For the past two years, every small and medium-sized business and mid-sized company we’ve come across has been told the same thing: “You have to adopt AI.” Many have launched a pilot project, seen an impressive demo, and then realized that nothing had changed in their operations. The problem is almost never the model itself. It lies in the choice of process.

The right question isn’t “Which AI?” but “Which process?”

A successful AI automation project rarely looks like an AI project. It looks like a project to streamline a tedious task, in which AI replaces a specific step—often one involving reading, classification, or writing—within a sequence of deterministic rules.

In other words: you’re not automating “customer service.” You’re automating “the qualification and routing of incoming emails, with a pre-written response for the six most common reasons for contact, which account for 70% of the volume.” It’s much less impressive in a meeting. It’s infinitely more profitable.

The qualification grid: five criteria, and all of them are required

Before writing a single line of code, we run every automation candidate through this grid. A process that fails on even one criterion isn’t ready.

The volume is real and recurring. With fewer than 50 occurrences per week, automation will cost more than performing the task manually. We always start with high-volume, repetitive tasks.
The result must be verifiable. We must be able to say “it’s correct” or “it’s incorrect” without debate. If the quality of the result is a matter of opinion, you won’t be able to test, measure, or correct it.
The input data already exists in a usable format. Scanned PDFs in a shared folder aren’t usable data—it’s a digitization project in disguise.
The cost of error is limited. A poorly written product description can be corrected. A mistakenly initiated bank transfer cannot. We start with processes where errors are visible and reversible.
A human remains in the loop at the start. For the first three months, the AI makes suggestions, and the human approves them. This approval rate becomes your trust metric and determines how gradually you grant autonomy.

Five automations that pay for themselves in a few months

These are the projects we see most often come to fruition among retailers, e-commerce businesses, and industrial SMEs.

Automation	What AI Actually Does	Implementation
Product Listing Enrichment	Generates titles, descriptions, and structured attributes from raw supplier data	2 to 4 weeks
Processing supplier files	Reads spreadsheets in various formats and maps them to your data model	3 to 6 weeks
Qualification of incoming requests	Classifies, extracts key information, routes to the appropriate team, and suggests a response	4 to 6 weeks
Data reconciliation	Detects duplicates and discrepancies between two data sources (catalog, inventory, accounting)	3 to 5 weeks
Content translation and adaptation	Translated while preserving industry-specific terminology and adhering to format requirements for each channel	1 to 3 weeks

What these five projects have in common: they involve data you already have, produce a result that a human can review in a matter of seconds, and fit into an existing process without requiring a reorganization.

What You Shouldn’t Automate Right Away

Anything involving money without validation. Refunds, credit memos, rate adjustments: AI makes suggestions, a human validates them—until you have months of historical data.
Anything based on dirty data. Automating on top of an inconsistent catalog just leads to faster errors. Clean up the data repository first—that’s often where the real project lies.
Anything without a business owner. Without someone whose daily work is visibly improving, the pilot project will have no one to defend it during budget negotiations.
Rare, high-stakes decisions. AI excels at repetitive, high-volume tasks but performs poorly in one-off cases where every parameter matters.

The real budget: what no one quantifies at the outset

The cost of inference—the infamous tokens—is generally the smallest line item in the budget. What really costs money is:

Integration with existing systems (ERP, PIM, CRM, email accounts): often 50 to 70% of the project workload.
The test dataset: about a hundred annotated real-world cases—without which you’ll never know if a change improves or degrades the result.
Observability: logging of every decision, cost per execution, and human intervention rate. Without this, you’re flying blind.
Handling failed cases: there will always be 5 to 15%. They must end up in a queue handled by a human, not disappear silently.

A 90-Day Plan That Works

Days 1–15 — Scope definition. A single process, selected using the grid above. Baseline measurement: how much time, how many errors, how many people. Without this baseline measurement, no ROI can be demonstrated.

Days 16–45 — Closed prototype. The AI runs on real data but doesn’t write anything to any system. We compare its output to that of humans on the test dataset. Goal: achieve over 90% agreement on standard cases.

Days 46–70 — Assisted rollout. The AI makes suggestions, and the team validates them with a single click. We measure the correction rate and use it to prioritize improvements.

Days 71 to 90 — Gradual rollout. Case categories where the correction rate is close to zero are processed automatically. The others remain in the validation queue. This threshold shifts every month.

Frequently Asked Questions

Do you need a data lake before getting started?

No. That’s the best way to delay the project by eighteen months. An initial use case requires direct access to one or two sources—not a complete overhaul of the information system.

Do you need to train your own model?

In the vast majority of business cases, no. A well-trained general-purpose model, fed with your data at the time of the query, meets the need. Specific training is justified for very large data volumes or a domain far removed from everyday language.

Does our data go to the AI provider?

That depends entirely on the chosen architecture. It’s possible to work with your own vendor credentials, route sensitive processing to a model hosted in France, or run an open model on your own infrastructure. This is an architectural decision that must be made during the scoping phase, not afterward.

How long before we see results?

With a well-chosen process, the team notices a difference as early as the assisted phase, around the second month. Quantifiable gains become evident after a full quarter of production.

Conclusion

AI automation isn’t a technology project; it’s an operations project with a technology component. Successful companies aren’t the ones that chose the best model; they’re the ones that chose the best process—well-defined, measurable, and with a motivated business owner.

For the next part of this series, read why 88% of AI agent pilots never make it to production. And if you’re looking for the right first process to automate in your organization, let’s talk: we conduct this scoping exercise regularly, and it takes just half a day.

AI Automation: Where to Start If You're Not a Big Tech Company