To decide whether an AI tool is worth keeping, compare the result it creates with its full cost: subscription, setup, training, review, corrections, and attention. Do not begin with a vendor case study or a list of impressive features. Begin with the task your business wants to improve and the number that describes its current state today right now.
This matters because a tool can feel productive without improving the business. You may spend an hour experimenting, produce a polished draft, and still have the same number of unanswered leads. You may save ten minutes on a report and lose twenty minutes checking facts. A fair AI tool ROI check includes both sides of that exchange.
Start with a baseline
Write down what happens before the tool is introduced. Pick one task that repeats often enough to observe. Record how long it takes, how many items are completed, how often corrections happen, and what happens next. If the task affects customers, note the customer-facing measure too, such as response time or missed follow-up.
Keep the baseline modest. You do not need a complicated dashboard. A spreadsheet with dates, task counts, minutes, corrections, and next outcomes is enough for many small teams. The goal is to make your before picture visible so your after picture has something to answer.
Give the tool one job
Do not evaluate a tool for everything it claims to do. Evaluate it for one job with a clear input and output. “Help my business grow” is too broad. “Turn each approved call summary into a follow-up draft within one business day” is testable.
Choose a task that is repetitive, low enough risk for a controlled trial, and connected to a real business priority. The selection method in Your First AI Workflow is useful here. A narrow starting point gives you a better signal than a wide experiment where every variable changes at once.
Run a 60 to 90 day test
Use the same basic process during the test. Changing the prompt, person, tool, and workflow every few days makes the result difficult to read. Keep a short log of what the tool produced, what a person changed, and whether the final work was used.
Set a review date before the trial begins. During the test, ask three questions: Is the task happening more reliably? Is the output good enough after review? Does the improvement create time or capacity for a valuable next action? If the answer to all three is no, the tool has not earned a permanent place.
Measure quality as well as speed
Speed is easy to notice, but quality protects the business. Count factual corrections, missing context, awkward customer language, duplicate work, and cases that needed a full restart. If an assistant produces a draft faster but requires a complete rewrite, the apparent gain may be false.
For customer-facing work, include a human approval step. The person approving the output should know what the tool was asked to do and what it cannot know. A review is not a failure of automation. It is part of the process that keeps a fast draft from becoming a careless promise.
Count the full cost
Write down the plan price and any usage or seat cost. Then add the less visible costs: learning time, maintenance, data cleanup, integrations, prompt revisions, and review. If only one person knows how to operate the tool, record that dependency too.
Now compare that total with the result. A tool may be worth keeping because it helps you respond consistently during a busy week. It may not be worth keeping if it produces occasional convenience but has no important job. A small business AI stack should earn trust through repeated use, not through novelty.
The article The 5-Tool AI Stack for a Service Business gives you a way to assign tools by job before you add another subscription. That assignment makes the review easier because every tool has a defined place in the flow.
Use a keep, change, combine, or cancel decision
- Keep: the tool has an owner, a repeated job, and a result that justifies the full cost.
- Change: the job matters, but the setup or instructions are producing avoidable problems.
- Combine: another tool already covers most of the same work with less upkeep.
- Cancel: the job is rare, the owner is unclear, or the evidence does not support renewal.
Do not let a past payment make the next decision. The money already spent cannot be recovered by keeping a weak tool. Review the next 90 days, not the last invoice.
Questions to ask before renewal
- Who used the tool, and for which repeated task?
- What was the baseline before it was added?
- What changed in time, quality, capacity, or customer experience?
- How much human review did the result require?
- Would the business miss the tool next week if access ended?
The final question is a useful reality check. If nobody can name the next important job, the subscription is probably serving the idea of an AI business rather than the business itself. For a broader audit, read AI Tools You Can Cancel Right Now.
Evaluate tools with the same care you give any contractor or process. Your own numbers, your own standards, and your own customer commitments are more useful than hype. If you want a system for choosing and managing practical AI workflows, learn the AI System.