The subscription invoice tells you what the tool costs. It tells you very little about what AI is costing your business, or what you're getting back.
Your team may be using it every day. There may be promising projects, faster drafts, and demonstrations that would have seemed impossible a year ago. You can appreciate all of that and still want an answer to a reasonable business question: is this worth what we're putting into it?
That includes money. It also includes your people's time, the attention pulled from other priorities, and the work required to make the result dependable.
If you're still deciding where to invest, the same question applies before you spend. AI deserves a fair chance to prove useful, with the same business judgment you'd bring to any other investment.
Your people's experience belongs in that assessment from the start. They should help decide what changes, where their judgment matters, and what better work would look like.
The expense that doesn't appear on the invoice
A subscription is easy to see. The hours spent trying it, correcting its output, and starting over are harder to find.
Imagine an operations manager experimenting with AI to prepare a weekly report. The first version looks good. The next week, the source file changes. Someone spends an hour fixing the output. Another person creates a different version because they didn't know the first one existed.
Nothing here requires a spectacular failure. It's ordinary work happening around an unfinished idea.
Some experiments leave only a forgotten file. Others leave active subscriptions, duplicate processes, and someone quietly keeping the whole thing alive. Those costs belong in the investment decision, even when they sit outside the software budget.
Before you spend more, look at what you're already paying to keep going.
Give curiosity room, and give it a boundary
Exploration matters. People need space to try things before they can tell you exactly what those things are worth.
A short experiment might teach a team that a tool can't handle its documents reliably. That can be a useful result. So can discovering a simpler way to do the work without AI.
I'd give that learning a small, explicit budget of time and money, along with a question to answer. Could this help us respond to customer requests faster? Can it handle the exceptions we see every week? What would someone have to check?
Could it help us offer a service customers would pay for, serve more demand, or improve the margin on work we already do? What would we track to see whether that value materializes? For a new service, I'd track paying customers and the full cost of delivery.
Then set a date to look at what you've learned. Continuing, changing direction, and stopping are all legitimate decisions.
The problem starts when the experiment has no boundary, or when learning quietly turns into a production process that nobody agreed to support.
Saved time needs somewhere to go
Suppose, purely as an illustration, a team estimates that AI saves ten hours a week.
That's promising. It still leaves a few questions.
Does the estimate include checking the output and correcting mistakes? Is the time coming back in usable blocks, or as scattered minutes? Are people able to spend it on work that matters?
Ten hours of capacity doesn't automatically reduce payroll by ten hours. The same people may still be employed at the same cost, which may be exactly what you want.
The value might come from following up with customers who previously waited, reducing overtime, clearing a backlog, or handling more work without adding pressure to an already stretched team. Better service and a less exhausting day can matter before either shows up neatly in a financial report.
But name the intended benefit. If you're claiming a cash saving, identify the expense that will actually go away. If you're claiming more capacity, decide where it will go and check whether that happens.
Consider a hypothetical quoting team. An AI tool shortens the time it takes to prepare a draft quote, but every quote still waits two days for approval. Drafting got faster. The customer may see no difference.
Now suppose the team uses that recovered time to resolve missing information earlier, while the approver gets a clearer view of the exceptions that need a decision. That gives the project a plausible route to faster responses. You can measure it.
Whether that produces more profitable business still needs to be tested. Faster quotes alone don't prove additional revenue.
Price the work you will have to live with
The initial build is one part of the cost.
Someone needs to check the output, manage access, deal with exceptions, and keep the process working when the underlying tools or business requirements change. Your people will need time to learn it. A project that saves effort in one department can create work in another.
For a proposed AI project, I'd want a rough view of the first year's costs: setup, software, internal time, review, and maintenance. I'd also want to know who is responsible when it breaks.
Precision can improve as you learn. An honest range, with its assumptions visible, is more useful than a confident return calculation that leaves out half the work.
This is where financial discipline and respect for people belong in the same conversation. The person who understands the difficult customer, the unusual order, or the exception in the policy helps you see costs that a demonstration won't reveal.
Involve that person before you've decided what to automate. Their judgment may be the reason the process works at all.
A useful project can be the next step
You don't have to resolve every question about AI across the company before doing useful work.
A focused project can give you something concrete to learn from: one costly delay, recurring task, or missed opportunity; a person accountable for the outcome; and a clear way to compare what happens before and after.
Start by looking at the work as it exists. How often does it happen? What does it cost in time, rework, or lost opportunity? Where does human judgment matter? What would make the result worth the effort?
Then decide what kind of change it needs. You might use a capability you already pay for, simplify the process, build something specific, or decide the current approach is good enough.
For the quoting example, I'd first test a simpler approval process without AI. Otherwise, we could end up buying a tool and crediting it for an improvement we could have made anyway.
Then compare the AI-assisted version against that simpler process on representative work. Track turnaround time, corrections, and total staff effort, including review. Separate minor edits from errors that could put a bad price in front of a customer. A lower average correction rate doesn't excuse an occasional costly mistake.
Suppose the team gets drafts faster, but checking them takes just as long as preparing them used to. Customers aren't getting answers sooner, and the team has another subscription to manage. I wouldn't expand that version. I'd give a specific, fixable problem one bounded attempt, or stop.
If the AI-assisted process does reduce total effort and customers get accurate quotes sooner, there is a reason to continue. Whether there is a reason to expand depends on what that improvement is worth compared with the ongoing cost, and whether it holds up on difficult cases. Keep human approval for pricing commitments; a promising pilot doesn't justify handing that authority away.
That's the judgment I want to see before the next investment. An impressive output is a reason to investigate. Evidence that the work gets better is a reason to keep funding it.
Even when a project works, decide what another team would need to repeat it: usable instructions, the right access, someone to maintain it, and people who understand where its limits are. That's how one useful project begins to become a system your business can run.
You are allowed to ask for a better answer
If you're skeptical, curious, or wondering whether you're doing enough with AI, you don't need to settle that by counting tools or projects.
Look at one piece of work your business cares about. Ask what improvement would matter, what it would take to achieve it, and how you'd know.
There may be a very good AI project inside that conversation. There may also be a subscription to cancel or a process to simplify.
Any of those can be progress. The test is whether the business works better after the decision.
