AI won't replace you. But someone with AI will.

I finished the Master's in Artificial Intelligence in Communication and Media at UCM four months ago. It was my honest attempt not to be one more consultant saying "let's do something with AI" while confusing Claude with a search engine. After that Master's, and after rolling out AI in six projects with real clients since then, I have an uncomfortable conclusion for the market: most of what's being sold as "AI for SMEs" today adds nothing.

Not because AI doesn't work. It does. But the "have a GPT on your website" industry has created an expectation that doesn't match what actually delivers ROI. The result is SMEs paying €8,000 for a "smart assistant" and switching it off two months later because nobody uses it.

Here's what I've seen actually work — and what I haven't — over the last four months with real clients.

Three cases where AI did deliver clear ROI.

Case 1 · Out-of-hours lead replies — private clinic.

Context: a sports physiotherapy clinic in Sant Cugat with six practitioners. 60% of their leads were coming in between 7pm and 11pm, when patients were heading home from work and searching on their phones. The clinic's reply: the next morning, if whoever managed appointments happened to make it in.

Result: 38% of those leads no longer replied to the return WhatsApp. Too many had already gone looking elsewhere.

What we built: an AI agent (Claude with an n8n wrapper) that replies automatically on WhatsApp in under 30 seconds. It doesn't diagnose anything. It doesn't practise medicine. It does four specific things:

• Greets the person by name (if they filled out the website form).
• Confirms the message arrived and that a human will respond first thing tomorrow.
• If the person asks about opening hours, indicative pricing or what a treatment includes, it answers from a knowledge base curated by the clinic.
• If they ask anything medical or personal, it says literally "someone on our team needs to answer that, we'll get back to you first thing tomorrow".

Result 90 days later: the loss rate for afternoon/evening leads dropped from 38% to 12%. Setup cost: €1,200 + €35/month in tokens. Estimated return in the first quarter: 15-20 additional patients booked. Investment paid back in six weeks.

Case 2 · Generating product descriptions — home textiles e-commerce.

Small distributor in the Penedès with 1,800 product references. Each new reference needed an 80-150 word description for the website and three bullets for the Google Shopping feed. The person doing it took 12-18 minutes per product, writing at the end of the day. Perpetual backlog of 200 products not yet live.

What we built: an n8n workflow that takes the product's technical specs (composition, size, colour, use) and generates three outputs in Catalan and Spanish: SEO-friendly web description, Shopping bullets, meta description. The person reviews, corrects nuances of tone, signs off in 90 seconds.

Result: time per product 12 minutes → 2 minutes. Backlog cleared in six weeks. New product live on the website half a day after physical arrival (previously: 2-3 weeks). Cost: €600 setup + €12/month in tokens.

Important: no product goes live with the AI description as-is. The person validates everything. The difference is that they now validate instead of writing from scratch. The creative work has been preserved; the mechanical work has disappeared.

Case 3 · Automatic classification of sales inbox.

Industrial distributor with 18 years in the sector. The senior sales rep was getting 80-120 emails a day, of which only 15% were real opportunities. The rest: complaints that admin could resolve, generic information requests a junior could answer, sales spam.

What we built: an AI classifier that reads each incoming email and places it in one of five categories. Labels in Gmail, Slack alerts for "hot" ones, auto-replies with standard information for "generic info" requests. At the end of the day, the rep's inbox only contains the 12-15 that genuinely matter.

Result: 1.5 hours a day recovered for the rep. 22% increase in closed opportunities the following quarter (not because more were coming in, but because the ones that did arrive got answered faster). Total cost: €800 setup + €8/month in tokens.

Three cases where AI didn't deliver ROI (yet).

Case 1 · "Smart" chatbot on the main website.

A healthy-food brand hired me to put a chatbot on the homepage. The idea was that it would answer product questions, help pick out diets, and so on. Five weeks of work, integration with the catalogue, curated knowledge base.

