AI process automation: what each department can actually automate

AI process automation breaks down by department: in sales and marketing, answering leads and drafting copy; in finance and purchasing, reading invoices and orders; in HR, handling paperwork; in customer service, resolving repeat questions. The AI reads and writes, and rules do the data entry.

In 2025, 20.27% of Spanish companies with ten or more employees used at least one AI technology, nearly double the year before (11.31%), according to Eurostat. Across the European Union, companies that use AI put it to work mostly in marketing or sales (34.70%) and in organizing administrative processes (31.05%), and barely at all in logistics (6.08%). This article goes from that big picture down to the actual tasks: what gets automated in each department, how much effort it takes, what it's worth, and whether it calls for plain rules, an AI workflow, or an agent. It closes with a worked example of where to start, drawn from a questionnaire the leadership of a retail chain filled out for us, and with the reasons many European companies that have considered AI still haven't made the move.

Which departments use AI the most?

Marketing and sales is where the most European companies use AI: 34.70% of the companies that have it use it there, according to Eurostat's 2025 data. Next come organizing administrative or management processes (31.05%), accounting and finance (23.11%), production (20.76%), IT security (19.77%), and R&D (19.05%), with logistics far behind (6.08%). Company size reshuffles the order. At large companies, administration moves to the top (43.38%) and IT security reaches 47.51%, compared with 14.51% at small ones. Industry matters too. In retail, 48.18% of companies using AI apply it to marketing or sales, and in transportation and warehousing, logistics rises to 19.07%. The averages hide a lot, so it's worth checking your own industry's row before deciding where to start.

Our read on those numbers is a practical one. Marketing and sales lead because many tools those teams already use, from email platforms to the CRM, come with AI built in, so getting started doesn't require connecting anything. Logistics trails because the thing that pays off most there, demand forecasting, needs a clean history of data that few small companies have. Administration and finance sit in between, and most of the work there is reading documents and keying their data into a system. That's exactly the kind of task where a language model adds the most, because it can read different formats without anyone writing a rule for each supplier. That's why, at the companies that come to us, administration and purchasing are usually the first candidates, even though they're rarely the department people talk about when they talk about AI.

Which processes can you automate with AI?

You can automate the processes that repeat many times a month, work with text or documents, and whose data already lives in some company system: invoices, purchase orders, customer emails, or reports. To decide whether a specific process qualifies, we ask four questions. How often does it happen each month? A task done three times a month rarely justifies the work of building it. What form does the information arrive in? If it's text or a document that varies, you need a language model; if it's already structured data, rules will do. What does a mistake cost? If an error has legal or payroll consequences, the decision stays with rules and with a person. And which systems does it have to talk to, whether that's the ERP, the CRM, or the document repository? An automation that can't connect to them never gets past the demo stage.

The answers to those four questions also decide the tool. If fixed rules are enough, a classic automation does the job, with no AI involved. If a step requires reading or writing variable text inside a fixed path, you build an AI workflow: the model handles that step and the rest runs on rules. And if the steps change from case to case and cut across several systems, an agent makes sense, because it chooses its own path within set limits. We cover the difference between the three options in detail in what an AI agent is and what it can do for your business. The table applies that test to the tasks we see most often in each department, with our estimate of effort and impact for a company with 20 to 200 employees.

What can you automate with AI in sales and marketing?

In sales, the biggest payoff is replying right away to every lead that comes in through the website or by email and sorting it by fit, so reps see the promising ones first. That's an AI workflow: the model reads the message and classifies it, while assigning the lead, alerting the rep, and logging it in the CRM run on rules. The second task with a good effort-to-impact ratio is summarizing every call or meeting and recording the next step in the CRM, something many sales teams don't do consistently. Both take little effort if the CRM supports integrations, which the widely used ones do. What shouldn't be automated is the call on pricing or discounts, which stays with the rep or their manager, along with any commitment the company has to stand behind later.

In marketing, AI already drafts copy, product descriptions, ad variations, and performance summaries, always reviewed by someone who knows the brand before anything goes out. What changed this year is something else. Search engines now answer with AI and cite some websites and not others, so marketing teams have started working to get mentioned in those answers too, which is what we call generative engine optimization, or GEO. Part of that work automates well, and it's the measurement: asking several engines the same questions every month and logging who gets cited, as we explain in how to check whether ChatGPT cites your site. Marketing case studies with numbers from our own accounts are getting their own article in this series, which is in the works.

What can you automate with AI in finance and administration?

In finance and administration, the most automatable task is reading supplier invoices and purchase orders, which arrive in every format imaginable, and entering the vendor, amounts, dates, and line items into the accounting system or ERP. Reading is work for a language model, which can make sense of one supplier's PDF and a photo of another's delivery note without a rule for each. Entering the data, on the other hand, runs on rules, with a person reviewing anything the model flags as uncertain. Bank reconciliation is different. Matching transactions to invoices is mostly rule-based work, and AI only helps with the mismatches, when someone has to read a garbled description. Cash flow forecasting needs historical data and a predictive model, and it's a project of a different size altogether.

This is the department where the gap between companies that already keep their data in a system and those that don't shows up most. In a questionnaire from July 2026, the leadership of a retail chain described its administration as having "lots of outdated processes" and listed a "homegrown ERP" and an "accounting program" as its only digital tools, rating its own digital maturity at 1 out of 5 (their answers were in Spanish). Starting from there, automating invoice reading is doable, because the data has somewhere to go. What determines the effort is whether that homegrown ERP accepts data from outside, either through imports or through an integration. That's the first technical question to answer, before any AI model enters the picture.

What can you automate with AI in HR?

