How to choose an AI tool for your company (and when to say no)
The previous lesson had you write a one-page usage policy, and its first line was: approved tools. But how do you decide what goes on that list? In most companies a tool lands on the list because one employee tried it and liked it, or because the name is everywhere this month. That choice can work by luck, but it does not survive the first serious question: who owns the data we typed in? And what happens if the tool shuts down or changes its price overnight?
Start from the task, not the tool
The common mistake is to start with a tool and then look for work to give it. The correct order is the reverse: define the task first in one measurable sentence, such as turning the supervisor's end-of-day voice notes into a standard report. Only then look for a tool that does that sentence. A defined task gives you a yardstick; a tool looking for a task always finds someone to justify it.
And never drop the free option: using no tool at all. If the task takes a few minutes a week, the time spent learning, monitoring and reviewing can cost more than it saves.
Five criteria for comparing candidates
Once you are down to two or three candidates, compare them against written criteria rather than impressions. These five cover most decisions in small and mid-sized companies:
| Criterion | The practical question | Warning sign |
|---|---|---|
| Fit for the task | Did you test it on real cases from your own work, not on a prepared demo? | One impressive trial, then a decision |
| Data handling | What happens to what you type in? Is it used to train models, and can that be turned off? | Vague terms, or a privacy page that answers nothing |
| Fit with your systems | Does the output reach the system your team actually works in? | Manual copy and paste every single time |
| Real cost | What does it cost per month at your actual volume, not at trial volume? | Usage-based pricing with no clear ceiling |
| Exit path | Can you export your data and your instructions and move elsewhere? | No export, or a closed format nothing else opens |
The data criterion deserves a longer pause. Before you put customer or employee data into any external tool, review your obligations under the Personal Data Protection Law, and take that from the official text published by the Saudi Data and AI Authority (SDAIA), not from the vendor's marketing page. As the data-safety lesson showed, many tasks can be done with anonymised inputs, and that alone widens the list of tools available to you.
A concrete example
A small contracting company wants to turn daily voice notes from its sites into a written report. The first candidate is a cheap general tool that writes excellent text but connects to nothing, so someone copies the result across by hand every evening. The second candidate costs more but sends the report straight into the system where the manager already follows projects.
Judged on writing quality alone, the first one wins. Judged on the whole task, one standard report arriving in one place with no manual step, the second wins: the price gap is paid back by employee time that repeats every day. That is what comparing by task rather than by tool means.
A tool you cannot leave is not a tool. It is a long-term commitment you made without noticing.
Checklist before you approve a tool
- You tested it on a sample of your real cases and recorded where it failed, not only where it shone.
- You read how it handles your data from the official source, not from a marketing summary.
- You know the cost at your expected volume months from now, not in the first week.
- You named the person who owns the account and know how it is recovered if that person leaves.
- You know your exit: export of data and instructions, and a fallback if the service stops.
- You set a review date to decide later: keep, change, or drop.
Record the decision in one line
You do not need a large document. A small register is enough: tool name, the task it was approved for, the type of data allowed in it, the person responsible, and the review date. That register keeps the approved-tools list from swelling until it means nothing, and it answers the customer who asks where your data goes.