AI Process Automation for Businesses

What automating a business process actually means

Automating a process means the system takes over work that a person currently does by following the same steps every time: logging an insurance claim that arrives by email, matching a carrier's invoice against the month's orders, copying the data from a document into a spreadsheet. People stop retyping and deal with the exceptions.

AI is needed when the incoming information has no fixed structure, such as a Spanish land registry extract (nota simple) or an email written any old way. When the data is already structured, a well-built integration is cheaper and more predictable than a language model, and it makes fewer mistakes.

If what you need is an application where today there is none, that is custom software. And if you want AI to operate inside your tools with its own permissions, the way a team member would, see AI agents for businesses.

RPA, traditional automation and AI: how they differ

They get sold as if they were the same thing, but they solve different problems. In our projects the norm is traditional integration, with AI only in the step that needs it.

What to look at	RPA	Traditional integration	AI
How it works	A bot repeats a person's clicks and keystrokes on screen.	Systems exchange data through APIs, webhooks, the database or scheduled file exports.	A model reads text or documents and returns structured data or a classification.
Data it handles	Whatever is visible on a screen.	Structured data in a known format.	Free text, PDFs, emails and images.
What breaks it	A screen redesign, a pop-up or a field that moves.	An API change or an expired credential.	Documents unlike the ones it was tested on, and answers that look right but aren't.
Running cost	Usually tool licences plus a machine to run it on.	Little beyond the server.	A cost per document or query that has to be measured.
When we use it	As a last resort: no API, no data access and no exports.	Whenever the data is already structured.	When something needs interpreting that a rule can't handle.

RPA has its place with closed systems, but all it takes is a vendor updating a screen for the bot to stop. Before going there we look for other ways in. Many ERPs without an API have a database that can be read in read-only mode: that is what we did for the Navarro Hermanos custom CRM, which syncs with the company's ERP twice a day.

AI, for its part, needs checks around it. When reading land registry extracts for Gesttia, every amount the model returns comes with the exact passage it was taken from, and a verifier compares it with the PDF text. Anything that can't be confirmed is flagged for review instead of being accepted.

Which processes to automate first

Don't start with the flashiest one. Start with one that meets these four conditions:

It happens a lot

Dozens or hundreds of times a month. A process that happens five times a month will struggle to pay back the build and the upkeep.

It follows clear rules or handles similar documents

If two people on the team would handle it the same way, it can be automated. If every case gets debated in a meeting, not yet.

Mistakes cost money

An overcharge nobody checks, a claim stuck in an inbox or an order that never reaches the ERP. That is where the return shows up first.

The data is already digital

A PDF or a spreadsheet is fine. If the information is on paper or in someone's head, the first step is a different one.

Processes that tend to work well as a first project

Checking supplier or carrier invoices against what was actually ordered or shipped.
Extracting what PDF documents contain, such as land registry extracts or insurance claim reports, into a spreadsheet or database.
Creating records in the ERP or CRM from what comes in by email, web form or WhatsApp, without anyone retyping it.
Keeping customers, products, stock and orders in sync between the ERP and your other tools.
Alerting the right person when something looks off, instead of going through reports every day.

What we wouldn't automate (or not yet)

Saying no is part of the job too. These are the cases where we usually recommend waiting or taking a different route:

Processes nobody can explain
If everyone does it their own way, automating it just locks in the mess. First agree on how it should be done.
Big legal or financial decisions with no human review
AI can prepare the work and flag what looks doubtful. Approving a large payment, signing or answering a formal notice is still a person's call.
Very low volumes
If the process takes two hours a month, maintaining the automation can cost more than carrying on by hand.
Bad source data
If the ERP has duplicate customers or empty fields, automation spreads that error to every system. At Navarro Hermanos the customer type was empty or inconsistent in more than half of the records, so it was decided to manage it in the CRM rather than copy it from the ERP.
Customer service with no way through to a person
A bot that can't hand the conversation to someone on the team ends up causing more frustration than savings.