In HR, anything that's repetitive paperwork automates well: onboarding, offboarding, and contract changes sent to the payroll provider, training records, and answers to employee questions about the labor agreement, vacation, or pay. The same leadership team summed up the problem in their HR department as "lots of paperwork, lots of hours with no value." Onboarding and offboarding don't need AI. The steps are fixed, and personal data calls for rules and an audit trail, which is why that company already had a workflow connected to its payroll provider. Where a language model does fit is in answering employee questions from internal documents, the labor agreement and company policies, with each answer citing the document it came from and a person in HR for anything that isn't written down.

Candidate screening is the case that calls for the most care. The EU's Artificial Intelligence Act, in Annex III, classifies as high-risk any AI system used to recruit or select people: placing targeted job ads, analyzing and filtering applications, and evaluating candidates. The same goes for systems that make decisions about promotions or terminations, or that monitor and evaluate employee performance. A high-risk system comes with documentation, human oversight, and data quality obligations that a small or midsize company can rarely meet on its own. Our recommendation is to let AI help prepare the work, sorting résumés or summarizing interviews, and to keep every decision about a person in human hands. That keeps the efficiency gains and leaves the legal exposure where the company can manage it.

What can you automate with AI in purchasing and logistics?

In purchasing and logistics, the first thing to automate is order paperwork: reading delivery notes, confirmations, and orders from customers or suppliers that arrive by email in different formats, and loading them into the system. It's the same AI workflow as for invoices, with the same split between what the model reads and what the rules enter. Demand forecasting is the other big task in this area, and it's the one fewest companies have figured out. In Europe, only 6.08% of companies using AI apply it to logistics, although in transportation and warehousing that number rises to 19.07%. Forecasting demand is a job for a predictive model trained on the company's own sales history, and that model is only as good as the history behind it.

That history is usually what's missing. The retail chain's leadership described their logistics problem this way: "no forecasting, no metrics." It's an accurate diagnosis, and it explains why logistics comes late to AI. Before you can forecast, you have to measure: sales by store, by product, and by week, kept in one reliable place long enough for seasonal patterns to show up. At a company like that, the first logistics project is a dashboard that pulls those numbers together every week. Once the data exists, forecasting becomes a contained project. Without the data, any forecast a vendor promises is a guess with a nice presentation, and the first quarter of real sales will show it.

What can you automate with AI in customer service and management?

In customer service, repeat questions get automated, and this is the clearest case for an AI agent: a question that requires checking the order, the shipment, and the invoice before answering, where the steps change depending on what turns up. The agent checks all three systems, answers or sets up the fix, and hands off to a person anything that crosses a limit, like a refund above a set amount or a customer who has already complained twice. Questions that always have the same answer, such as hours, delivery times, or return policies, don't need an agent. An assistant that answers from the company's own documents does that job at lower cost and lower risk, and it's easier to keep accurate when a policy changes.

In management, the automatable task is the regular summary: pulling sales, cash, and operations numbers into a dashboard every week and having a model write up what changed and why. The data side runs on rules and integrations, and the AI only writes about numbers that have already been calculated. In operations, the most instructive case we've seen is an environmental inspection firm that asked to automate its technical reports with AI. When we compared them, nearly half the text was identical from one report to the next, and that part was solved with templates, no AI needed. The language model only kept the writing based on calculated data, and the part covered by the firm's accreditation still runs on code.

Where should you start automating processes with AI?

Start with a single process: the one that combines the most manual hours, data that's already in a system, and mistakes that are cheap to fix. The retail chain's questionnaire works as an example. Asked about their three biggest operational bottlenecks, the leadership answered "HR, administration, logistics." Asked where the most time is lost on manual tasks, "administration, purchasing, HR." And asked which area would benefit most in the short term, "purchasing, HR, logistics." HR shows up in all three answers, but its heaviest task, onboarding and offboarding with the payroll provider, was already handled by a workflow they'd built themselves. Purchasing and administration show up in two answers each, and they share the same job: reading supplier invoices and purchase orders.

That would be our first pilot: an AI workflow that reads supplier documents and loads them into the homegrown ERP, with a metric set before starting, such as administration hours per month or data entry errors. Logistics, the third candidate, comes later, because its stated problem is the lack of metrics, and that gets fixed with data first. That's how our AI consulting works: an assessment, a contained pilot, and scaling only what meets its metric. The assessment starts with an inventory of systems. When a pet food group asked us for the list of standard automations by department, we gave it to them with a written caveat: a list like that is a working hypothesis, and before setting priorities you need to see which CRM, ERP, and document system the company uses.

What holds companies back from automating with AI?

What holds companies back most is a lack of in-house know-how. Among European companies that have considered AI but don't use it, 70.89% point to a lack of relevant expertise, according to Eurostat (2025). Next come unclear legal consequences (52.52%), concerns about data protection and privacy (48.83%), problems with the availability or quality of the data they'd need (44.37%), and incompatibility with the equipment, software, or systems they already have (41.39%). Cost comes in sixth, at 33.34%, and only 20.68% think AI wouldn't be useful to them. In other words, most of the obstacles are about knowledge and data, and those are the ones an assessment is built to work through.

Those numbers match what we hear in conversations. Asked what worried them about rolling out AI, the retail chain's leadership answered: "doing it without a plan and not actually being able to create value or increase productivity." What worried them was order, which is what the systems inventory and a pilot with a metric address. Legal clarity and data protection get handled in the same assessment: which personal data each process touches, where it gets processed, and which tasks fall into the high-risk categories of the EU regulation. And no single department makes the call. At the companies that have come to us, decisions about AI usually sit with a committee made up of HR, IT, and finance.

FAQ

Which process would you automate first?

We start with an inventory of your systems and the process with the most manual hours, and we tell you which one needs rules, which one an AI workflow, and which one is better left alone for now. See our AI consulting.