
AI for business comes down to four routes: general tools like ChatGPT, the AI already built into the software you pay for, AI workflows on one specific process, and agents. Leadership picks a route after setting the problem, the data, the legal risk, and the metric that decides if it continues.
In 2025, 20.27% of Spanish companies with ten or more employees used at least one AI technology, according to Eurostat, nearly double the year before. For the person running a company, the useful question is which decisions are theirs to make and in what order. This guide is written for that person. It lays out the five questions worth answering before spending anything, compares the four routes and what each one costs, works through an example using prices published by OpenAI and Anthropic, explains who usually decides inside the company and what the EU AI Act requires today, and proposes a 90-day schedule with a built-in decision point. Prices are in euros or dollars as each vendor publishes them, and every one carries its date, because they change without notice.
What is AI for business?
AI for business means putting artificial intelligence models to work in an organization's day-to-day operations: drafting and summarizing, sorting documents and email, answering customers, pulling data out of invoices, or preparing analysis. In practice it arrives through four channels. The first is general tools, like ChatGPT or Copilot, that each person uses at their desk. The second is the AI that ships inside software the company already pays for, such as the CRM, the ERP, or the office suite. The third is AI workflows, which chain fixed steps on a specific process and put a model in one of them. The fourth is agents, which decide the next step on their own. Each channel calls for a different investment, different data, and a different level of oversight, so the decision starts with the problem, not the tool.
How many companies already use AI?
In 2025, 20.27% of Spanish companies with ten or more employees used at least one AI technology, up from 11.31% in 2024, according to Eurostat. The EU average was 19.95%, and among large EU companies it reached 55.03%. Those numbers say two things to an executive. First, four out of five Spanish companies that size still don't use AI, so you're not late. Second, use clusters in marketing or sales (34.70% of European companies that use AI) and organizing administrative processes (31.05%). The breakdown by department, with the specific tasks in each, is in our article on what AI automates in each department. This guide covers the decision that comes before any of that.
What does leadership need to decide before investing in AI?
Leadership needs to answer five questions before spending on AI: what specific problem gets solved and what it costs today, whether the data exists, what legal risk the use carries, which route needs the fewest moving parts, and which metric decides if it continues. The order matters. Start with the tool and there's no baseline to compare against later; skip the stop metric and there's no way to end a pilot that isn't paying off. The first question needs a number: hours per month, errors, response time, or lost sales. Without that number, there's no way to tell later whether AI actually helped.

The second and third questions touch the obstacles companies cite most, according to Eurostat. Among European companies that have considered AI but don't use it, 70.89% point to a lack of relevant expertise, 52.52% to unclear legal consequences, 48.83% to data protection and privacy concerns, 44.37% to problems with the availability or quality of the data, and 41.39% to incompatibility with their existing systems. Cost comes in sixth, at 33.34%, and only 20.68% think AI wouldn't be useful to them. The practical takeaway for an executive is that budget is rarely the first obstacle. The first obstacle is knowing what data the company has, where it lives, and who's accountable for it, and that's something to find out before choosing a vendor.
Tool, AI workflow, or agent: which route makes sense?
The route that makes sense is the simplest one that solves the measured problem: a general tool for each role's text work, AI in software you already pay for if it covers the task, an AI workflow for a process with fixed steps, and an agent only if the path changes case by case. Every step up adds integration, maintenance, and risk. A general tool can be rolled out in a few weeks with usage rules, as we cover in ChatGPT for business. An AI workflow already means connecting systems and deciding which step a person reviews. An agent also needs permissions, limits, and a log of what it does, which is why we wrote about when you don't need an AI agent.

The route people forget is the second one. A company that already pays for a CRM, an ERP, or an office suite may have AI features included in its plan, so before building anything it's worth checking what the current contract includes and what it would cost to turn on. Sometimes the best answer isn't AI at all. In an assessment for an accredited environmental inspection firm that wanted to automate its technical reports with AI, we compared three real reports and found that nearly half of the body text was identical from one report to the next. That's a job for templates, not a model. AI kept the part of the writing that changes with each case. Choosing the simplest route lowers the cost of the project and shrinks what has to be monitored afterward.
