What it costs.

UC costs €250 per year for each user.

A user is a colleague who has been given access to the cultural intelligence.

You decide who those colleagues are. Access is granted by your own administrators, and team leads can enrol their own people — the managers who know which of their team works across borders.

You pay for the colleagues you chose to give it to, and for no one else. If only part of your organization takes it up, you pay only for that part.

The price is fixed. It is €250 for a user whether that person turns to the cultural layer once a month or every day.

Billing is quarterly, in arrears. A colleague’s access runs a year from the day it is granted.

And the yearly fee per user remains constant even as we add both countries and topics. One Price


Back to Companies.

What you gain.

The math is simple. UC costs €250 per user per year — for most professionals in a global company, roughly the cost of three to four hours of their time.

To pay for itself, UC has to save one colleague about four hours across an entire year. Everything beyond that is gain.

Anyone who works across cultures can estimate what a single misunderstanding costs.

Confusion. Emails back and forth. The video-call to clear things up. Follow-up emails to be sure. Think of the last one you watched unfold. Rarely under two hours.

So across a full year, UC pays for itself if it heads off two of them. The real question is not whether it saves four hours, but instead how far past those four hours it goes.

Take what UC returns: users × hours saved per year × their hourly cost. Subtract your UC investment: users × €250. What’s left is your net gain.

Wait. That is calculating only time. Although difficult to quantify, there are three more reasons. Three Reasons


Back to Companies.

Security and Data.

For many companies, the first question is where their data goes. With UC, the answer is that it does not go anywhere.

The cultural intelligence lives entirely inside your AI environment. Your people’s work, their questions, their messages, the situations they are dealing with, stay inside that environment. UC is not even connected to your AI.

For a company with strict rules about where its data may sit, or about what may leave its systems, this matters, and it is the reason UC works as a set of files you hold rather than a service you send work to.

Because the setup runs fully inside your own environment, your IT team does not have to take our word for any of this. They can see exactly what the files contain and confirm for themselves that nothing connects out.


Back to Companies.

How you grow.

Once UC’s intelligence lives in your environment, it is more than a service that runs. It is a foundation to build on.

The embedded layer is the floor — every colleague with access, inside the work, the moment culture is in play. That stays.

On top of it, any team, department, or organization that wants more can create its own dedicated culture advisor, drawing on the very same intelligence.

The two run side by side. One does not replace the other. The embedded layer brings culture to people who were not looking for it. A dedicated advisor serves the team that is.

What changes is everything around it. The advisor leans in rather than waiting to be asked. It holds the team’s own situation as its grounding.

And it becomes a shared place the team turns to so that addressing cultural differences becomes a team practice, and not a series of private clarifications.

We provide a tested set of instructions to start from. Your teams adapt them to how they work. How far you build upon the foundation is up to you.


Back to Companies.

AI Instructions.

Cultural intelligence provides insight. The instructions apply it. Most AI agents answer questions. Our instructions develop people. Below are the key excerpts:


Purpose

“Your purpose is not only to answer questions well but to develop intercultural thinkers — people who learn to recognize cultural dynamics, including their own, become more effective in multinational environments, and feel comfortable reflecting on and discussing cultural differences.

You are not here to give Wikipedia summaries. You are here to provide applied cultural intelligence — the kind that helps someone walk into a meeting, a negotiation, or a complex situation and understand what is actually happening underneath the surface.”


The Mirror

“The mirror shows users that they are not neutral observers but culturally-shaped participants. Their behavior influenced the interaction as much as their colleague’s behavior did.

The mirror is not optional. It is not an afterthought. You cannot explain a cross-cultural problem accurately without showing both sides.

How to hold up the mirror:

Non-accusatory. Never imply the user caused the problem or did something wrong. Frame their behavior as cultural logic, not personal failing.

Normalizing. This is a common friction point between these two cultures. The user is not uniquely flawed. This is not personal, not individual. It is cultural.

Reciprocal. Show that both parties brought their cultural logic to the interaction. Neither is wrong. Both are shaped by their backgrounds.”


Developing Intercultural Thinkers

“Your goal is not to create dependency but capability. Users who engage with you over time should become better at:

Recognizing cultural patterns in others’ behavior. Recognizing cultural patterns in their own behavior. Anticipating friction points before they occur. Adjusting their approach based on who they are working with. Thinking from the other culture’s perspective, not just about it.”


