In 2025, almost 20% of companies in the European Union with at least 10 employees used AI technologies in their work. In Romania the figure was 5.2%, the lowest in the Union, according to Eurostat.
The gap is not about access to tools. The same tools are available to anyone, at the same price. It comes from the fact that in most companies AI arrived as a chat window where people ask questions, not as a change in how the work gets done. And the real value is not in the chat. It is in the processes that repeat hundreds of times a month, where people are effectively doing the job of an interface: reading something in one place and typing it into another.
That is what this article is about. Not what AI can do in general, but how you find the processes in your company that are worth automating, and how you automate them without creating a new problem.
What has actually changed
Process automation is not new. Approval workflows, integrations between systems, rules in the ERP and software robots (RPA) that click buttons instead of a person have been around for years. They all share the same limit: they only work when the input is structured and predictable. A field, a value, a rule.
The problem is that much of the work in a company does not look like that. It looks like an email from a customer asking for "the same as last time, but delivered faster". Like a PDF from a supplier, in a format different from the other 40 suppliers. Like a 30-page contract in which five clauses need checking. For all of these, classic automation needed a person to read, understand and turn the information into structured data.
That is what has changed. Language models read unstructured text and understand it well enough to extract what matters, classify it and suggest a next step. The simple judgement that kept the process with a person can move to the system. The decision stays with the person.
Not all automation is AI
One point we make at the start of every conversation: a good share of what gets requested as "AI automation" does not need AI at all.
The clearest example is e-Factura, Romania's national e-invoicing system. Invoices received through it arrive as XML, which is already structured data. Importing them automatically into your inventory or accounting system is a classic integration, with no language model involved. Putting AI there means paying for a layer that can make mistakes, in a place where a simple rule never does.
The rule we apply is simple: if the data is structured and the rule can be written down, you use classic automation. AI comes in only where the input is text, a document or natural language, and where a person would otherwise have to read and interpret it. The best solutions combine the two: AI turns the document into data, and from there everything runs on rules, verifiable and predictable.
Which processes are worth automating
The most useful test we know is the copy-paste test. Look at what people do in a week and find the moments when someone reads something in one window and types something into another. Every one of those moments is a candidate.
Then four criteria decide whether it is worth it:
- Volume. A process that happens three times a month rarely justifies automation. One that happens three hundred times almost always does.
- Type of input. The less structured the input (emails, PDFs, scanned documents, free-form messages), the more AI adds over classic automation.
- Verifiability. The result must be easy to check: the invoice total matches the order, the email was classified correctly or it wasn't. If a mistake cannot be detected, the process is not a good candidate for full automation.
- The cost of a mistake. An email routed to the wrong department costs a few minutes. A wrong payment costs money. The higher the cost, the closer the person stays to the decision.
On these criteria, the processes that come up most often are the same in almost every company:
Incoming documents. Supplier invoices outside e-Factura, delivery notes, order confirmations, statements, transport documents. AI extracts the data, checks it against the order or contract and flags only the differences. People stop typing in data. They check the exceptions.
Shared inboxes. office@, orders@, support@. Every message is read, classified and sent to the right person, with a summary and, where possible, a draft reply built from the customer's history and the company's procedures. Response times drop because nobody triages by hand any more.
Quoting. Requests for quotes arrive by email, in all sorts of shapes. AI extracts what is being asked for, finds the matching products or services, applies the company's pricing rules and prepares the quote for approval. The salesperson reviews and sends it instead of building it from scratch.
Generated documents. Quotes, contracts and annexes, client reports, handover reports, job descriptions, project documentation. In many companies each one starts from "the last document of the same kind", copied and edited by hand, with mixed fonts, old clauses forgotten in the text and another client's name left in the footer. Here automation splits the work in two: AI writes the content, from the client's data and the company's knowledge, while an approved template controls the form, not the model. Visual identity, structure, mandatory clauses and numbering are enforced by the template, and the document is checked automatically against the company's rules before it reaches a person. The person reads and signs a document that is right the first time, instead of fixing the formatting.
Document checks. A set of contracts checked against a list of mandatory clauses. A funding application checked against the applicant's guide. A file checked against a client's requirements. AI does not decide, but it does the first pass and shows exactly where the person needs to look.
Recurring reports. The Monday-morning report, built from three sources, with the same comments every time. AI prepares it, including the written part, and the manager reads it instead of composing it.
The foundation: company memory
All of the processes above share one need: context. To prepare a correct quote, the system has to know the prices and the discount rules. To generate a contract, it has to know which clauses the company uses now and which have been replaced. Without a company knowledge base, organised and kept current, every automation builds its own context, and AI confidently repeats outdated information. We wrote separately about how to build a company memory, a knowledge base AI can actually use.
