Before You Invest in AI, Ask These 7 Questions.

AI demos have a way of making difficult work look effortless. A customer asks a question, a system produces a polished answer, and the room starts calculating all the hours it could save.

Then someone asks a less exciting question: Who checks the answer?

Perhaps the information came from an outdated policy. Perhaps the customer’s request is unusual. Perhaps a confident but incorrect response would be more costly than the time saved on ten straightforward ones.

The demo may still be excellent. It just hasn’t answered the investment question.

That gap matters as AI moves from experimentation into ordinary business decisions. Whether you’re considering an off-the-shelf assistant, a predictive tool, or a custom system, you need to know more than what the technology can do under favourable conditions.

You need to know what it will improve, what it will demand from your organization, and how you’ll recognize when it isn’t working.

In brief:

• Define the business problem before selecting a technology.
• Compare AI with a simpler alternative.
• Test performance, workflow fit, risk and full cost before scaling.

QUESTION 01 / 07: What problem are we actually trying to solve?

“We want to use AI” is an ambition, not a problem statement. “We need to be more efficient” isn’t much better.

Consider a company whose staff spend days preparing monthly performance reports. It seems sensible to automate report writing. But where does the time go? Finding information across systems? Cleaning inconsistent figures? Checking calculations? Writing explanations? Getting approvals?

If collecting and reconciling data takes most of the effort, a tool that drafts the final paragraph may be impressive and largely irrelevant.

Describe the current process before choosing a solution. Record the time it takes, where mistakes happen, which decisions depend on it, and who is affected.

A stronger objective would be: Reduce the time required to prepare monthly reports while preserving accuracy and improving the information managers can act on. That leaves open the possibility that data integration, a different workflow, or a smaller automation is the better investment.

QUESTION 02 / 07: Why AI instead of something simpler?

The meaningful comparison isn’t AI versus doing nothing. It’s AI versus the best available alternative.

An organization might want AI to categorize incoming requests. But could a clearer intake form, a rules-based workflow, or a feature already included in its software deliver most of the benefit?

There are good reasons to consider AI, particularly when tasks involve messy documents, natural language, changing patterns, or variation that is difficult to capture with fixed rules. Even then, the answer may be a combination of AI, conventional software, and human judgment.

Google’s Rules of Machine Learning makes a similar engineering point: begin with a clear objective and a robust, simple approach before adding complexity.

Complexity has a carrying cost. It creates dependencies, maintenance work, and more ways for something to fail.

AI has to earn its complexity.

Before selecting a product, put a simpler alternative on the table. If it can deliver most of the expected improvement, be clear about the additional value AI would need to provide.

QUESTION 03 / 07: Do we have suitable information, and are we allowed to use it?

“Lots of data” does not necessarily mean “AI-ready.”

Imagine building an assistant to answer employee questions from company documents. The files exist, but some are outdated. Two procedures contradict each other. Different teams use different versions of the same form. The system might retrieve all of it efficiently and still give an unreliable answer.

You need to understand what information the tool requires, where it comes from, who maintains it, and what happens when it is missing or wrong.

You also need to know what information the tool is permitted to process. Customer records, employee details, contracts, and confidential business material bring privacy and security responsibilities. Questions about vendors’ data retention, access controls, processing locations, and secondary uses should be answered before real records are shared.

The Office of the Privacy Commissioner of Canada emphasizes privacy safeguards, transparency, appropriate authority to use personal information, and limiting unnecessary disclosures.

Some projects can begin with a small set of carefully prepared documents. The objective isn’t to perfect every database in the company. It is to make sure the information for this specific use is reliable, appropriate, and governed.

QUESTION 04 / 07: How will we test it under conditions that resemble our work?

A good pilot should be capable of changing your mind.

If the system will summarize technical reports, don’t test it only with clean, familiar examples. Include long documents, contradictory evidence, ambiguous wording, and cases where the correct response is to acknowledge uncertainty.

Measure the complete task, not just the speed of generation. A summary produced in ten seconds isn’t a time saving if reviewing and repairing it takes longer than the original task.

