Using AI does not mean handing the company over to a chatbot. It takes repetitive work off the team, prepares decisions better and frees up time for what needs context, experience and a human relationship.
Start with a better question
Most SMEs have already tried. Someone asked an assistant to write an email, summarise a PDF or suggest a social media post. It is a start, and it rarely changes the operation.
What usually goes missing is method: a concrete task, an expected result and someone who reviews before use. I write this from project and operations management, from the side of someone who measures tasks rather than sells tools.
Instead ofHow can we use AI?
AskWhich repetitive, low-risk task with simple review is eating the team’s time today?
Portugal is below the European average, above all in medium enterprises
In 2025, 11.5% of Portuguese enterprises with 10 or more people used AI. The EU average was 20.0%. The gap varies with size: 7.6 percentage points in small enterprises, 12.2 in medium ones and 5.8 in large ones.
- Portugal
- EU (average)
Eurostat counts several AI technologies declared by the enterprise (for example text, image and speech analysis and generation), so the figure covers far more than conversational assistants. All values are rounded to one decimal place.
See the data as a table
| Size | Portugal | EU (average) | Gap |
|---|---|---|---|
| All enterprises (10 or more) | 11.5% | 20.0% | 8.5 pp |
| Small (10 to 49) | 9.4% | 17.0% | 7.6 pp |
| Medium (50 to 249) | 18.2% | 30.4% | 12.2 pp |
| Large (250 or more) | 49.2% | 55.0% | 5.8 pp |
When Eurostat asks enterprises that considered AI and did not adopt it why not, the most frequent answer is a lack of skills. Legal doubts and data protection follow.
See the data as a table
| Reason | % |
|---|---|
| Lack of relevant skills | 70.9% |
| Uncertainty about legal consequences | 52.5% |
| Data protection and privacy | 48.8% |
| Not useful for the enterprise | 20.7% |
These are obstacles of know-how and of rules, not of access to tools. Those are what this article focuses on.
AI does not replace a process that does not exist
AI summarises, organises, classifies, suggests and writes first drafts. It does not know how your company should work. If everyone records customers their own way and nobody knows which numbers to review each week, you get the same disorganisation, faster.
Treat it as a very fast, always available colleague who does not know the context, misreads vague instructions, writes wrong things with conviction and does not answer for the consequences. The gain shows up when three conditions are in place.
- A minimally defined process. Someone can describe the steps and what counts as done well.
- A repeated task with a clear result. A summary, an email, a list of actions, a first draft.
- Human review before use. A named person validates and owns what leaves the company.
7 tasks AI can prepare in an SME
The right verb is prepare. AI moves the work forward and a person decides. The table shows what AI delivers, who validates and the typical mistake to avoid.
Tasks, validator and warning sign| Task | What AI prepares | Who validates | Warning sign |
|---|---|---|---|
| Sales follow-up and first draft of a proposal | Draft email with the points discussed and the next step. Proposal structure, deliverables and assumptions to confirm. | The salesperson responsible for price, scope and commitments | Sending without checking price, deadline or promise |
| Meeting minutes and actions | Decisions, owners, deadlines and blockers from notes or a transcript | Whoever ran the meeting | An owner assigned that nobody confirmed |
| CRM notes | Context, problem, stage, next action, follow-up date and objections | The salesperson who made the call | A deal value invented by the model |
| Triage of customer requests | Classification (quote, technical, billing, complaint, urgent) and a draft for repetitive requests | Whoever replies to the customer | A sensitive complaint handled with a template reply |
| Repurposing content | One article turned into posts, an FAQ, a checklist or a video script | Whoever speaks for the brand | Generic text and overstated promises |
| Internal procedures | A document with objective, owner, steps, exceptions and control points, from notes or a recording | Whoever performs the task day to day | Steps nobody actually follows |
| Briefings and preparatory analysis | Summary, risks, deadlines and questions for the meeting. Variances and line items to investigate in already validated data. | The manager, plus a lawyer or accountant when the subject calls for it | Deciding or signing on the summary alone |
An example instruction
Context, objective, tone and format. The better the instruction, the fewer the corrections afterwards.
Draft a follow-up email for an industrial company. In the meeting we discussed delays in monthly reporting, scattered Excel files and difficulty tracking margins by project. Propose a 20-minute call next week. Professional, direct tone, without excessive sales language. End with a specific question. Do not invent numbers.
| Field | Example |
|---|---|
| Decision | Review the proposal process |
| Next action | Map the current sales steps |
| Owner | Sales director (confirmed in the meeting) |
| Deadline | Date set and recorded |
| Blocker | Data on open proposals is missing |
AI organises. A person confirms who took ownership and by when.
The greatest value is not in the one-off request. It is in turning the task that worked into a template anyone on the team can use. Keep five things: the information needed before starting, the instruction, the format of the result, who validates and where the final version lives.
5 decisions that stay with a person
AI is useful when it cuts operational work and improves preparation. It becomes a risk when it decides, communicates or acts alone where the consequences are real.
