What AI actually costs a small business per month
The short answer
For most small businesses the tool subscriptions are the small number and the integration is the large one. Off-the-shelf AI tools run roughly Rs 1,500 to Rs 3,000 per user per month, API usage for a modest automation is often under Rs 5,000 a month, and the build is a one-time cost measured in weeks of work. The line nobody budgets is maintenance, and AI subscriptions are already the fastest-growing category of software waste with a 41% waste rate.
The three cost layers
Almost every AI bill breaks into these three, and confusing them is why budgets come out wrong.
Seat subscriptions
ChatGPT, Claude, Copilot and similar, priced per user per month. Easy to buy, easy to forget, and the layer where waste concentrates because nobody notices an unused licence renewing.
API usage
What you pay when your own software calls a model. Priced per token, so cost scales with volume and with how much context you send. A support bot handling a few thousand queries is small; a system summarising long documents is not.
Build and integration
Connecting the model to your systems, defining permissions, handling failures, testing. The largest line for anything beyond a subscription, one-time, and followed by ongoing maintenance.
Three realistic monthly pictures
Ranges rather than quotes. Volume and context size move API costs more than anything else.
| Setup | What it involves | Monthly cost |
|---|---|---|
| Team using AI tools individually | 5 seats, no integration. Value depends entirely on whether people actually use them. | Around Rs 10,000 |
| Support bot on structured questions | API usage plus hosting, after a one-time build. | Rs 5,000 - Rs 10,000 |
| Agent integrated with your order system | Higher API usage, hosting and monitoring, plus meaningful build and maintenance cost. | Rs 15,000 - Rs 35,000 |
Why the API number is so variable
Token pricing means you pay for what you send as well as what you receive. Two businesses running an identical support bot can see a threefold difference in bill purely because one sends the entire knowledge base with every question and the other sends the three relevant paragraphs.
This is the single largest controllable factor in AI running cost, and it is almost never discussed during vendor selection. Retrieval quality is a cost decision as much as an accuracy decision: better retrieval means less context, which means a smaller bill and usually a better answer.
The second factor is model choice. Most classification, extraction and routing tasks do not need the largest model available. Using a smaller model where it performs adequately is not a compromise, it is basic engineering, and the saving compounds every month.
Where the money actually leaks
- Unused seats. AI tool subscriptions carry a 41% waste rate, the highest of any software category, because licences are requested for an evaluation that quietly ends without a decision.
- Sending too much context. Token cost scales with what you send, so passing an entire document when three paragraphs would do multiplies that portion of the bill for no benefit.
- Premium tiers nobody needs. Bought during evaluation to test a feature, never reassessed afterwards.
- Pilots that were never measured. The most expensive category, because without a baseline you cannot tell whether to keep paying or stop.
- Duplicate tools. Three AI subscriptions doing overlapping work because three people each solved the same problem separately.
The line nobody budgets
Maintenance.
Models get deprecated. Prompts that worked stop working when a version changes underneath them. Your own systems change and break integrations. Somebody has to watch output quality, because a degrading AI system does not throw errors, it just gets quietly worse.
Budget a small ongoing allocation from the start. AI automation without maintenance degrades silently, and silent degradation is worse than an outage because nobody notices for months, by which point the trust in the system is gone and rebuilding it costs more than maintaining it would have.
This is also the honest answer to why an agency quote for AI work is not just the build. Anyone quoting a build with no maintenance line has either not done this before or is planning to hand you a problem.
How to keep the bill sane
Audit seats quarterly
Pull the list of who has a licence and who logged in this month. Reclaim anything untouched for sixty days. This takes an hour and is the highest-return hour in the budget.
Cap API spend at the provider
Every major provider offers a hard spending limit. Set it. A runaway loop in your own code should not be able to produce an unbounded bill.
Send the minimum context
Measure the average tokens per request. If it is high, invest in retrieval rather than accepting the cost. Halving context roughly halves that portion of the bill.
Use the smallest model that works
Test the cheaper model on your actual task before assuming you need the expensive one. For extraction and classification, the gap is often invisible to users.
Measure the process you automated
Time taken before, time taken after. Without this you cannot tell whether the spend is buying anything, and unmeasured spend is the first thing cut in a review.
From our own projects
On an e-commerce platform we built, AI generates the question bank. The measurable outcome was time: work that took a full day now does not. What made that spend justifiable was not the size of the saving, it was that we could put a number against it at renewal.
For our own team, we pay for Claude, ChatGPT and Gemini alongside roughly fifteen other tools. Three AI subscriptions for a small team is exactly the pattern that produces waste elsewhere, so our usage is deliberately narrow. AI helps write code. Logic, security and performance decisions stay with the team, and nothing goes in without being verified several times.
That narrow scope is also a cost decision. When a subscription supports one named task, the renewal question answers itself. When it supports a vague sense of being useful, it renews forever by default.
Frequently asked questions
Working out what AI would cost you?
We scope AI work against a measurable process rather than a technology, and we say so when the answer is that you do not need it.