7 Questions to Answer Before You Implement AI
Key takeaways
- The fastest way to waste an AI budget is to start with the tool. Start with these seven questions instead.
- Most "no, not yet" answers point at a fixable input — a process, a dataset, an owner — not at AI being wrong for you.
- Run the list against one specific workflow, not "AI for the business" in the abstract.
- Seven clear answers is a green light. Any blank is the first thing to work on.
Most AI projects don't fail at the model. They fail at the questions nobody asked before the model showed up. The good news: those questions are predictable, and you can answer them in an afternoon for any given use case.
Run this list against one specific workflow — the invoice approval, the support triage, the proposal drafting — not "AI for the company." Abstractions can't be assessed; a concrete workflow can.
1. What specific problem are we solving?
"We need an AI strategy" is a mandate, not a problem. A problem is "our team spends a day a week re-keying order data between two systems." If you can't state the problem in one sentence a frontline employee would recognize, you're not ready to pick a solution — you're shopping for one. Name the friction first.
2. How will we measure success?
Decide the number before you build, because it shapes everything that follows. Hours returned per week, error rate, response time, cost per transaction, percentage of cases handled without a human. A measurable target keeps the project honest and tells you when to stop tuning. If the only available metric is "it feels faster," treat that as a warning sign, not a result.
3. Is the process stable enough to build on?
AI encodes a process; it doesn't invent one. If five people do the task five different ways, there's no single thing to automate — and whatever you build will be wrong for most of them. Where a process is genuinely unstable, standardizing it first is usually the higher-return move, and sometimes that alone captures most of the value you were chasing.
4. Is the data reachable and trusted?
Two parts, both required. Reachable: can you actually get to the data without a six-month integration? Trusted: do the people who'll use the output already believe the underlying data? You don't need perfect data — you need data good enough for the decision at hand and trusted enough that someone will act on what the AI produces. Output built on data the team quietly distrusts gets ignored, no matter how good the model is.
5. Who owns it after launch?
An AI workflow is a living thing. Inputs shift, new edge cases appear, prompts and rules need adjusting. If ownership ends when the project team leaves, quality decays until people route around the tool. Name an owner on the operations side — someone who uses the workflow and has the authority to change it — before you launch, not after it starts drifting.
6. What happens when the AI is wrong?
It will be wrong sometimes. The question is what that costs and who catches it. A misfiled internal note is cheap; an incorrect number sent to a client is not. Decide in advance where a human reviews output, where the system can act on its own, and how mistakes get caught and corrected. Designing for the error case is what makes an AI system safe to trust with real work.
7. Does the value beat the true cost?
The license fee is the visible cost. The real cost includes integration, the process and data cleanup, change management, and ongoing ownership. Put the honest total next to the measurable value from question two. If value clearly wins, build. If it's close, the use case probably isn't first in line — and that's useful to learn now rather than three months in.
Where teams trip up
The trap is treating these as paperwork to clear on the way to the tool you've already chosen. Answered honestly, they sometimes tell you the best next step isn't AI at all — it's standardizing a process or fixing a data source. That's not a failure of the exercise; that's the exercise working. A "not yet" that saves a failed deployment is worth more than a "yes" that rushes into one.
You don't have to run this alone. Working through these seven against your real workflows is most of what an AI opportunity assessment does — with the benefit of having seen which answers tend to be optimistic.
Frequently asked questions
What should you decide before implementing AI?
Before implementing AI, decide the specific problem you're solving, the measurable outcome that defines success, whether the target process is stable, whether the data is reachable and trusted, who owns the result after launch, how you'll handle the cases the AI gets wrong, and whether the value clearly beats the total cost. If you can't answer those, you're not ready to build yet.
Do you need clean data before implementing AI?
You need data that is reachable and trustworthy enough for the task — not perfect data. The right question is whether the people who will rely on the output already trust the underlying data today. If they don't, fixing that comes before implementation, because AI built on distrusted data produces output nobody acts on.
Who should own an AI workflow after it launches?
A named person on the operations side who uses the workflow, not just the team that built it. AI systems drift as inputs and edge cases change, so someone has to monitor quality and adjust. Naming that owner before launch is one of the strongest predictors of whether an AI implementation lasts.