How much does AI cost a company?
The cost of AI for a company has five parts: per-user licenses, model usage through an API, integration work with existing systems, the team's time, and training. The first two are public, and for most uses they're the cheap part. ChatGPT Business costs €21 per standard seat per month billed annually and €26 billed monthly, according to OpenAI's pricing page for Spain, checked on November 17, 2026. On Anthropic's API, Claude Sonnet 5.5 costs $2 per million input tokens and $10 per million output tokens, and Claude Haiku 5.5 costs $0.10 and $0.50 for prompts up to 100,000 tokens, according to its pricing page on November 17, 2026. What usually costs real money is the work of designing the workflow, connecting it, and keeping it running.
A worked example makes the point. A 40-person company gives ChatGPT Business to its 12 office roles: €252 a month on annual billing, €3,024 a year. It also builds an AI workflow that reads the 2,000 emails and documents it receives each month, sorts them, and leaves a summary for whoever handles each one. Assuming 1,500 input tokens and 300 output tokens per document (a couple of pages of text and a short summary), the month adds up to 3 million input tokens and 600,000 output tokens. On Sonnet 5.5 that's about $12 a month; on Haiku 5.5, about 60 cents. The model barely registers on the bill. What drives it is the integration and review work, which is quoted per project. We'll cover Aurora's own pricing in a separate article on AI consulting.
Who makes the AI decision inside a company?
At the midsize companies that came to us this year, the AI decision sat with a committee of HR, IT, and finance, not with marketing or a single person. At a pet food group, the marketing manager who asked us for automation examples passed them to HR, which was running the change together with IT. The head of a healthcare services company did the same with finance and HR. Each one looks at something different: HR at training and what changes in each role, IT at security, access, and integration, and finance at cost and return. A project that reaches that committee with the problem measured, the data located, and the stop metric already written arrives with the answers each of them will ask for.
The CEO has a different job: deciding which problem gets tackled first and who answers for the result. When nobody owns that, every department tries tools on its own, and the company piles up personal accounts, scattered data, and no results anyone can compare. At two of the companies that came to us this year, a retail chain and an environmental consultancy, AI use had started before anyone decided on it, with no written rules about what information could go in. Fixing that doesn't take a data team. It takes one person with the authority to decide, a written usage policy, and a first process chosen for its measured cost. That first process is what later justifies the next one, or doesn't.
What does the law require of a company that uses AI?
A company that uses AI in the EU has to comply with GDPR for any personal data that enters its systems and with the EU AI Act, whose first obligations have applied since February 2025. Article 4, as rewritten by Regulation (EU) 2026/1744 (the Digital Omnibus), in force since July 27, 2026, requires anyone deploying AI to take measures to support AI literacy among their staff, without requiring any specific level. Annex III of the Act classifies as high-risk, among others, systems that screen applications, evaluate candidates, or are used to decide promotions and terminations, and that category calls for additional safeguards. For everyday office uses, the practical steps are to sign a data processing agreement with the vendor, write down what data can go in, and train the team. We cover what data to keep out of ChatGPT in a separate article.
Where do you start with AI in a company?
You start with a two-week assessment that measures the candidate processes, follow it with a pilot of about six weeks on a single process, and close in week twelve with a decision: expand, adjust, or stop, based on the metric set at the start. That schedule is our recommendation for a company of 20 to 250 people, and it can be compressed if the problem is already measured. What matters is the order. The assessment looks at processes, the systems already in place, and the data, and ranks use cases by impact and effort. The pilot runs on real operations, with a small group of users and someone reviewing what the AI produces. The final decision is made on the pilot's numbers.

The assessment exists so you don't build what you don't need. When a pet food group asked us for the most common automations by department, we sent a list covering marketing, sales, HR, operations, quality, internal communications, and logistics, and presented it as a working hypothesis, not an assessment: before ranking anything, we had to confirm with HR and IT which systems they already used (CRM, ERP, document management). That check changes the order of the use cases, because a process that looks simple can depend on data spread across three different programs. If you'd like to work through it with us, our AI assessment follows that method and ends with ranked use cases and a proposed pilot, or with a recommendation to wait if the company isn't ready yet.
