automatización procesos framework pyme

Which processes to automate first: The 2-axis framework

2026-08-24T07:15:04.122+00:00 · CompaniesAutomation

The framework for deciding which processes to automate first: cross hours/month with human judgment, apply the process design filter, and conduct a 5-step inventory with your team this week. Including what to expect from the first deployment.

To decide which processes to automate first, you only need to cross two axes: how many hours per month the process consumes and how much human judgment each case requires. The winning quadrant is always the same — many hours and low judgment — because it combines the fastest return with the lowest risk. Everything else, including what makes the most noise on LinkedIn, comes later.

Most companies don't fail at automating; they fail at choosing what to automate. They start with the flashy (a website chatbot), with what the vendor of the moment proposes, or with an idea from the last committee — and six months later there is no measurable saving to show. This article provides the complete framework for choosing with data: the two axes, the third filter that almost everyone skips, the signs of a good and bad candidate, and a 5-step inventory exercise you can do with your team this week.

What are the two axes for prioritizing processes?

The first axis is volume: how many person-hours per month the process consumes today. The second is judgment: how much human judgment is required to resolve each case, scored from 1 (fixed rules, always the same) to 5 (expert judgment, every case is different). With those two numbers per process, prioritization practically does itself.

The hours axis translates directly into money. A process that consumes 60 hours/month of an administrative profile costs, with total employer costs in Spain, between €1,500 and €2,700 per month — between €18,000 and €32,000 per year. That number is your savings ceiling and the basis of any serious AI ROI calculation. Without it, any automation decision is a gamble.

The judgment axis measures risk. A process with level 1-2 judgment (validating invoices against orders, recording data, answering questions with documented answers) can be automated with confidence. A process with level 4-5 judgment (negotiating with a supplier, resolving a delicate complaint, deciding on a commercial exception) is not automated: at most, it is assisted.

The quadrant map: where the money is

Crossing the two axes produces four quadrants, and each requires a different play. Knowing the four avoids the two classic mistakes: automating what shouldn't be and not automating what should.

Many hours + low judgment: automate now. The winning quadrant. Invoice reconciliation, order registration, recurring reports, answering repetitive queries. Measurable return in the first quarter and low risk, because the rules are clear.
Many hours + high judgment: assist, don't replace. The agent prepares 80% of the work — gathering data, drafting the proposal, suggesting the decision — and the human decides. Most hours are saved without surrendering judgment. This is the right play for quotes, offers, and complex incident management.
Few hours + low judgment: group for the second wave. Individually they don't justify a project, but five tasks of 8 hours/month add up to 40 hours. They are grouped and automated in batches once the first deployment has paid for itself.
Few hours + high judgment: leave it alone. Strategic decisions, one-off negotiations, key relationships. Automating here costs more than it saves and adds risk where it's least needed. This is where human judgment yields the most: let it stay there.

The third filter: is the process well-designed?

Before automating a process, ask if the process deserves to exist as it is today. If you automate a chaotic process, you get chaos faster: automation amplifies what is there, good or bad. This filter discards or reorders more candidates than it seems, and skipping it is the most expensive mistake on the entire list.

The signs of a poorly designed process are recognizable: approvals that bounce between three people without any of them adding value, data copied by hand between two systems because "it's always been done that way," steps that exist to compensate for an error no one remembers anymore. The correct sequence is to redesign first and then automate — and sometimes the redesign alone recovers half the hours. In that review, another uncomfortable truth emerges: not everything needs AI. If a case is solved with a fixed rule, a fixed rule is cheaper, faster, and more predictable than a model. Use the simplest tool that works.

What signs mark a good automation candidate?

A good candidate combines four signs: high volume, mostly stable rules, measurable cost, and accessible data. The more signals it checks, the shorter the path to ROI.

High volume: tens or hundreds of cases per month. Savings per case are multiplied and the investment is diluted.
Mostly stable rules: 80-90% of cases follow the same pattern and exceptions are identifiable. You don't need 100%: exceptions are escalated to a person.
Measurable cost: you can say how many hours and how many euros it costs today. Without a baseline, you won't be able to prove savings — and what isn't proven doesn't get budget for the next phase.
Accessible data: the information lives in digital systems (ERP, CRM, email, spreadsheets) that an agent can connect to. They don't need to be perfect; they just need to be there.

