How to implement AI in your business processes

A Brazilian packaging manufacturer, 420 people, three plants. The board approved an AI project in February with a sentence that sounded reasonable: start with the most critical process, which was quoting for special accounts. Seven months later the project had burned budget, produced three demos, and not one extra quote per week.

The diagnosis, once someone finally made it, had nothing to do with technology. The quoting process crossed five departments, had two steps that existed only to check the previous step, and nobody could say how long it took — internal estimates ranged from two to nine days. There was nothing to improve because there was nothing to measure.

01The right order is an old one #

Michael Hammer published an article in the Harvard Business Review in 1990 whose title was already the thesis: 'Don't Automate, Obliterate'. The argument is that computerizing a bad process merely mechanizes the waste, and that the meaningful gain comes from redesigning the work before applying technology to it. Thirty-six years later we swapped the word 'computerize' for 'add AI' and the mistake is identical.

Thomas Davenport, in The AI Advantage (MIT Press, 2018), reached the same point by another route. Studying corporate adoption, he argues for what he calls an incremental approach: instead of betting on one transformational project, accumulate small, tightly scoped uses that actually reach production. In the article he wrote with Rajeev Ronanki for HBR, also in 2018, the authors record that the projects with the lowest failure rate were those automating structured back-office tasks, not those doing sophisticated prediction.

The two sources together fit into one sequence: map the process, strip the waste, measure the baseline, and only then put AI on a small scope, with an owner and a success criterion defined before anything is switched on.

02The four classic mistakes #

1. Starting with the hardest case. The most critical process has the most exceptions, the most departments involved and the most people willing to block it. It is the worst place to learn how to do this.
2. No baseline. Without the number from before, the result afterward becomes opinion. And opinion about an AI project tends to be generous for three months and brutal from the sixth on.
3. No owner. An AI project without a person accountable for the process — not for the tool — does not survive the first exception. IT cannot own a commercial process.
4. No decision about what happens when the agent is wrong. It will be wrong. The question is whether the error is caught, by whom, how fast, and what gets reversed.

The fourth mistake costs the most internal credibility. An agent wrong in 4% of cases can be excellent, if those 4% are caught the same day and reversed with no damage. The same agent is unacceptable if the error only surfaces on the customer's invoice 40 days later. The difference isn't in the model: it's in how the check was designed.

03The 90-day roadmap #

Weeks 8 through 10 are the ones usually skipped, and the most valuable. Running in shadow costs little, exposes no customer, and reveals within two weeks whether the agent understands the problem — including the cases where it gets the right answer for the wrong reason.

Want to see how this looks inside a real operation? Explore the platform.

04The pilot's math, with assumptions in plain sight #

The scale of that packaging manufacturer — a mid-size Brazilian company (figures in Brazilian reais) — after re-picking the process: they moved from special quoting to triaging recurring purchase orders. Assumptions stated; redo it with your own numbers:

On those assumptions, the scope covers roughly 748 orders a month and frees up something close to 75 hours of triage per month. It is not a revolution: it is one and a half people's worth of capacity handed back to the part of the work that needs judgment. The point of the roadmap isn't to promise more than that — it is to make sure that number exists and can be checked, instead of becoming a slide-deck estimate.

05What sustains the second case #

The first case teaches the company how; the second decides whether it becomes practice. Three things sustain it: recording work where it happens, so the baseline for the next process already exists; documenting scope and permissions, in the spirit of the govern and measure functions of the NIST AI Risk Management Framework 1.0, from 2023; and a quarterly review comparing what was promised against what was observed, including switching off whatever didn't pay.

It is also worth choosing the next case by the same criteria as the first, rather than by the enthusiasm of whoever watched the demo. High volume, stable rule, reversible consequence. The hard processes come later, once the company knows how to measure and how to switch things off.

At Relevanti, mapping the process, measuring the baseline and putting the agent inside it happen in the same place — it is worth rereading what an AI agent is — and what it isn't before choosing the pilot; see how on the platform, by department under solutions, or bring a specific process to a conversation.

Hammer's advice survives intact: don't automate, redesign. Davenport's completes it: start small, but ship it. Between the two sits a quarter of unglamorous work, which is exactly what separates a pilot from a result.