How to Find Business Processes Worth Transforming with AI

“Where can we use AI?” is the wrong question. It sends you looking for a place to put a tool you’ve already decided to use. The useful question is different: which part of the operation costs more than it should — and what kind of solution actually fixes it? This guide shows how to answer it with four criteria and a worksheet you can run with your team this week.

What makes a process worth transforming

A process is worth transforming when three things are true at once: changing it moves a business outcome you can measure; your data, systems and people can support the change today; and the cost of getting it wrong is manageable. If one is missing, the idea isn’t bad — it just isn’t ready.

Transforming a process doesn’t mean adding AI to it. You can also transform it with software that has no AI in it, by automating repetitive steps, by redesigning how the work gets done — or by cleaning up its data first.

Why decide before you build

• Ideas are cheap; budget and engineering hours aren’t. Every project you pick crowds out another one.
• The most visible opportunity is rarely the most valuable. A chatbot demos well; a reconciliation that eats hours every week doesn’t. The value usually hides in the boring steps.
• The expensive mistake isn’t failing to use AI. It’s carefully building something that never moved a number.

The framework: four criteria, four decisions

Criterion	The question	It’s high when…	It’s low when…
Impact	How much does it move a measurable business outcome?	It changes cost, revenue, time or risk in a way leadership would notice	The improvement is cosmetic or can’t be measured
Effort	How much work, change and cost does it take?	It needs new systems, several teams or a data migration	It builds on what exists and changes few habits
Risk	What happens if it’s wrong, and who catches it?	Errors are silent, costly or regulated	Errors are visible, reversible and reviewed by someone
Feasibility	Do today’s data, systems and people allow it?	Data is recorded reliably and the process has an owner	There’s no data, no owner, or the process changes every month

Those four criteria put every candidate into one of four decisions:

• Do now: high impact, contained effort, manageable risk, feasible today.
• Plan: high impact but more effort. It deserves a definition and a budget, not an improvised experiment.
• Test small: the impact is uncertain. Run a bounded experiment — with a date and a metric — before investing.
• Not yet: feasibility or control is missing. It isn’t a permanent no; it’s the list of what has to change first.

We deliberately avoid decimal scores and weightings. Rating high, medium or low forces a conversation about why; a 7.4 implies a precision nobody has.

How to apply it, step by step

1. Where does the operation hurt — not where does the technology shine?

Talk to the people who do the work, not only the people who manage it. Look for five signals: waiting, rework, errors someone fixes by hand, decisions that depend on who’s on shift, and information copied from one system into another.

2. Can you write each candidate in one line, with its metric?

Use this format: “[Process] currently [takes, costs or fails] [how much]; if it changed, [metric] would improve.” If you can’t fill in the last part, it isn’t an opportunity yet — it’s an annoyance.

3. How much impact, and how much effort?

Rate each one high, medium or low. When two people disagree, don’t average — talk it through. Disagreement almost always exposes different assumptions about the process.

4. How risky and how feasible is it — before you fall in love with it?

Ask what data you’d need and where it lives, who would notice an error and what it would cost, and who owns the process. This is the question that stops the most ideas — and it’s better to stop them here than halfway through a project.

5. Which quadrant does it land in?

Place each candidate in Do now, Plan, Test small or Not yet. Within each group, risk and feasibility set the order.

6. What kind of solution fits best?

Choose the kind of solution before you choose the tool:

• Software without AI: a clear rule or a well-designed workflow solves it.
• Automation: the work is repetitive, stable and well defined.
• Process redesign: the problem is how the work gets done, not which tool does it.
• Data first: the information you need isn’t recorded, or can’t be trusted.
• AI: there are patterns a rule can’t capture, enough data, and someone to verify the output.
• Test small or nothing yet, when that’s the honest answer.

7. What’s the first step, and how will you know it worked?

For every opportunity in “Do now,” define what gets built or changed first, who owns it, and which number has to move. Without those three answers, it isn’t ready to start.

An illustrative example

A mid-sized distributor (an illustrative example, not a client) reviews five candidates. Ratings follow the order impact · effort · risk · feasibility.

Candidate and metric	Ratings	Solution type	Decision
Matching invoices to payments — weekly hours of manual work	High · low · low · high	Software without AI (matching rules)	Do now
Hand-built sales reports — days of delay on the monthly report	Medium · low · low · high	Automation	Do now
Demand forecasting — stockouts	High · high · medium · medium	AI, after cleaning up sales history	Plan
Customer self-service portal — order-status calls	Uncertain · medium · low · high	Software	Test small
AI support assistant — response time	Medium · medium · high · low	AI	Not yet: there’s no reliable knowledge base

Of five candidates, two use AI — and neither is in “Do now.” That isn’t a verdict against AI. It’s what normally happens when you start with the problem.

A real case: the metric first, the foundation before the model

The same pattern shows up in real projects. At Grupo Entre Ríos, one of Guatemala’s three largest rubber exporters, pickup trucks left using only 52% of their capacity on average. The success metric was truck utilization, not model accuracy. And before any model, the team built the foundation that makes one possible: the platform that records rubber intake, budgeting and maintenance. With that foundation in production, an AI model began planning weekly pickup from production, weather and demand — and utilization rose to at least 85%.

That project didn’t come out of an assessment in this format. We show it because it follows the same order: the metric first, and the data before the model.

Common mistakes

• Starting with the tool and looking for a problem afterward.
• Picking the most visible idea instead of the most expensive problem.
• Not writing down the metric. Without it, there’s no way to know whether it worked.
• Assuming the data exists without asking who records it and how reliably.
• Automating a broken process as-is instead of redesigning it.
• Launching pilots with no owner and no end date.
• Treating “Not yet” as a failure. It’s often the recommendation that saves the most money.