If you have already had an AI pilot that never made it to production, some of these will sound familiar. We go through them in more detail in 5 mistakes that sink an AI project (in Spanish). And because of the last point, handover to a person is built into AgentDesk.

Systems we work with

What matters about each system is how data can get in and out. These are the ones where we already have automations running:

ERPs and databases

ERPs with an API and ERPs without one, reading their MySQL or PostgreSQL database in read-only mode. Also spreadsheets doing an ERP's job.

Online stores

Shopify and PrestaShop: catalogue, stock, prices, orders and shipping rates calculated at Shopify checkout.

Email, WhatsApp and carriers

Microsoft 365 and IMAP mailboxes, WhatsApp for alerts and conversations, and the MRW and GLS APIs for shipments, pickups and tracking.

Documents

Text-based or scanned PDFs, with OCR and Azure OpenAI models deployed with EU data residency.

For simple workflows we use n8n. When the logic grows, we move it into our own tested code: Crearal's quoting bot started in n8n and is now a TypeScript service.

Real process automation cases

Three automations in production, picked because each one tackles the problem differently: one uses no AI, one depends on it, and the third mixes rules and AI depending on what comes in.

Electrotodo: auditing carrier invoices

Multi-store e-commerce

The problem: Nobody was comparing what the carriers invoiced with what customers had been charged in the store. The first review was done by hand: download the invoice spreadsheet, pull 17,035 orders from Shopify and check them against the agreed rates.
What we built: A module inside their AppCentral. The carrier's invoice is uploaded and every line is matched to its Shopify order by reference. It flags duplicate shipments (including across invoices from different months), cancelled orders that were billed, parcels that actually went with another carrier, surcharges that were never agreed, billed weights above the real ones and charges above the rate card, which is stored as versioned code. Findings arrive in the morning report. It uses no AI because it doesn't need any: the data is already structured.
The detail that made the difference: The rule that stuck: match on the order reference only. During the manual review, matching by postcode and weight produced a false positive that nearly reached the client: a 1.5 kg washing machine seal appeared to have been billed as a 16 kg box, when that order had actually gone with another carrier. The review made something else clear too: the carrier's invoice was essentially correct, and what needed fixing was the store's international shipping prices, which had never been worked out from cost. The manual audit and the first version of the module were done on the same day.

No AIShopify Admin APIInvoice-to-order matchingDaily alerts

Gesttia: automated reading of land registry extracts

Legal services

The problem: For each portfolio, every land registry extract (nota simple) has to be broken down into owners, property details and every charge on it (mortgages, liens, amounts, ranking) and entered into the client's spreadsheet, which has more than 120 columns per row.
What we built: An engine that processes batches of extracts in PDF. If the PDF contains text, there is no OCR: the model receives the transcription plus page images and returns the data in a fixed schema, with each amount tied to the passage it came from. Deterministic code then calculates totals, normalises dates and ID numbers, converts amounts written out in words and queries the Spanish Cadastre. The output is the client's own spreadsheet filled in, one row per charge, plus a report of cells worth reviewing. 1,603 extracts have been processed so far across four portfolios.
The detail that made the difference: In the first portfolio OCR only lost data in 2 of 176 documents; almost all the errors came from interpretation. For example, 34 of those 176 extracts gave amounts in both pesetas and euros, and the model picked the pesetas. In the latest portfolio, 320 extracts received and delivered on the same day, the client's review found 98.39% of cells correct. The most frequent error turned out to be a rule (a contract ID copied onto every charge) and it was fixed in code without calling the model again: 98.77%.

