Artificial intelligence applied to business processes

1. What applying AI to a business process means

Artificial intelligence applied to business processes means inserting AI models into a workflow the company already runs, so they take over the steps that until now required a person to read, interpret or decide. It is not a proof of concept or a tool off to one side: it lives inside the process, connects to the ERP or the CRM, and is measured by its effect on that process.

The difference from "using AI" individually matters. Having the team consult a conversational assistant to write better is useful, but it does not change the process: the hours are still there, so are the errors, and nothing is recorded. Applying AI to a process means the step disappears from someone's working day.

A business process with AI in it usually has three pieces:

• An unstructured input that previously had to be read: a supplier PDF, an email, a call transcript, a photo of a job sheet.
• A model that interprets that input and returns something structured: extracted fields, a category, a summary, a proposed reply.
• An integration that executes the consequence in the business system, with a person supervising the doubtful cases.

2. AI, classic automation and RPA: what each one does

This is the most common confusion and the one that wastes the most budget. They are not competing: they solve different parts of the same process.

Classic automation and RPA execute deterministic rules: if a condition is met, an action fires. They are fast, cheap and perfectly predictable as long as the input always has the same format. The problem is variation: one supplier changes its invoice layout and the rule stops working.

AI contributes exactly what is missing there: the ability to interpret input that does not come in a fixed format, and to resolve ambiguous cases with judgement. In exchange it is probabilistic rather than deterministic, which means the process must be designed on the assumption that it will sometimes be wrong.

That is why the projects that work almost always combine the two: AI interprets and decides, rules execute and validate. A good sign that a project is badly framed is that it tries to solve everything with a model, with no rules or validation around it.

3. Which processes are good candidates

Not every process improves with AI. The ones that do normally meet these four conditions at the same time:

1. Volume. It repeats many times a month. A process that happens three times a year is not worth automating, however tedious it is.
2. Skilled time on mechanical tasks. Someone paid for their judgement spends hours copying data, sorting emails or searching through folders.
3. Unstructured information. There are documents, free text, emails or images involved. If everything is already in clean tables, classic automation is probably enough.
4. Tolerable, verifiable error. There is a way to tell whether the output is correct, and an occasional mistake can be caught and fixed without serious consequences.

When the third or fourth condition is missing, it usually means the process does not need AI but something else: rules, an integration or a redesign.

4. Cases by business function

The processes that most often become a first project, by function:

Administration and finance

• Invoice and delivery-note entry: extracting fields from a PDF or photo, matching against the purchase order and posting to the ERP. It is the most repeated case because volume is high and the result can be verified against the order.
• Reconciling receipts and payments against free-text bank descriptions.
• Preparing recurring reports that are assembled by hand today from several sources.

Customer service

• Classifying and routing inbound email: every message tagged by type, urgency and owner before anyone reads it.
• A proposed reply for repetitive queries, with a person reviewing and sending.
• Automatic conversation summary and CRM entry when the case is closed.

Sales and marketing

• Lead qualification using public company information and your own history, so sales spends time on the ones worth pursuing.
• Preparing quotes from a catalogue and the customer's terms.
• Spotting churn-risk signals in existing accounts.

Purchasing and suppliers

• Reading quotes and price lists in heterogeneous formats to compare them in a single table.
• Checking that what was received matches what was ordered and what was invoiced.
• Tracking expiry dates and renewals inside contracts.

Operations and logistics

• Digitising job sheets handwritten or photographed in the field.
• Classifying incidents by root cause from the operator's free-text description.
• Demand and stock forecasting supported by historical data.

Internal knowledge and quality

• Natural-language search over internal documentation: procedures, manuals, applicable regulation, project history.
• Checking a document against the template or procedure before it is issued.
• Supporting new joiners, who ask instead of searching.

5. How the rollout works, phase by phase

The order matters more than the technology chosen. This is the method we follow in our projects:

1. Map the current process. How it actually runs today, not how the procedure says it runs. This is where the invisible steps nobody had counted show up.
2. Measure the baseline. Current hours per month, number of cases, error rate and lead time. Without this number there is no way afterwards to show the project achieved anything.
3. Prioritise. Each candidate process is scored by impact and by effort. The first one should have the best ratio between the two, not be the flashiest.
4. Scoped pilot. A single process, a real subset of cases, and a success criterion written down before starting.
5. Go live with supervision. The system enters the real flow, with human review of the cases the model flags as doubtful and a record of everything it decides.
6. Measure and scale. The numbers are compared against the baseline and, if the result holds, the pattern is replicated in neighbouring processes.

The phase most often skipped, and the one that causes most trouble, is the second. Without a baseline, the discussion about whether the project worked ends up being a matter of opinion.

6. How the return is measured

An AI-on-processes project is justified with four metrics, all compared against the baseline:

• Hours freed per month and their equivalent cost. It is the direct saving and the easiest to defend to management.
• Error rate before and after, and the cost of each error: rework, credit notes, penalties, lost customers.
• End-to-end lead time. It is often worth more than the hours saved, because it unblocks invoicing.
• Capacity: how many more cases can be absorbed without hiring.

As a reference for what we typically see: around 20 hours per month per employee on automated tasks, error reductions of up to 90% in administrative processes, and payback under six months on the first project. These are ranges commonly seen depending on scope and starting point, not a promise: the number that counts is the one that comes out of measuring your own process.

7. When NOT to apply AI to a process

Saying no in time saves more money than most projects make. Cases where we advise against applying AI:

• The process is broken. If nobody knows who owns each step, or everyone does it differently, it has to be redesigned first. Automating a broken process only makes it fail faster and at greater scale.
• The rules are clear and stable. Classic automation is then cheaper, faster and more predictable. Adding a model only adds cost and uncertainty.
• Volume is low. Below a certain number of cases a month, the project never pays for itself even if it works perfectly.
• There is no way to verify correctness. If you cannot tell whether the output is right, you cannot detect when it starts to degrade either.
• The error is not tolerable. Where a mistake has serious, irreversible consequences, AI can propose, but the decision stays with a person.

8. What the regulation requires

Applying AI to internal processes carries specific obligations, and they are best resolved at design time rather than at the end. Two frameworks overlap:

Regulation (EU) 2024/1689 (the EU AI Act) classifies each system by risk tier. Most internal company processes — invoice reading, email classification, document search — fall under the transparency tier, not high risk. There are exceptions that do qualify and are worth identifying early: recruitment, performance evaluation and decisions affecting workers' rights.

And if the process handles personal data, the GDPR also applies, supervised in Spain by the AEPD, with its requirements on minimisation, legal basis and informing the people affected.

The full detail — risk tiers, application timeline and who supervises in Spain — is on our AI compliance page.

9. Frequently asked questions