The right measures depend on the application, but may include accuracy, the severity of errors, net time saved after review, consistency, human intervention, and whether decisions or customer outcomes actually improve.

For safety-related or other high-consequence applications, a small percentage of serious errors may be unacceptable even when average performance looks strong.

NIST’s AI Risk Management Framework addresses evaluation and ongoing monitoring in the context in which AI is deployed, not merely whether a model performs well in isolation.

QUESTION 05 / 07: What will change for the people doing the work?

Software doesn’t automatically create a better process.

If AI prepares a customer-service response, who checks it? Which requests need escalation? Who maintains the source material? Can the reviewer verify the underlying facts, or are they being asked to approve an answer they cannot properly assess?

An added review step can erase the time saved during drafting. A new interface can shift work onto employees whose schedules are already full. An automated recommendation may arrive too late for anyone to act on it.

Talk to the people doing the job before deciding how the tool should fit into it. Ask what they currently do when information is missing, when priorities conflict, or when an unusual case appears. Those informal practices may be essential to making the work run.

A system isn’t adopted because people have been told to use it. It earns a place in the workflow by helping them perform their responsibilities better.

QUESTION 06 / 07: What happens when the system is wrong?

Every tool can fail. With AI, the challenge can be recognizing failure when the output looks plausible.

A faulty draft of an internal event announcement and an incorrect recommendation about equipment maintenance have different consequences. They should not face the same level of scrutiny or operate under the same controls.

Think through the harm an error could cause, how likely staff are to notice it, who is authorized to act on outputs, and how the organization would respond. Also consider security, confidentiality, fairness, and changes in performance over time.

NIST’s Generative AI Profile identifies concerns including inaccurate outputs, information integrity, privacy, and overreliance on generated material.

“Human in the loop” isn’t enough on its own. Oversight must be meaningful: the person reviewing an output needs the expertise, evidence, authority, and time to intervene.

For higher-risk uses, define a clear process for monitoring, escalation, and stopping the system if required.

QUESTION 07 / 07: What will it cost after the demonstration, and what would make us stop?

A subscription price is not the full cost of AI.

Depending on the application, costs may include organizing data, integration, configuration, employee training, security reviews, vendor management, quality checks, ongoing monitoring, and support. A custom system can also create long-term maintenance obligations. Usage-based charges may rise as more people adopt the tool.

The benefits deserve the same scrutiny. Time saved is not automatically money saved. If employees reclaim an hour each day, will it let them accomplish more valuable work, avoid overtime, improve service, or reduce a measurable expense?

Agree on a decision rule before the pilot begins. What evidence would justify expanding it? What problems are fixable? Which outcomes would lead you to pause or stop?

The discipline to walk away matters as much as the enthusiasm to begin.

What a sensible first AI project might look like

Return to the customer-service assistant from the opening.

Instead of trying to automate every response, a company could test a narrow task: can the system accurately classify routine enquiries and suggest drafts that employees can verify quickly?

It could use appropriately prepared examples, compare results with the current process, and measure errors and review time. Customer-facing actions would remain under defined human control during the initial test.

If classification works but drafting proves unreliable, keep the useful part. If simple routing rules work just as well, use those instead. If the results depend on documents nobody maintains, solve that problem before expanding.

This kind of pilot may look less impressive in a sales presentation. It is far more useful when deciding what to do next.

Make the investment decision before falling in love with the technology

AI can help organizations work with complex information, find patterns, and redesign tasks that previously required substantial manual effort. Those opportunities are worth serious attention.

But accessibility doesn’t make every use case worthwhile. As tools become easier to buy and build, the quality of the decisions surrounding them matters more, not less.

The strongest starting point may be a modest application with a clearly defined problem, representative tests, realistic costs, and people who know how it will fit into their work.

At Inozee, we approach AI as one possible route to a useful outcome, alongside research, data analysis, design, and process improvement. The technology matters. So does knowing when it is unnecessary.

A compelling demonstration can earn an idea your attention. A credible business case should earn it your investment.