AI with deterministic rulesAzure OpenAI in the EUClient's own spreadsheetDoubtful cases reviewed

Crearal: insurance claims that log themselves

Ceramic replicas · Malaga

The problem: Jobs from insurers arrived by email, each company with its own format, and had to be typed into a shared spreadsheet before a job could be opened.
What we built: An n8n workflow reads the mailbox, identifies the insurer by sender and extracts the claim details: with regular expressions when the email has a fixed format, and with an AI model when the information comes in an attached PDF. It then creates the job in their custom ERP through a webhook. It covers seven insurers and doesn't duplicate a claim that already exists.
The detail that made the difference: According to the May 2026 usage report (last 30 days), 57% of new jobs had come in this way, with nobody typing anything. At first the workflow also kept writing to the old spreadsheet, so nobody had to change how they worked overnight. Requests from individuals and trade customers are handled separately by a bot on the website and WhatsApp, which opens a draft job in the same ERP.

n8nRules and AIWebhook into the ERPNo duplicates

How we build an automation

Small steps work better than a months-long project delivered in one go. This is the order we follow:

Measure the process as it is

How many times it happens a month, how long each one takes and how many errors there are. Without that baseline there is no way to tell later whether it was worth it.

Test on real cases

Before promising anything we work on a real sample of invoices, documents or emails. For Gesttia we set up a test bench of 30 extracts with known correct results, and every change was measured against it.

First version with human review

The system proposes and a person approves. That way edge cases surface without any of them reaching a customer or a supplier.

Automatic, with exceptions

Once results are stable, whatever passes the checks goes through on its own and anything doubtful reaches a person along with the reason.

Ongoing monitoring

Error logging, a morning report and AI usage measured per process. An automation that fails silently is worse than none.

How to tell whether the automation was worth it

With a handful of numbers measured before and after. The full method, with a formula and an example, is in ROI of AI automation: how to measure the real return (in Spanish). These are the ones we watch most:

Hours freed up, minus review time
The hours the team used to spend on the process minus the hours now spent reviewing doubtful cases. Leave out the review and the numbers come out inflated.
AI cost split by process
Each automation draws on its own Azure resource. At Electrotodo the catalogue, the search engine and the website and WhatsApp bots each have their own, and at Gesttia each portfolio does. That way the bill shows what each thing costs.

AI process automation FAQ

What is the difference between RPA and artificial intelligence?

RPA repeats a person's clicks on screen: it works with structured data and stops when the software's interface changes. AI interprets information with no fixed format, such as a PDF or an email, but it can get things wrong and needs checks. In between sits integration through an API or database, which is the most stable option when the data is already structured.

Which business processes should be automated first?

The ones that happen dozens or hundreds of times a month, follow clear rules or handle similar documents, cost money when something goes wrong, and already have their data in digital form. Checking supplier invoices, extracting data from documents and logging what arrives by email are usually good first projects.

Can I automate if my ERP has no API?

Almost always. You can read its database in read-only mode, work with the exports it already produces or, as a last resort, use RPA. The Navarro Hermanos CRM syncs twice a day by reading their ERP's MySQL database directly.

How much does it cost to automate a process with AI?

There are two parts: building the integration, which depends on the systems involved, and AI model usage, which depends on the volume of documents or queries. We measure usage separately for each process so you know what it costs every month. The initial assessment is free.

Does AI make mistakes? How is that controlled?

Yes, and the design has to assume it. We ask the model to back every value with text from the document, check results against rules and other data, and flag anything that can't be confirmed for a person to review. In the land registry extract work for Gesttia, the latest portfolio reviewed by the client reached 98.77% correct cells.

What about data protection when using AI?

We work with Azure OpenAI, where prompts and completions are not used to train the base models. When documents contain personal data, like the land registry extracts for Gesttia, we use deployments with EU data residency. Each project also runs on its own resource, kept apart from the rest.

How long does it take to automate a process?

It depends on how many systems are involved and what needs validating. Electrotodo's carrier invoice audit had its first version live on the same day as the manual audit. Crearal's automatic claim intake went live with the first phase of their ERP, about three weeks after work began.

AI process automation