And the signs of a bad candidate, mirrored: low volume that will never pay off the project, expert judgment in every case, a process in constant flux (you'll automate a version that won't exist by the time you're done), and data scattered on paper, unscanned PDFs, or in two people's heads. If your favorite candidate fails two or more of these, drop it from the list without regret: the list of good ones is longer than you think. To calibrate what is typical in each area, it helps to review these AI agent use cases by department.

The 5-step inventory exercise (to do this week)

This exercise requires a 60-90 minute meeting with department heads and a spreadsheet. The result is a prioritized list of 2-3 processes with numbers behind them — the starting point for any serious automation.

List repetitive tasks (20-30 minutes). Each manager lists tasks their team repeats every week or month, without filtering yet: a typical mid-sized company finds between 15 and 30. The question that unlocks the list: "what do you do every week that bores you?"
Estimate hours/month for each. Rough estimate: people involved × hours per week × 4. Don't look for consultant-level precision; look for an order of magnitude. Distinguishing a 10-hour process from an 80-hour one is enough to prioritize.
Score judgment from 1 to 5. 1 = always resolved the same way, with rules that fit on one page; 5 = each case requires expert judgment. If two people on the team disagree by more than one point, the process isn't clear — take note of that, as it is also a finding.
Cross the two axes. Sort the sheet by hours descending and filter for judgment ≤ 2. What remains at the top is your winning quadrant. Add a column for annual cost (hours × 12 × hourly cost) to talk in euros, not feelings.
Choose 2-3 winners and apply the final filters. Is the process well-designed or does it need a redesign first? Is the data accessible in systems? The ones that survive are your first candidates — not ten, not one: two or three, to concentrate effort and be able to compare results.

What to expect from the first deployment?

A well-chosen first process is automated in 4-8 weeks, not quarters. The reasonable sequence: measure the baseline (current hours, cost, and error rate), deploy the agent on the scoped process, and compare against that baseline after 4-6 weeks of operation. The goal of the first deployment isn't just savings: it's to prove the method with numbers, so the next phase is funded by the savings of the first. How to link those phases is detailed in the 90-day roadmap for implementing AI in your company.

We apply this same framework in our own businesses before anyone else's — we are our own first client — and the constant lesson is the same: success is decided in the choice of the process more than in the technology. A poorly chosen process with the best AI yields mediocre results; a well-chosen process with standard technology yields savings that show up on the bottom line.

If you prefer to do this inventory with someone who has done it dozens of times — and walk away with validated numbers and 2-3 dimensioned candidates — that is exactly what we do in the diagnostic phase of our AI consultancy: 2-4 weeks, an inventory prioritized by ROI and a first deployment plan that stands up before any committee.

Frequently Asked Questions

What processes are automated first in an SME?

High-volume administrative ones: invoice reconciliation and registration, order entry, recurring reporting, responding to repetitive customer inquiries, and updating data between systems. These are processes with many hours, stable rules, and data that is already digital — the winning quadrant in almost any sector.

How many processes should be automated at once?

Two or three in the first deployment. Just one leaves you without comparison and at the mercy of that process's particularities; more than three scatters the team's effort and complicates measurement. With 2-3 scoped processes, you prove the method and decide on expansion with data.

Do I need to have perfect data before automating?

No. You need accessible data — living in systems an agent can connect to — not perfect data. In fact, the agent itself often uncovers and corrects inconsistencies while processing history. Waiting to "clean the data" is the excuse that delays projects for years; cleaning is done on the live process.

Does automating processes mean firing people?

In SME practice, almost never: the freed-up hours are reassigned to higher-judgment work — sales, customer care, improvements — that was being neglected. The common scarcity in a typical company is qualified time, not a lack of pending tasks. The high-judgment quadrant always has room.

What if my process changes constantly?

Then it's not a candidate yet. Automating a process in constant flux means rebuilding the automation every month. Stabilize the process first — or automate only the core that doesn't change — and leave the volatile part in human hands until it settles.