Most companies know they should be doing something with AI. The hard part is knowing where and how to start.
The bottleneck we see, over and over, is almost never a lack of good ideas. It’s a lack of a shared, structured way to compare them and pick the best ones. Some processes sound exciting but have no data behind them. Others have great data but no clear owner. A few would save real money in theory, but nobody in the room can name a number.
To help leadership teams get past that bottleneck, we built the AI Opportunity Canvas: a one-page worksheet that walks a single candidate process through the specific questions we’ve found matter most. Fill it in for one process, and you have a written answer you can defend at the next board meeting.
The canvas is designed to be honest. Not every process that goes through it turns into a greenlight, and sometimes the most valuable answer is “not yet.”
In this article we walk through each of the seven sections, share the helper questions we ask in the room to get past vague answers, and use examples to show what a strong response actually looks like.
That way, by the time you sit down with the canvas yourself, you already know what to aim for.
1. Process: What Are You Actually Evaluating?
Almost every workshop we run starts the same way. Someone on the leadership team says they want to use AI for something broad, like customer service, marketing content, or contract review. The room nods. It sounds like a plan.
It isn’t. Each of those is a category, not a process. Under “customer service” sits ten different workflows, run by four different teams, using four different systems. You can’t pilot AI for a category. You have to pick one of the workflows inside it.
That’s what the first section of the canvas forces you to do. It asks for one specific, repeatable process that ends with a document or a decision. The key word here is repeatable. AI needs examples to work with.
A helper question we often use in the room: “If I gave this to a new hire on day one, could they tell me where it starts and where it ends?” If not, you don’t have a process yet. You have a category of related work, and it needs to be broken into pieces first.
So what should a good answer to “What are you actually evaluating?” look like? The exact workflow, in one or two sentences, with a clear trigger, a clear cadence, and a clear artefact. Something like:
“The responses our customer support team writes to product questions coming in through the help desk, roughly 40 tickets per week.”
Once you have that one workflow named clearly, the rest of the canvas has something concrete to work with.
2. Owner: Who’s Responsible for This Process?
After you’ve named the process, the next question is who runs it. This sounds like an easy one. In most workshops, it’s actually where the room gets quiet for the first time.
In many organisations, the honest answer is “everyone and no one” — multiple teams touch the process, nobody officially owns it. And a process without an owner today won’t suddenly get one just because AI is added to it. If anything, AI makes ownership matter more.
When AI joins a process, someone has to make judgment calls the software can’t:
Is this output good enough to send?
What do we do the first time it gets something wrong?
Without someone with the authority to answer those questions, the pilot stalls the moment the first unusual case shows up.
A helper question we use in this case is: “If the AI drafted something wrong tomorrow, who on your team would spot it, and who would decide what to do about it?” If people look at each other instead of naming someone, you have your answer.
So what should a good answer to ‘who owns this process?‘ look like?
For our support ticket example:
“Anna, Head of Customer Support. She sets response quality standards, reviews escalated tickets, and reports team performance to the COO each month.”
A department name won’t do. A role won’t do either. You need an actual person, named, with a sentence about what they’re already responsible for today.
3. Volume & Cost: What Does This Process Cost You Today?
The next question is about size. How much time, money, and effort does this process eat up right now?
This is where many leadership teams realise they’ve never actually measured it. The process happens, deadlines get met, people are busy. But the specific numbers behind it live in nobody’s head.
The canvas asks for rough figures across four things:
How many people work on this process?
How many hours does one cycle take?
How long does it run from start to finish?
What happens when something goes wrong?
You don’t need exact numbers here. Rough is fine. But you do need real ones, not guesses.
Why does this matter for AI? Because without a baseline, you have no way to prove the pilot worked. If nobody knows how long the process takes today, nobody can tell whether AI made it faster tomorrow. The pilot ends, everyone shrugs, and the budget quietly goes somewhere else next year.
Try to answer this question: “If this process disappeared tomorrow, how many hours a week would open up on your team’s calendar?” If the answer is “I honestly don’t know,” the first job isn’t AI. It’s basic measurement.
So what should a good answer to “what does this process cost you today?” look like?
For our support ticket example:
“Two full-time support agents handle around 40 tickets a week. Each ticket takes roughly 20 minutes to draft a response. When a response is wrong, the customer usually replies again, and that second round takes another 15 minutes plus a manager review.”
Rough numbers, but real ones. And now you have a baseline you can measure the pilot against.
4. Data: Where Does the Information Live?
Now it’s time to look at where the raw material sits. AI works on the documents, records, and data those processes produce. This section decides whether a pilot is actually possible in the next few months, or whether you’re looking at a much longer runway.
The canvas asks three things about the data behind your process:
Where does it physically live?
What language is it in?
Would anyone outside the company have a hard time accessing it?