What is not worth automating (yet)
What you leave out matters just as much.
A process that doesn't really exist. If three people do the same thing three different ways, automation will reproduce the chaos, only faster. Clarify the process first, then automate it. We wrote more about this in the article on when you need custom software: software makes a process faster, it does not make it exist.
Decisions about people. Screening candidates, evaluating employees, granting credit or a commercial condition based on a person's profile. GDPR restricts decisions based solely on automated processing that significantly affect a person, and the EU AI Act treats some of these uses as high-risk. Here AI can prepare the information, but the decision stays explicitly with a person.
Processes where mistakes go unnoticed. If nobody can say after the fact whether the result was right, you have no way of knowing whether the automation works. Without verification you don't have automation, you have hope.
Low-volume, high-variation processes. A complex document that comes up once a quarter, different every time. Here a person with a good AI assistant is more efficient than any automated flow.
How much you let the system do on its own
In the article on bringing AI into your company without losing control of it we described three levels of access: chat, delegated tasks and processes that run on their own. In automation the same logic applies to each process separately:
- The system proposes, the person acts. AI extracts, classifies and prepares drafts. The person approves every step. Every automation starts here.
- The system acts, the person approves. The invoice is entered, the email is drafted, the quote is built. The person sees the result and clicks "send" or "correct".
- The system acts on its own, the person sees the exceptions. Standard cases go through automatically. Only those that don't fit the rules reach a person.
A process moves from one level to the next on data, not on trust. If after a few hundred cases at level 2 corrections have become rare and predictable, the process can move to level 3 for standard cases. If not, it stays where it is.
What a project that delivers looks like
Automation projects rarely fail because of the technology. They fail because nobody measured the starting point, so at the end nobody can say whether it was worth it. The order that works:
Measure first. How many documents, how many emails, how much time each, how many mistakes and what a mistake costs. Without these numbers, any result is an impression.
Pick a single process. The one with high volume, unstructured input and an easily checked result. Not the most complicated one, nor the most visible, but the one where success can be shown clearly.
Run a pilot on real data. On the company's real documents and emails, not hand-picked examples. With the people who run the process today, because they know the exceptions no document describes.
Go to production with written rules. Which data goes into the system, where it is processed, who checks it, what happens when the system is not sure. The same data classification rules we described for AI adoption apply here, all the more so because an automated process runs without anyone looking at every step.
Measure monthly. Hours saved, errors avoided, response time, running cost. The numbers decide which process comes next and whether the current one can move up a level of autonomy.
One more thing we say from the start: the AI Act already requires companies that use AI systems to take measures to train the people who work with them. Training the team is not a bonus of the project. It is part of it.
Which tools
No single tool fits every process. The choice depends on where the data lives and how close to the company's systems the automation has to run.
A team AI assistant, for tasks a person delegates: document checks, reports, consolidations, documents generated from the company's templates. We recommend Claude on the Team plan: shared projects keep the company's knowledge in one place, and Cowork carries out tasks on files and applications instead of just answering questions.
Automation inside the platform you already have, for companies that work entirely in Microsoft 365: Power Automate for flows and Copilot Studio for agents built on the company's procedures, with the permissions groundwork done first.
A custom-built integration, for processes that have to run inside the company's systems, at volume, without a person starting each step: a service that reads the inbox, extracts the data with a language model, validates it against the company's rules and writes it straight into the inventory system or CRM. Here AI is a component of a system, not a separate tool.
In practice, the best results come from the combination: the assistant for what people do, the integration for what repeats at volume.
What we do
The AI consulting we do starts from processes, not tools. A workshop with the people who do the work, an analysis of processes by volume, input type, data sensitivity and estimated gain, then building the company memory from the sources that matter, a pilot on the chosen process, the move to production with written rules, and monthly reporting on hours saved, errors avoided and running cost. For documents, that means the company's templates brought to a single standard, with automatic checking rules, from which quotes, contracts and reports are generated.
Sometimes the conclusion is that a process needs AI. Sometimes it is a classic integration or a setting in something you already have. All of these are good answers, as long as the numbers chose them.
If your company has people who spend a good part of the day copying information from one place to another, that is the first process to automate. The details of how we work are on the AI consulting and process automation page, and a conversation starts with one simple question: which process costs you the most time today?
Data source: Eurostat, use of artificial intelligence in enterprises, 2025 (companies with at least 10 employees).
