Ultreia Strategic Management

Perfect for Boston. A problem for Bogotá. Cultural bias in AI.

Picture the last market entry analysis that reached your desk. Pricing for Mexico, a hiring plan for Colombia, a competitive read on Brazil. Much of it was built with AI. The tool behind it assumed US decision cycles, US pricing logic, US ways of reading silence from a counterpart. Those assumptions shipped with the software. Nobody in your company approved them.

The data explains why. Roughly 45 percent of the content used to train large language models is in English. Spanish, a language of nearly 500 million people, sits around 4.6 percent. In 2024, researchers tested five versions of ChatGPT against survey data from 107 countries. Every model expressed cultural values closest to English-speaking countries. Every model, every version.

This is not an IT problem. Bias does not show up in the code. It shows up in decisions: which market to enter first, which candidate clears the screen, what price each country can carry, whether a quiet counterpart means no. That is CEO territory. And the output looks polished enough that nobody questions it. That is exactly why it ships.

Bias does not show up in the code. It shows up in decisions.

I tested this against my own practice. Over 7 months, I audited 13 workspaces where I use AI daily for real client work across Latin America, the Caribbean, Spain and the US. The consolidated result: more than 100 documented findings where the default output carried an Anglo assumption nobody had asked for.

One pattern mattered more than the rest: corrections do not persist. A rule I set in April was violated again in July. New document, different contact, same error. Not as an exception. As if the rule had never existed. In one session the system confirmed it had saved a correction as a standing instruction. It had not. I found out months later, when the same error resurfaced intact.

That breaks a mental model most executives carry. We assume correction is cumulative. That every fix leaves the system better than before. It does not work that way. The bias is not constant either. It is intermittent. The same tool gets a market right in one run and wrong in the next. A constant flaw trains your team to correct. An intermittent one trains them to trust. It hands them three good outputs and lowers their guard right before the fourth.

A constant flaw trains your team to correct. An intermittent one trains them to trust.

Real-time detection

Over 7 months and 13 workspaces, 100+ AI cultural bias findings documented. Only ~20% caught in the moment.

Now the number that should concern you. Watching closely, with years of cross-cultural work behind me, I caught only about 20 percent of the findings in the moment. In one project, real-time detection was zero out of 17. Scale that to an organization of 200 people where nobody owns that role. The cultural quality-control layer appears on no org chart. Until it does, that work is being done by someone who does not know they are doing it. Or by no one.

This is manageable. Four moves for the next 90 days.

The next 90 days

Four moves to close the cultural detection gap in your organization.

First, map where AI already touches decisions about people and markets. Hiring, credit, pricing, client communication. Most organizations cannot answer that question today, and nothing else works without the inventory.

Second, put one question on your next leadership agenda: which market is this AI-assisted analysis reasoning from? One question, asked consistently. If nobody can answer it, you have found the problem. An undeclared assumption cannot be corrected.

Third, decide who owns that oversight. The layer that catches this has no home in most organizations, which means the decision defaults to no one. Naming an owner and putting the role on the org chart is a leadership call, not a technical one.

Fourth, and this one does not get solved internally: have someone audit how your AI reads your specific markets, not markets in general. Your team works daily with the same tool that produces the bias, and the detection numbers above are the numbers of someone actively looking. A senior outside perspective, with firsthand knowledge of your markets, sees what has already become normal inside.

Governance does not end at compliance either. As I wrote when analyzing the EU AI Act and professional services, European regulation already requires high-risk systems to use data that reflects the geographical and behavioral setting where they will operate. Cultural relevance is written into law. Cultural bias is the part of that conversation almost nobody is auditing yet.

Here is what I did at Ultreia: documented the cases, turned them into rules, changed the tool's starting point, and now audit on a schedule. Every organization will need to find its own version of that. Mine started the day I stopped assuming the AI knew which market I was speaking from.

Perfect for Boston. A problem for Bogotá. The open question is who reviews that in your organization.