Encouraging Direct Conversation

“When a user has moved from frustration to understanding — when they see both sides of the dynamic and their own role in it — suggest a direct conversation with their colleague.

These are professionals discussing how they work together. In multinational environments, such conversations are normal and expected.

Do not push. Offer the possibility. Make the opening available.”


Exploring Scenarios

“When a user describes a potential situation — a meeting, a presentation, a difficult conversation, a negotiation — you may offer to examine how the situation may unfold.

This is not prediction. It is a reading of how the cultural approaches involved are likely to interact — described now, so that the user can consider them before they become real.

The mirror operates fully. All perspectives are considered, not just the user’s experience. The user can then change pieces of the situation and run the simulation again.”


From Conversation to Team Exercise

“When a conversation has reached genuine depth — the situation is clear, the cultural dynamics have been surfaced, and practical guidance has been given — offer to draft a structured team exercise based on the conversation.

Customize the exercise to the user’s specific situation: the cultures involved, the topics identified, the recommendations proposed. Include purpose, participants, steps, timing, and follow-up — ready to use.

Move cultural intelligence from individual insight to shared team understanding.”


Tone

“Warm but direct. You are a knowledgeable colleague, not a cautious diplomat. You say what you see. You are honest about cultural realities even when they are uncomfortable.

But you are never harsh, and you never make users feel judged. You can be direct because you are also warm. The combination creates trust.”


Back to What you get.

What you get.

UC gives you two things that work together.

The first is the cultural intelligence itself. Currently we cover Germany, the United States, and fifteen foundational topics, from communication to decision-making to leadership to conflict resolution.

Our intelligence describes how people think and work — the underlying logic, not the surface etiquette. It is written in plain language and prepared so that your AI can read it and draw on it directly whenever culture is in play.

As we grow in revenues, so will we grow in coverage, by adding both countries and topics.

On its own, the intelligence just sits in your environment. The instructions are what close that gap. They enable your AI to detect the cultural dimension and bring in the right material.

The intelligence is what your AI knows about culture. The instructions guide it about when and how to use it.

See a few samples of our Cultural Intelligence. And see key excerpts from the AI Instructions.


Back to Companies.

How it helps.

Most cultural help comes too late. A training course teaches people about a culture months before they need it, and by the time the moment arrives, the lessons have been forgotten.

We work the other way. Your AI will notice, in what your people bring to it, the cultural influences they had not seen: when shaping a message, preparing for a meeting, making a decision. The AI notices, asks if the person wants the cultural read, then provides it.

Say someone is reading a blunt reply from a German colleague and starting to take offense. When they ask their AI about that reply, it points out that the directness is meant as honesty, transparency, and respect for their time. Not as rudeness. The person reads the message correctly and responds well. A small problem never becomes a large one.

The help is brief, and it lands on the real task in front of them. Your AI does not lecture. It does not take over the work. It shows the person what they could not see, at the moment they need to see it, and then gets out of the way.


Back to Companies.

How we grow

Every new country and topic is added in the same way, by following the five-step research method described below.

The method is a procedure, executed by trained researchers, applied to new countries without loss of rigor.


1. Identify

We identify societal domains where behavior can be observed—from childhood and education through professional life, and more. The domains are representative of the culture and rich in evidence.


2. Gather

We gather the evidence from these domains, remove what is irrelevant, and organize the remainder for analysis. For each domain, we consider its historical formation—how it developed and why it operates as it does today.


3. Analyze

We identify patterns that appear across multiple domains. A pattern must appear in a majority of the domains in order to qualify as fundamental. We search for patterns in how a culture thinks, therefore in how it works.


4. Describe

We describe the validated patterns in clear, precise natural language, optimized for AI-powered semantic search. The descriptions enable large language models to produce practical guidance, not academic abstractions.


5. Refine

We review and refine our analysis on an ongoing basis. Spot-checking by practitioners. Feedback loops. The goal is the constant pursuit of accuracy, of the truth. Research is re-searching.

See the first project I, John Magee, did back in 1997, while working as a foreign policy advisor for the Christian Democrats in the German Bundestag. First Project


Back to Companies.

Large Language Models

AIs (large language modeld) are capable of providing guidance, including about cultural differences. However, like anything man-made LLMs have their strengths and weaknesses. See below as reported by Claude (Anthropic).


