What AI Automation Actually Costs a Small Business
The projects that pay for themselves are boring ones with a number attached. The ones that fail usually started from the technology instead of the task.
Most AI projects that fail did not fail technically. They failed because nobody could say what would be different afterwards. The ones that succeed almost always start from a task somebody currently does by hand, slowly, many times a week.
The work that is genuinely worth automating
- Reading invoices, purchase orders or delivery notes into a system.
- Answering the forty questions that make up most of your support inbox.
- Qualifying inbound leads before a person spends time on them.
- Pulling structured data out of PDFs that arrive by email.
Each of those has a number attached — hours spent, or errors made — and that number is how you will know whether it worked. If you cannot state the number before starting, the project has no definition of success.
What it costs to build
A focused automation handling one process typically runs PKR 250,000 to 700,000 to build, depending on how many systems it has to talk to. Something spanning several departments with human approval steps costs considerably more, and should be built in stages rather than all at once.
What it costs to run
This is the part most quotes leave out. Hosted models charge per token, so the running cost scales with usage rather than being fixed.
- A support assistant handling a few hundred conversations a month: roughly PKR 5,000 – 20,000 in model costs.
- Document processing at a few thousand pages a month: usually more, and worth measuring on a sample before committing.
- Self-hosted open models: no per-token cost, but a server to run and maintain, which is its own budget line.
When the honest answer is no
If the task must be exactly right every time and there is no way to check the output, a language model is the wrong tool. If it happens twice a month, the automation will cost more than the work it replaces. If your data is scattered across WhatsApp messages and paper, fix that first — automating a mess produces a faster mess.
We turn down roughly a third of the AI enquiries we receive for one of those three reasons. That is not modesty; it is that a project which cannot pay for itself damages both sides.