any developer how much text goes into a typical request, and ask for the monthly token estimate in writing.
One question, three routes
A useful way to choose is to take one real question and see what each route can do with it. Say a customer asks the florist: "Can you deliver the large seasonal bouquet to the hospital on Friday before 11am?"
- Route one (staff project): a member of staff pastes the question in, and the project answers from the delivery policy: "Hospital deliveries go to the main reception; Friday morning slots end at 11am; the large seasonal bouquet is in the current range." Staff still check the van schedule and reply. Saves a few minutes per enquiry; the customer waits for a human.
- Route two (website builder): the bot answers from the FAQ: yes, hospital deliveries are possible, morning slots run until 11am, and here's the order link. It can't confirm Friday's schedule has room, so it says so and offers to pass the question to staff. The customer gets an instant, mostly complete answer.
- Route three (bespoke): the assistant checks the delivery system for Friday morning capacity and the stock system for the bouquet, confirms both, and takes the order. The customer gets a definite answer at 10pm. That's the only route that can, and it's the only one that costs thousands to build.
If most of your questions look like the route-two case, where a good general answer plus a handover is fine, you don't need route three.
Five questions to ask before paying anyone
- "What exactly counts as a message, credit or conversation?" Get the unit in writing, and a worked example at your volume.
- "Where does my knowledge live, and can I export it?" If the answer is "in our platform", ask how you'd move it.
- "What happens when it doesn't know?" You want a clear handover to a person, not a confident guess.
- "Which model does it use, and who pays when that model changes?" Especially for bespoke builds.
- "What will I need to do every month?" An honest supplier will describe upkeep. One who says "nothing" is either selling something very simple or not being straight with you.
Three budgets side by side
| Scenario | One-off | Monthly | First-year total |
|---|---|---|---|
| Cheap: staff assistant in an existing Claude Team plan (funeral director) | About $150 of internal time | $0 extra software, about $30 of upkeep time | About $510 |
| Mid: website assistant on a builder (florist) | About $600 of internal time for setup and FAQ | About $120 software plus $90 of upkeep time | About $3,100 |
| High: bespoke stock-and-hold assistant (bookshop) | About $6,000 development plus $600 internal | About $30 tokens and hosting, $125 support, $90 upkeep | About $9,500 |
The high scenario is only worth it if holds placed by the assistant turn into sales the shop would otherwise lose. At a typical margin on books, that means a lot of holds. Many bookshops would be better served by the mid route plus a clear "we'll check and message you" handover to staff.
Costs that turn up after launch
- Out-of-date knowledge. An illustrative florist raises its delivery charge in March, updates the website, but forgets the chatbot's FAQ. For three weeks the bot quotes the old price and staff honour it on 60 orders. That's a real cost of skipping upkeep.
- Review time. Someone should read a sample of conversations every month. Twenty conversations take about 30 minutes and catch most problems, such as a question the bot keeps dodging or an answer that has drifted from your policy.
- Vendor change. Builders get acquired, change pricing or shut down. Clockwise, an AI calendar tool, shut down in March 2026 and deleted user data rather than transferring it. Keep your FAQ and documents in your own files so moving is a day's work, not a rebuild.
- Model changes. Providers retire older models. On a bespoke build, that can mean a developer revisiting prompts and testing again. Ask your developer how they handle it.
- Data and privacy work. For anything customer-facing, check the provider's data processing terms and tell customers they're talking to an AI. If you sell to customers in the EU, the EU AI Act's transparency duty for chatbots has applied since 2 August 2026.
How to keep the cost down
- Start with route one. Many "we need a custom assistant" requests are really "we need our documents in one place and a way to ask questions of them". A project or Gem proves the idea for almost nothing.
- Write the FAQ before choosing a tool. If you can't list the 30 questions it should answer, no tool will fix that, and the list is what you'll test with.
- Use the cheapest model that passes your test. Run the same 30 questions through a low-cost model and a mid-range one. If you can't tell the answers apart, stay cheap.
- Keep the knowledge portable. Documents in your own Drive or SharePoint, not typed only into a vendor's settings screen.
- Cap spending. Set a monthly limit on any API account and a credit ceiling on any builder before launch.
If your assistant mainly needs to answer questions from a large set of documents, what a chat-with-your-documents system costs goes deeper on that specific case, including how retrieval keeps each request small.
Custom assistant costs: follow-up questions
Should I build my assistant as a custom GPT?
No. OpenAI is retiring custom GPTs; they stop running on 11 December 2026 and are being migrated into plugins and skills. For a staff assistant, use a ChatGPT or Claude project, or a Gemini Gem, which are designed for shared business context. For a customer-facing assistant, use a chatbot builder or a bespoke build on an API.
Does my ChatGPT or Claude subscription cover API costs?
No. Subscriptions such as ChatGPT Plus or Business and Claude Pro or Team don't include API use. A bespoke assistant built on the API is billed separately, per token, on the developer platform, so budget for it as its own line and set a spending limit on the account from day one.
How long does a custom assistant take to set up?
A project or Gem for staff takes an afternoon to set up and a week of testing. A chatbot builder on your website takes one to three days including testing. A bespoke build usually takes several weeks, most of it spent preparing the knowledge, handling edge cases and testing rather than writing code.
What's the cheapest model that works for an assistant?
For answering questions from your own documents, the low-cost tiers are often enough: Claude Haiku 4.5 at $1 per million input tokens, or OpenAI's gpt-6-luna at $0.10. Test the cheap model on 30 real questions first and only move to a more expensive one if the answers are noticeably worse.
Further reads
- Should a Small Business Build a Custom AI Tool or Buy One? — The wider build-or-buy decision before you spend.
- What Does the ChatGPT API Cost for a Business Automation? — Token maths for OpenAI's models in more depth.
- AI Knowledge Base Options for Small Businesses Compared — Where the assistant's knowledge should live.
- How Much Does a Website AI Chatbot Cost, and Will It Pay Off? — Whether a customer-facing assistant pays back.
- What If Your AI Vendor Shuts Down? Checks Before You Commit — Protect yourself if a builder platform disappears.
- How to Build Gemini Gems for Repeat Business Tasks — The Google Workspace way to build a staff assistant.
- Per-Seat SaaS Fees vs One Custom Tool: Five-Year Costs Compared — Five-year costs of per-seat software against a custom AI tool, year by year, with a break-even seat count and the upkeep costs quotes leave out.
- AI Assistant vs Virtual Assistant: Which Should You Pay For? — A side-by-side comparison, a food truck owner's week sorted task by task, real costs of each and when a small business needs both.
- How Much Does an AI Agent Cost a Small Business? — The five ways AI agents are priced, the cost-per-task sum that makes them comparable, and the supervision time that usually costs more than the software.
- No-Code AI Tools vs a Custom Build: When Templates Stop Working — Six signs a Zapier or Make workflow has outgrown no-code, the middle options to try first, and a worked cost comparison for a care agency's shift offers.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Anthropic and OpenAI API pricing (per million tokens), ChatGPT Business and Claude Team pricing, Chatbase pricing page, Google Workspace Gem sharing announcement, OpenAI help centre on custom GPT retirement; checked September 2026. Developer rates are illustrative only.