Hiring, promoting, appraising
AI can structure a job description or prepare interview questions. Stays human: who gets in, who moves up, who is appraised badly or ruled out.
Prices, discounts and commitments
AI can prepare scenarios and summarise the customer history. Stays human: approving price, deadline and terms. That needs knowledge of the margin and the relationship.
Delicate communication
AI can organise the facts and propose a first draft. Stays human: dismissals, conflicts, serious complaints, health, significant delays.
Validating critical information
AI can point out what to review first. Stays human: accounts, contracts, legal questions, technical requirements, security, tax obligations.
What leaves the company
AI can work with anonymised data or in an enterprise version. Stays human: deciding what information goes into a tool and where it is processed.
An AI answer can look confident and well written and still contain errors or omissions. The greater the impact of the decision, the less acceptable it is to treat generated text as truth.
A five-line policy is enough to start
Before putting a document into an AI tool, answer three questions. Does it contain sensitive personal, financial, contractual or commercial data? Do I know where the information is processed and whether it can be used to train models? Can I anonymise it or use an enterprise version with suitable controls?
Minimum usage policy| Rule | Why |
|---|---|
| No sensitive personal data and no full customer lists in public tools | The GDPR applies whenever the tool processes personal data. |
| Nothing is published or sent without human review | A person owns what goes out under the company name. |
| No automated decisions about people | Hiring, appraisal and dismissal require judgement and accountability. |
| Numbers and sources confirmed before they enter important documents | The model can invent a plausible figure. |
| In doubt about data, security or the law: ask for help | It costs less than fixing it later. |
What changed in the AI Act in 2026
- 27 July 2026. The digital AI Omnibus, Regulation (EU) 2026/1744, entered into force, published in the Official Journal on 24 July.
- Article 4. The AI literacy obligation, applicable since 2 February 2025, now requires measures supporting staff training rather than a guarantee of a specific level.
- 2 December 2027. Obligations for high-risk systems under Annex III were postponed (previously 2 August 2026) and, for regulated products, to 2 August 2028.
General information, not legal advice. Check your own case with a lawyer.
Measure the time. The feeling misleads.
Two studies show why it pays to measure before believing. In an experiment published in Science in 2023, college-educated professionals did writing tasks from their own jobs (difficult emails, press releases, reports). With ChatGPT, time fell by 40% and rated quality rose by 18%.
In 2025, METR followed 16 experienced developers across 246 real tasks. With AI, they took 19% longer. After the tasks, they estimated they had been 20% faster.
- Measured value
- Perception of those who used it
Different contexts and tools. Do not compare values across studies.
See the data as a table
| Study | Context | Type | Time |
|---|---|---|---|
| Noy and Zhang, Science, 2023 | Professional writing | Measured | −40% |
| METR, 2025 | Experienced developers, after the tasks | Perception | −20% |
| METR, 2025 | Experienced developers | Measured | +19% |
The careful reading is short: the gain depends on the task, the user and the context, and the impression of the person using it is not proof. METR itself warns that its most recent estimates suffer from selection bias, because many developers already refuse to work without AI. That is why step 2 of the plan below asks you to measure the starting point, and step 4 to compare.
The 30-minute test
Pick a task the team repeats and answer honestly.
Tick the yes answers to see the reading.
4 or 5: a good candidate for a pilot. 3: possible. 0 to 2: pick another task.
Five steps, in this order
-
1
Pick a concrete task Avoid goals like "use AI in sales". Prefer "prepare the follow-up after every meeting" or "turn notes into tasks".
-
2
Measure the starting point for two weeks How long it takes and how often it repeats. How many errors or delays it causes and who performs it. What the impact is when it does not get done.
-
3
Write a reusable instruction Context, objective, recipient, tone, format, what it should not include and what needs validation.
-
4
Test with human review Treat AI as a junior assistant: fast and useful, not yet autonomous. Compare it with the previous way on time saved, quality, number of corrections, risk of error and team acceptance.
-
5
Document what works or drop it If it works, record where the template lives, who uses it, who validates and when it should not be used. If it does not work, adjust the instruction, pick another task or stop. The goal is to improve a concrete part of the operation.
Conclusion
Start with a small, repeated task that is easy to review.
Build the method, measure the result and only then move to the next one. Whoever combines technology, process and human judgement in a simple, consistent way will gain more than whoever buys more tools.
Sources
- Eurostat, "20% of EU enterprises use AI technologies" (11 December 2025): EU average in 2025 and 2024.
- Eurostat, "The use of AI technologies in the European Union" (KS-01-26-009): Portugal data by size, Table 1 of the document.
- Eurostat, "Use of artificial intelligence in enterprises" (Statistics Explained): EU by size and reasons for not using AI.
- Noy, S. and Zhang, W. (2023), "Experimental evidence on the productivity effects of generative artificial intelligence", Science. Summary.
- METR (2025), "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity". Method update (February 2026).
- European Commission, AI Act Service Desk, Article 4.
- Entry into force of the AI Omnibus: White & Case and Lewis Silkin.
Data and dates verified on 21 September 2026.