Mondragon Corporation: How to Research It with AI Without Getting Burned

The #1 mistake people make when researching Mondragon Corporation – the Spanish federation of co-operatives – is asking ChatGPT to summarize it and trusting whatever comes back. The model will confidently tell you Mondragon has 110 cooperatives, 80,000 workers, and €12 billion revenue. Some of those numbers were accurate in 2019. None are accurate now.

Mondragon held its 2026 annual Congress on 20 July at the Kursaal Palace in San Sebastián, with 465 representatives, and the topic is circulating across LinkedIn and cooperative economics circles. This tutorial shows the correct way to research a fast-moving organization like this using AI – with Mondragon as the working example.

Why Mondragon breaks LLMs

Think of it like trying to photograph a flock of birds. The moment you snap the shot, the formation has already shifted. Mondragon is a worker-cooperative federation founded on 14 April 1956 by Father José María Arizmendiarrieta in the Basque Country – members own shares, elect leadership, vote on strategy. But the membership itself is not fixed.

The catch: cooperatives join, merge, and sometimes leave. Ulma Group and Orona voted to depart on 16 December 2022, dropping the workforce by 13% and group sales by roughly 15% overnight. Any LLM trained before mid-2023 still counts them in – and repeats that count with total confidence.

What the actual numbers look like right now

Here’s where most sources drift. Revenue figures of €12-13 billion floating around social media come from pre-2022 data. The 2024 figures from Mondragon’s annual report:

The co-op count (~92) comes from the International Cooperative Alliance’s most recent site visit note – but that number fluctuates as cooperatives enter and exit.

The 5-prompt workflow

Use this in ChatGPT, Claude, or Perplexity. The logic is identical across all three. Each prompt has exactly one job.

Prompt 1: force a knowledge-cutoff disclosure

You are a research assistant. Before I ask about Mondragon Corporation,
state your training cutoff date and confirm you cannot access live data
unless I paste it in. If you have web search, use it and cite URLs.
Do not proceed until this is acknowledged.

Skipping this step is why most people get bad output. Without this box, the model happily invents 2025 figures from 2021 training data.

Prompt 2: paste the fresh source, then ask

Grab a paragraph from Mondragon’s official site or a recent TU Lankide article and drop it into the prompt as context. Then ask your real question.

Context (paste from mondragon-corporation.com, dated July 2026):
[paste 2-3 paragraphs here]

Based ONLY on the context above, summarize Mondragon's current
strategic priorities. If the context doesn't answer, say so.
Do not use prior knowledge.

The “only from context” instruction separates a useful summary from a hallucination. Test it: remove that line and watch the model smuggle in old facts.

Prompt 3: structure the entity

Now ask the model to map what Mondragon actually is. Four divisions to identify: Finance, Industry, Retail, Knowledge (per Mondragon’s official cooperatives page). Key co-ops to name: Eroski (retail), Laboral Kutxa (bank), Mondragon Unibertsitatea (university). Note: Fagor Electrodomésticos, often listed as a flagship example, went bankrupt in 2013 – any model that treats it as current is running on stale data.

Ask the LLM to build a table matching each division to its major co-ops. This forces the model to commit to specifics you can fact-check, rather than hiding behind generalities.

Prompt 4: challenge the model

What's the strongest critique of the Mondragon model?
Cite at least one specific event where the model showed weakness.
Do not hedge with 'some critics say' - give me the actual event.

A good answer surfaces the 2013 Fagor Electrodomésticos bankruptcy or the 2022 Ulma/Orona departures. Filler answers give you “critics argue…” vagueness. Filler means bluffing.

Prompt 5: cross-model verification

Run the same factual question through two different models – ChatGPT and Claude, say – and compare. Same number from both? Moderate confidence. They disagree? One is wrong. Check Wikipedia or the annual report directly.

Quick rule: When an LLM cites a specific Mondragon number, follow up with “what year is that figure from?” Nine times out of ten it’ll admit: 2020 or 2021.

What’s genuinely unclear is how well this workflow ages. Mondragon holds its Congress annually, which means the numbers shift every year. The 5-prompt structure stays valid – but the pasted source content needs refreshing each time. Worth asking: at what point does a federation that changes composition this often require a fundamentally different research approach?

Pitfalls

Trusting the co-operative count. Seriously – different sources quote 80, 92, 95, 102, 110. All real, all from different years. If your article needs that figure, cite the exact source and date, not just the number.

Copying the 1:9 pay ratio without context. Turns out the Young Foundation study puts Mondragon’s internal salary ratio at 1:9 versus 1:129 for FTSE 100 firms – but that applies to worker-owner co-op governance specifically. Mondragon’s non-cooperative subsidiaries abroad don’t necessarily follow the same ratio. Models quote the 1:9 number constantly and almost never add that caveat.

Treating departing co-ops as failure. According to Peeters and Schouteten’s analysis, co-ops that leave Mondragon tend to be financially strong – they exit because they no longer want to subsidize weaker members through the solidarity fund. That’s the opposite of the narrative most AI summaries produce.

When to skip this workflow

Conceptual research – the values, the philosophy, the history of cooperativism. LLMs are good at that. Mondragon’s founding principles haven’t shifted since 1956.

Casual conversations. Nobody needs the exact 2024 employee count at dinner. But a report, a pitch, a policy brief, anything that gets fact-checked? Do the full workflow.

Citation-grade academic work? Skip AI entirely. Go directly to Mondragon Unibertsitatea’s research output or the ICA’s published studies. LLMs paraphrase secondary sources without flagging that they’re secondary.

FAQ

Is Mondragon a company or a cooperative?

It’s a co-operative of co-operatives – a federation where each member firm is legally independent. Not one company.

Why do LLMs get Mondragon’s numbers wrong so often?

Training cutoffs are part of it – the 2024 revenue figure of €11.213 billion and the December 2022 Ulma/Orona exits aren’t in older models. But there’s a second problem that’s easy to miss: Mondragon is genuinely a moving target. Cooperatives join, merge, and exit regularly, so even a 2023 article’s “current” figures can already be stale by the time a model trains on it. Pasting a dated source directly into the prompt is the only reliable fix – smarter prompting alone won’t close that gap.

Can I use Mondragon as a case study for my own business?

Yes, but cherry-pick carefully. The one-member-one-vote governance model is portable. The solidarity fund between co-ops? That assumes a shared cultural context – Basque cooperative tradition built over decades – that doesn’t transplant easily. Study the mechanics. The cultural substrate is not a template.

Next step: Open your LLM now. Paste Prompt 1 above. Try it on any organization you need to research this week. If the model won’t disclose its cutoff or bluffs past the question – switch tools.