AI turned out expensive, and we don't know how to use it well

It's a conversation we have more and more often. The organization got excited about AI, bought licenses for the team, and a few months later the feeling is uncomfortable: you pay every month, but nobody's quite sure what you gained.

At Sumadots we've seen this in at least five organizations. And the problem is almost never the technology. The problem is using the biggest tool for everything, without first deciding what it's for.

Signs you're paying too much

• You pay for licenses per person, and nobody really knows what else they can do besides drafting emails.
• Documents get reviewed with the most expensive model before anyone has tested whether the task works.
• The tool fails with dark photos or documents in another language, and nobody knows why.
• Someone writes the same criteria from scratch in every conversation.
• One team uses three different assistants, each with its own account.
• What works with ten documents becomes unaffordable with a thousand.

If you recognize two or more, you're not alone. And the good news is that almost all of them have a simple fix.

Five tweaks that lower the bill

1. One model for each task

Not every task needs the most powerful model. Sorting survey responses, summarizing meeting minutes, or classifying emails is handled well by a lightweight model that costs a fraction. Save the big model for what really calls for it: a sensitive analysis, or a text a funder is going to read.

2. Prep your documents before sending them

A scanned PDF or a dark photo is hard work for AI: sometimes it misreads it, and sometimes it can't read it at all. Converting the PDF to text first (pulling the words out of the file) makes the answer cheaper and more reliable.

3. Save your instructions

If every week you explain the same criteria to the AI for reviewing a report, write them down once and save them. That way anyone on the team uses them the same way, and the result no longer depends on who wrote the question that day.

4. Use what you already have

Many nonprofits have office software plans that already include an AI assistant. Before buying another license, check what comes with what you already pay for, and put the whole team on a single tool.

5. Sometimes you don't need AI

If the task is always the same (for example, flagging as urgent anything that contains a certain word), a fixed rule handles it for free and without mistakes. AI is worth it when something needs interpreting, not when following an instruction is enough.

Narrowing it down is half the work

The AI tools that work best in the organizations we support have something in common: they do one thing, for a clear audience. They're not an assistant for everything. They're, for example, a space where a group of people practices a hard conversation before having it in real life.

That focus is what lets you know whether it works, what it costs, and whether it's worth growing.

Often it's better to train than to build

When an organization asks us to build something with AI, the most useful answer is often a different one: help the team learn to choose. Which task is worth it, which model to use, how to write instructions that can be reused, and how to tell whether the result got better. You can't buy that with a license. You learn it by doing it with a problem of your own.

If after that you still need to build, you build something smaller, cheaper, and that your team understands. That's how we work in SumaLab.