How to Know Which Processes in Your Company Should Use Artificial Intelligence (and Which Shouldn't)

The problem isn't the technology, it's choosing the right process

Most AI projects that fail don't fail because of a poorly trained model or a lack of budget. They fail because they were applied to the wrong process: one that runs three times a month, one that depends on criteria nobody ever wrote down, or one that simply wasn't hurting anyone.

Before asking which tool to use, it's worth asking a more boring but far more profitable question: which of my current processes actually justifies the effort? Answering that well saves you months of work and uncomfortable conversations with leadership.

This article offers a concrete filter. It's not a magic formula, it's an order of questions you can work through in a two-hour meeting with the people who actually do the work.

The five filters a process must pass

A process is a good candidate for AI when it meets most of these conditions. You don't need all five, but if it fails three or more, you're probably looking at an expensive project with little return.

• **Volume and repetition.** It happens dozens or hundreds of times a week. Classifying 400 support emails a day is worth it; classifying 12 isn't.

• **Available, digital data.** There's a history in some system, not in someone's head or in physical folders. If you have three years of tagged tickets, you have raw material. If you have a spreadsheet that each branch fills out its own way, cleanup comes first.

• **Describable criteria.** Someone can explain how the decision is made, even if the rules are fuzzy. If the answer to "how do you decide this?" is "it depends, you just know," there's documentation work to do first.

• **Contained tolerance for error.** An occasional mistake causes an inconvenience, not a lawsuit or a serious loss. Or there's a human review step that catches it before it goes out.

• **Measurable impact.** You can name the metric that would improve: person-hours, response time, error rate, cost per transaction. If you don't know which number should move, you won't know whether it worked.

Where AI tends to pay off fast

In practice, certain families of processes show up again and again as first wins in mid-sized and large companies across the region. They aren't the flashiest, but they're the ones that build the internal credibility to keep investing.

• **Document reading and extraction.** Supplier invoices, dispatch notes, policies, contracts. An accounting team keying in 600 invoices a month can go from 15 minutes per document to 2 minutes of review.

• **Classification and routing.** Emails, tickets, complaints, job applications. The model doesn't resolve the case, it sends it to the right team with a suggested priority.

• **First-line support.** Common questions about order status, hours, requirements or internal policies. It works well when there's a real document base and a clear escalation path to a person.

• **Internal knowledge search.** Technical or legal teams that lose hours digging through manuals, regulations or past records. An assistant trained on your own documentation cuts that time noticeably.

• **Draft generation.** Sales replies, product descriptions, meeting summaries. The key is that a human always edits and approves.

Signs that a process isn't right for AI (at least not yet)

Saying no in time is part of the job. These are the most common red flags we see when reviewing initiatives that already had budget assigned.

• **The process is broken, not slow.** If the business rules contradict each other or nobody agrees on what the correct outcome is, AI will only automate the mess faster.

• **A simple rule solves it.** If 90% of cases are decided by three conditions, write the three conditions. Traditional automation costs a fraction and is 100% auditable.

• **The cost of error is high and irreversible.** Credit decisions without review, diagnoses, dismissals, final tax calculations. AI can support there, but never decide alone.

• **There's no data, or it's on paper.** You can build it, but that's a different project, with a different timeline and a different budget. Better to call it what it is.

• **Nobody owns the process.** If there's no one with the authority to change how the work gets done, the technical solution won't be adopted. It's the most common reason pilots die in December.

• **Low frequency.** A quarterly process rarely justifies the maintenance an AI system in production demands.

How to prioritize when you have ten candidates

Once you open the conversation, more ideas almost always surface than you can execute. A simple matrix organizes the discussion and keeps the loudest voice from winning.

Ask each area to estimate two things per process: approximate annual impact (hours saved × hourly cost, or losses avoided) and implementation difficulty (data quality, required integrations, expected resistance). Place each candidate in a quadrant.

The first project should come from the high-impact, low-difficulty quadrant, even if it isn't the most interesting one technically. You need a visible result in 8 to 12 weeks, not a perfect platform in 18 months.

A concrete example: a distributor with 40 people in back office listed eight processes. The most appealing was a demand forecasting model; the one they chose was automatic validation of purchase orders against the catalog, which consumed 90 hours a week across three people. It shipped in ten weeks, freed up the team and funded the demand conversation the following year.

Define success before you start

A pilot without written success criteria is a pilot that never ends. Before the first line of code, put four things in writing: today's baseline metric, the target, the evaluation deadline and who decides whether it continues or stops.

It also defines the acceptable quality threshold. A classifier with 85% accuracy can be excellent if the remaining 15% lands in a human review queue, and unacceptable if it goes straight to the customer. That's a business conversation, not a technical one.

And budget for what comes next. An AI system isn't a project you deliver and forget: it requires monitoring, occasional retraining and adjustments when processes or vendors change. With no one looking after it, performance quietly degrades over months.

Finally, mind the human side. The people who know the process best are also the ones who most fear being replaced. Involving them from the diagnostic stage, and being honest about what changes in their day-to-day, does more for adoption than any demo.

A sensible starting point

If you're just getting started, you don't need a corporate AI strategy. You need an honest list of processes, five filters applied rigorously and a first use case you can measure before the quarter ends.

At Da2 Group we support this kind of diagnostic regularly: mapping processes, assessing the real quality of your data and separating what's worth automating from what simply needs fixing. Sometimes the recommendation is not to use AI yet, and that's a useful outcome too.

If you have a process in mind and want a second opinion before committing budget, let's talk. An hour of review usually clears up a lot.