Result 60 days later: 800 sessions started, 12 handoffs to human, 3 sales attributable to the bot. Practically nothing.

Why: by the time the user lands on the site they've already decided they're hungry or want a new diet. What they need is to see the menu, not chat with an assistant. The chatbot added an unnecessary step. We switched it off after 90 days. We reinvested the same budget in improving the catalogue filter and the product page. ROI showed up in the second month.

Case 2 · "Internal copilot" for a 12-person team.

A communications agency asked me to build an internal chat with access to all their files (Drive, Notion, emails). The idea: anyone on their team could ask "what do we know about client X" or "what was the last campaign brief".

Result 90 days later: 4 out of 12 people used it more than once a week. The rest said Notion's own search was fine. Setup cost was €4,200. Token and maintenance cost: €180/month. Measurable ROI: zero.

Why: never underestimate the friction of "changing a habit". If your people already have a system (Notion, Google, Slack, whatever) and that system works, adding a "copilot on top" doesn't add value unless the current search is bad. In this case it wasn't.

Case 3 · Founder's "AI avatar" for onboarding videos.

A startup wanted new clients to receive a personalised welcome video generated by AI, with a clone of the founder's voice mentioning the client by name. Tech: HeyGen + ElevenLabs.

Technically: it worked. Conversion from "watch the video" to "complete onboarding" was identical to a plain text email. The video took longer to load, the discomfort ("is this real?") held people back more than it helped. Cost: €2,000 + monthly subscriptions. We went back to the text email.

Why: there's a difference between "we can do it with AI" and "it should be done with AI". The first is technology; the second is value. If it doesn't improve a metric you actually measure, you don't do it.

The pattern I see.

After six projects and four months of experimenting, I'm left with one practical rule for SMEs that want to apply AI:

All three successful cases above share this:

• There was an existing job that someone was doing (badly or late).
• It was clearly bounded: replying to leads, generating descriptions, classifying emails.
• The human cost was visible: hours, lost leads, backlog.
• The person doing it felt relieved, not threatened.

All three unsuccessful cases shared this:

• There was no existing job: it was a new "initiative".
• The problem was fuzzy: "help the visitor", "improve onboarding".
• The cost of not doing it wasn't clear.
• People saw it as a new layer on top of what they were already doing.

Why "someone with AI" will replace you.

The title of this article isn't clickbait. It's literally what I see in the market. Your competitors who are investing in AI the intelligent way — removing boring jobs, not creating showy services — are recovering hours every week, cutting fixed costs, and freeing up more time for what actually needs a person.

If your competitor saves 10 person-hours a week on operational tasks thanks to three well-thought-out workflows, they can:

• Handle twice as many leads in the same time.
• Sell at a higher margin because their operating cost is lower.
• Or simply live better — which is also a valid choice.

In none of those three cases is "AI replacing you". What's replacing you is a person who's learned to use AI as a multiplier. The difference is semantic, but the practical outcome is the same.

What to do if you were planning to invest €5,000 in AI this year.

If you have a limited budget to try AI, here's the order of priority I'd give you:

1. Efficiency audit (1-2 weeks, €1,000-1,500). Identify the 3-5 jobs in your team that are repetitive, well bounded, and clearly boring. Without the audit, anything you build is a shot in the dark.
2. Implement 1 small workflow (2-3 weeks, €800-1,500). Pick the case with the most impact and least risk. Measure before and after.
3. Team training (1 day, €600-1,000). Make sure everyone knows how to use Claude/ChatGPT day-to-day for their individual tasks (writing, summarising, translating). That alone is a quiet multiplier.
4. Keep the rest in reserve (~€1,000-1,500). When the first workflow proves ROI, we build a second. If it doesn't prove ROI, we don't build any more.

If someone is proposing to "implement AI" without this order, they're selling the outcome without understanding the process. And, as we've seen, it's usually better to disinvest from a badly thought-out process than to invest more.