Why do AI projects fail at companies?
AI projects tend to fail because of rising costs, unclear business value, and inadequate risk controls. Those are the three causes Gartner gave in June 2025 when it predicted that more than 40% of agentic AI projects will be canceled by the end of 2027. In our view they apply to any AI project, and they matter more than which model you pick. All three can be headed off with the five questions from the start of this guide. Cost, by choosing the simplest route that solves the problem. Value, by measuring the problem before starting and setting the stop metric. Risk, by deciding what data goes in, who reviews what comes out, and which uses fall under the Act's high-risk category. A project that can't answer those questions before the pilot is unlikely to answer them later, once the budget is spent.
There's a fourth reason, harder to see: waiting for the technology to settle down. The CEO of an environmental consultancy wrote to us last July that something new came out every day that beat the last thing, and that they wanted to see how they adapted before taking another step. It's a reasonable objection, and the answer is to separate what changes from what doesn't. Models change every few months. A company's processes, its data, and its metrics don't. A well-measured assessment and pilot still hold up when a new model comes out: you swap the model and measure again. What you can't get back is a year spent without knowing which process deserved the investment.
Frequently asked questions
How much does it cost to implement AI in a company?
It depends on the route. A general tool costs its license: ChatGPT Business is €21 per standard seat per month on annual billing in Spain (OpenAI, 2026-11-17). Model usage through an API is usually small: sorting and summarizing 2,000 documents a month with Claude Sonnet 5.5 comes to about $12. Most of the cost is integration, review, and training, which are quoted per project.
What AI should a small company use?
The simplest one that solves a measured problem. For each role's text work, a general tool on a business plan with usage rules. Before buying anything, check which AI features the software you already pay for includes. AI workflows and agents are for repeated processes with a clear cost.
Is there public funding for AI in Spain?
Spain's Kit Consulting program offered digital advisory vouchers of €12,000, €18,000, and €24,000 for companies with 10 to 249 employees, but the Acelera pyme website lists its application round as closed (checked 2026-11-17). Confirm with Red.es whether a new round opens before counting on it.
Do you need a data team to use AI?
Not to get started. You need one person with the authority to decide, a written usage policy, and a clear picture of where the data for the target process lives. An in-house technical team only becomes necessary when the company builds and maintains several workflows or agents at once.
How long does AI take to show results?
With a measured problem, a single-process pilot produces comparable numbers in about six weeks, and the decision to expand, adjust, or stop can be made around week twelve. That's our recommended schedule, not a guaranteed timeline: it depends on the data and on the systems that need connecting.
Will AI replace the team?
In office work, AI takes repetitive hours off people's plates, like sorting, summarizing, or drafting, and leaves review and decisions to them. Our view is that AI works for the company and a person signs off on what goes out, especially when numbers, customers, or candidates are involved.
Eurostat, "Use of artificial intelligence in enterprises", Statistics Explained, data extracted December 2025, and its 2025 data file (accessed 2026-11-03).
OpenAI, ChatGPT plans and pricing (Spain), "Business and Enterprise," accessed 2026-11-10 and again 2026-11-17.
Anthropic, Pricing, API prices per million tokens, accessed 2026-11-17.
Regulation (EU) 2024/1689 on artificial intelligence, Article 4 and Annex III, and Regulation (EU) 2026/1744, Digital Omnibus on AI, accessed 2026-11-10.
Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027," press release, 2025-06-25, accessed through w.media's coverage (Gartner's site blocks automated reading).
Acelera pyme (Red.es), Kit Consulting, accessed 2026-11-17.
Aurora's own records, cited without names: emails from a pet food group (2026-07-06), a healthcare services company (2026-07-08), and an environmental consultancy (2026-07-08); a by-department use-case proposal (2026-07-09); a technical-report assessment for an environmental inspection firm (published on our AI services page).
Where would AI start in your company?
We start with an assessment of your processes, systems, and data, and give you back use cases ranked by impact and effort, with a proposed pilot and its metric. If it isn't worth it yet, we'll tell you that too. See our AI services.