In the first quarter of 2026, 33.6% of Spanish companies with 10 or more employees used artificial intelligence, according to INE, Spain's national statistics institute. Among companies with fewer than 10 employees, the figure was 21.4%. The next question for any owner is whether to build it with the team they have or hire someone who does it every day.
The available data suggests the answer depends on the case, which is why it is worth deciding with numbers. This article gathers what INE, the European Commission, MIT, and Gartner say, and ends with a method for deciding. We are IA4PYMES, a company that develops AI for SMEs, so one of the two answers benefits us. That is why we cite every figure with its source and explain when you do not need outside help.
1. How many Spanish companies already use AI?
INE published its ICT-in-business survey on October 22, 2026:
- 33.6% of companies with 10 or more employees used AI in the first quarter of 2026.
- Among companies with fewer than 10 employees, 21.4%.
- By sector: services 38.2%, industry 29.1%, and construction 22.6%.
INE itself warns that this year's data is not strictly comparable with earlier years, so we do not calculate a growth rate here.
On October 14, 2026, the European Commission published a Flash Eurobarometer based on 12,826 interviews with heads of companies with 1 to 749 employees across the 27 member states (fieldwork in March and April 2026). 42% use AI or algorithmic management, though only 12% extensively and 30% in a limited or experimental way. Another 21% plan to, and 35% have no plans. The three most cited barriers are investment cost (50%), legal uncertainty (49%), and lack of expertise (48%). We analyzed that study in our article on the 2026 Eurobarometer.
With a third of companies with 10 or more employees already using AI, trying it is no longer the hard part. The hard part is getting the result to the bottom line.
2. What does the evidence say about building in-house versus using a provider?
The report The GenAI Divide: State of AI in Business 2025, from MIT's NANDA project, drew on 150 interviews with leaders, a survey of 350 employees, and an analysis of 300 public deployments. According to Fortune's coverage of August 18, 2025:
- About 5% of generative AI pilots achieve rapid revenue acceleration. Most deliver no measurable impact on the profit and loss statement.
- Tools bought from specialized vendors and partnerships succeed about 67% of the time. Internal builds succeed one-third as often.
- The report attributes failures to a learning gap in both tools and organizations and to flawed integration, not to model quality.
- More than half of generative AI budgets went to sales and marketing, while the biggest return appeared in back-office automation.
Three caveats. We cite press coverage, not the original report. "Failure" here means not achieving rapid revenue acceleration, not that nothing happened. And we have not found an equivalent measurement for Spanish SMEs.
On June 25, 2025, Gartner published a forecast: more than 40% of agentic AI projects will be canceled by the end of 2027 because of rising costs, unclear business value, and inadequate risk controls. The firm also estimates that of the thousands of vendors presenting themselves as "agentic", only about 130 are genuine, because many have rebranded chatbots and automation tools. It is a forecast, not a measurement.
Our reading of both sources: projects go wrong more often because of integration with company systems and the lack of a value metric than because of the model chosen.
3. The model and the price change while you decide
On August 16, 2026, DeepSeek replaced its flat rates with peak and off-peak pricing. According to Engadget and HWBusters, V4-Pro output at peak hours went from $0.87 to $3.96 per million tokens, more than four times. We covered it in our article on the DeepSeek V4-Pro price hike.
A system that depends on a single model inherits every change in pricing, policy, or availability. Whoever builds AI alone needs someone to follow those moves and keep an alternative ready. The three basic protections are a layer that lets you switch models without rewriting the system, cost measured per task rather than per token, and a tested second option.
4. When can you do it yourself?
This table reflects our judgment, not a rule:
| Situation | Do it yourself | With support |
|---|---|---|
| Drafting text, summarizing meetings, translating | Yes, with a commercial tool | Not needed |
| Bounded task, one responsible person, no personal data | Yes | Optional, for review |
| Customer service on web or WhatsApp | Possible, with limited scope and handoff to a person | Advisable if it touches orders or customer data |
| Invoices, delivery notes, reconciliation | Only with human validation and testing | Advisable: it connects to the ERP and an error costs money |
| Personal data of customers or employees | Only if you know how to comply with GDPR | Advisable |
| Models on your own server | Only with a technical profile on staff | Advisable |
Doing it yourself safely takes three things: a person with assigned time, a success metric defined before you start, and a plan for when the model or its price changes. If one is missing, the risk goes up. 49% of European SMEs cite legal uncertainty as a barrier. Before entering customer data into any tool, review the data processing terms of your plan.