Why does this matter for AI? Because the state of your data decides three things at once:
Whether a pilot can start at all
How long it will take to set up
How much of the budget goes to plumbing versus actually doing the work
A helper question we often use is: “If we needed to give a new employee access to everything they need to do this process well, how many systems and passwords would they need?” If the answer is more than three, the pilot is going to spend real time on data access before it spends any on AI.
So what should a good answer to “where does the information live?” look like?
For our support ticket example:
“All incoming tickets come through Zendesk, in English, and stay there. The last three years of tickets and responses are searchable in the same system. No external parties need access.”
That’s a green light for a pilot. If the answer had involved four different tools, two languages, and PDFs of old emails from an acquired company, the honest answer would be “not yet.”
5. Risk Check: Questions Before You Start
Now the questions nobody wants to answer, and everyone has to, before the project moves forward:
Who in the company will sign off on decisions or documents that AI suggests?
What happens if AI makes a mistake?
Who’s responsible when it does?
How will we catch it?
Where will the data going into the AI be stored, and who will have access to it?
Why do these matter for AI specifically? Because AI doesn’t take responsibility. When the model gets something wrong, the accountability still has to sit with a human, somewhere in the org chart. And when data flows into a model, it usually flows through a vendor’s infrastructure, which raises real questions about security, privacy, and compliance.
If the room can’t answer these questions today, the pilot isn’t ready today. Not because the process is bad, but because the governance around it doesn’t exist yet.
A helper question is: “If the CEO asked tomorrow who’s accountable for what this AI produces, do you have an answer they’d be satisfied with?” If the answer is “we’d have to figure that out,” those decisions need to happen before the pilot, not after.
So what should a good answer to the risk check look like?
For our support ticket example:
“Escalated responses are reviewed by a team lead before they go out. If a response goes wrong, the team lead corrects it and logs the case for training. Ticket data stays in Zendesk, which is already covered by our GDPR agreement.”
Three short answers, but each one names a person or a boundary. That’s what makes them useful.
6. Measurable Goal: How Will You Know It’s Working?
With the process, owner, cost, data, and risk on the page, the next question is: how will you know the pilot actually worked?
This is where most AI pilots go wrong before they even start. Teams launch with vague ambitions (“make things faster,” “reduce errors”) and then, six weeks later, nobody can agree on whether the pilot succeeded. The result usually gets described as “promising but hard to say for sure,” which is a polite way of saying nobody’s going to fund the next phase.
The canvas asks for a concrete, measurable outcome you can check in 4-6 weeks. Not “it’ll be faster,” but by how much. Not “fewer errors,” but from what to what.
The canvas even gives you three examples to work from:
Document prep time drops from X days to Y days
Error rate decreases by X%
Information lookup time drops from X minutes to Y minutes
Why does this matter? Without a number written down at the start, those wins are invisible when it’s time to defend the budget.
Here’s a question that usually helps: “If someone asked in six weeks whether the pilot worked, what single number would you point to?” If there isn’t one, you don’t have a goal yet. You have a hope.
So what should a good answer to “how will you know it’s working?” look like?
For our support ticket example:
“Average response time drops from 20 minutes to under 10 minutes per ticket, with at least 80% of AI-drafted responses accepted by the agent without significant edits.”
A specific number, measured against today’s baseline, checkable in six weeks. That’s what makes the goal real.
7. Minimum Scope: What’s the Smallest Version You Can Test?
One last question, and it’s often what saves the pilot from itself.
The temptation here is always to go big: cover the whole department, handle every ticket type, every document, from day one. That’s how most AI pilots die.
The canvas pushes in the opposite direction. It asks for the smallest meaningful pilot: one process, one team, one document type. Something you can actually complete and measure in four to six weeks.
The point of a pilot isn’t to prove that AI works. It’s to find out what breaks when AI meets your real processes, data, and team. A smaller pilot surfaces those problems in weeks instead of quarters.
One question we use to shrink the scope: “If we had to launch this in two weeks instead of six, what would we cut?” Whatever survives that cut is the minimum pilot.
So what should a good answer to “what’s the smallest version you can test?” look like?
For our support ticket example:
“For four weeks, AI drafts responses to one category of tickets (returns and refunds), for one team of two agents. Every AI draft is reviewed by the agent before sending. At the end of week four, we measure response time and acceptance rate against the baseline.”
Small enough to actually finish, big enough to learn something real.
Ready to Find Out If Your AI Idea Is Actually Ready?
Pick one AI process you’ve been thinking about and run it through the seven questions in this article. The AI Opportunity Canvas walks you through them in about 45 minutes on a single page. By the end, you’ll have a clear answer: go, not yet, or not here.
If you’d rather work through this with someone who’s done it many times, check out our AI strategy workshops. We help leadership teams turn their AI ideas into concrete pilots by scoring their actual processes against the same framework, in a single working session.Artykuł AI Opportunity Canvas: How to Evaluate if an AI Project Is Worth Doing pochodzi z serwisu DLabs.AI.