5. How to decide with numbers
The calculation fits in a spreadsheet:
- Monthly savings = hours per month on the task × hourly cost × share that can be automated.
- Monthly cost = licenses or model usage + maintenance.
- Payback in months = project cost / (monthly savings − monthly cost).
An illustrative example, with our own assumptions and not a real client: a €9,000 project, €22 per hour, 70% automatable, and €150 per month in maintenance.
| Task | Hours per month | Monthly savings | Net after maintenance | Payback |
|---|---|---|---|---|
| Invoice and delivery note entry | 60 | €924 | €774 | about 12 months |
| Weekly report | 15 | €231 | €81 | over 9 years |
The same investment pays off in one case and not in the other. Before programming anything, we calculate that number with each client. If it does not work out, we do not proceed.
6. What we do and what you can take away even if you do not hire us
Here is what we offer:
- A 60-minute consultation with engineers: feasibility, technical architecture, recommended components (APIs, local or cloud models), and a phased roadmap with costs. It costs €299 + VAT.
- Full deduction of that amount if you later develop with us.
- Full refund if no possible benefit shows up in the first 15 minutes.
- Your decision: you can take the roadmap and build it yourself.
- Development: a diagnosis of 2 to 3 weeks and 1 to 4 months of build and launch. The code belongs to the client.
- Data: GDPR, a confidentiality agreement, and no training of public models with your information.
7. Six questions for any provider, including us
- Who owns the code and the data, and what do you keep if the relationship ends?
- Where is the data hosted, and is it used to train models?
- How and when is the return measured, and what happens if it does not materialize?
- Which part is a real product and which part is a rebranded chatbot?
- What happens when the model or its price changes?
- Who validates the answers, and how does a person step in?
For an independent way to check a provider, see our guide to the Kribaia Kriterion seal. If you are also looking for funding, review the AI360 Plan and the AI voucher.
What to do this week
List three repetitive tasks in your company and note the hours each takes per month. Apply the formula from section 5 to each. Drop the ones that do not pay back within a timeframe you find reasonable, and start with the one that saves the most with the least personal data involved. If that task touches invoices, customers, or your ERP, consider doing it with support.
Book your 60-minute technical consultation here and we will calculate the return on your case before writing a line of code.
Frequently asked questions
Is it better to build AI in-house or hire a provider?
It depends on the case. In the MIT study, tools bought from specialized vendors succeeded about 67% of the time and internal builds one-third as often, according to Fortune's coverage. For bounded tasks without personal data, a commercial tool may be enough. For ERP integration or customer data, specialist support reduces risk.
What percentage of Spanish companies use AI?
According to INE, 33.6% of companies with 10 or more employees used AI in the first quarter of 2026, and 21.4% of those with fewer than 10. By sector, services leads with 38.2%, followed by industry (29.1%) and construction (22.6%). INE warns that the data is not strictly comparable with earlier years.
Why do so many AI projects fail in companies?
The sources cited point to similar causes. MIT identifies a learning gap and flawed integration, not model quality. Gartner forecasts that more than 40% of agentic AI projects will be canceled by the end of 2027 because of rising costs, unclear value, and inadequate risk controls.
Can I use ChatGPT in my company without hiring anyone?
For tasks like drafting, summarizing, or translating, yes. Before entering customer or employee data, review the data processing terms of your plan and confirm you comply with GDPR. To connect AI to your ERP or automate processes that involve money, a design with testing and human validation is advisable.
How do I know if an AI provider is reliable?
Ask who owns the code, where the data is hosted, how the return is measured, and what happens if the model changes. Gartner estimates that only about 130 of the thousands of vendors presenting themselves as "agentic" are genuine, so ask for a demonstration with your own data before signing.