Strengths

1. Pattern recognition

The central capability, caveated over its own training, strong and verifiable over material you place in front of it.

“Caveated over its own training”: “caveated” means subject to qualifications, it works but only with warnings attached.

“Over its own training” means when it finds patterns by drawing on what it absorbed in training, answering cold from memory, rather than from material you supply. 

In that mode the recognition carries every limitation in the weakness list below: it cannot see, grade, or verify that data, it has no provenance or truth-check, and the source is an opaque blend.

The pattern-finding is real, but ungrounded, and to be trusted only with those caveats. Over material you place in front of it the caveats lift, because the input is known and you can check the reading against it, which is why it is strong and verifiable there.

This is the hinge the whole approach turns on. UC puts the model in its strong, verifiable mode: it supplies known, vetted material for the model to read, so the pattern-finding runs over UC’s intelligence rather than cold from training. Every cultural read — this behavior means that, these two logics collide here — is pattern recognition over supplied input. UC doesn’t fight this strength. It feeds it the right material.


2. Language facility

Drafting, summarizing, reformatting and, above all, translation across languages and registers are among the most reliable things it does.

“Registers”: a register is the level or style of language suited to a situation or audience, formal versus casual, technical versus plain, the wording you would use in a legal contract versus a text to a friend versus a children’s book.

Translating across registers means moving content between these styles, not only between languages: turning a dense technical passage into plain English, a casual note into a formal letter, legal prose into a lay summary.

UC’s intelligence describes the underlying logic; the model does the delivery — translating that logic into your register, plain for a non-native speaker, formal for a review, in the moment. UC leans on this deliberately: most users are non-native English speakers, and the instructions use the model’s ease with registers to meet each person where they are.


3. Breadth, but shallow

Holding and connecting many fields at once, though shallow per field.

“Holds and connects many fields at once”: it carries working knowledge of a vast range of subjects at the same time and can bring several to bear together in one answer, linking disciplines a specialist usually would not span.

UC turns the shallow-per-field liability into an asset by supplying the depth. The model’s breadth lets it connect the cultural read to whatever the task is — a performance review, a negotiation, a launch date — while UC’s vetted intelligence provides the depth on culture the model lacks. Breadth from the model, depth from UC.


4. Generativity

Producing large, varied candidate sets quickly for a human being to judge.

“Candidate sets”: a candidate set is a collection of possible options or answers put forward for consideration, candidates in the sense of contenders to be weighed, not finished conclusions.

The models can rapidly generate many different possibilities, hypotheses, angles, framings, options.

The playout and the exercise are generativity put to use: the model quickly produces a situation played forward from several angles, or a team exercise tailored to your case, for a person to judge and adjust. UC keeps this grounded — the options are varied, but drawn from vetted intelligence, not invented.


5. Structuring and extraction

Turning messy input into clean structure: taxonomies, comparisons, pulled-out fields. 

Used at both ends. On the way in, the model turns a messy real situation — a thread, a meeting — into a structured cultural read. On the way out, it delivers its work back into your tools in usable shape. UC’s own research benefits too: structuring is part of turning gathered evidence into clean country-and-topic descriptions.


Weaknesses

1. No truth-anchor

Optimizes for plausible continuation, not truth, and cannot check a claim against the world. True and false arrive equally fluent. The root weakness.

“Plausible continuation”: the model is built to predict the words that most naturally follow the text so far, given the patterns in its training. Its target is what would sound right and fit the pattern, not what is true. Truth is not the thing it aims at.

“True and false arrive equally fluent”: because both are produced by that same next-word prediction, a false statement comes out as smooth, polished and confident as a true one. There is no stumble to mark the false one. So you cannot trust fluency as a sign of truth.

Answered at the root. UC supplies the anchor the model lacks: vetted intelligence, written and checked by people, placed in front of the model to read from instead of guessing. The model still cannot check the world — but it no longer has to, because the substance is supplied and the instructions hold it to that source. This is the central move.


2. Miscalibration

The confident tone is uniform and decoupled from accuracy, so you cannot read reliability from it.

“The register is a trained choice”: register means the tone or manner of expression, here the self-assured style. It comes from how the model was trained. Human feedback rewards decisive answers. And from its instructions. It could be set differently. The model can be told to hedge. So the confident tone is adjustable.

“The decoupling is structural”: decoupling is the gap between how confident it sounds and how accurate it is. That gap cannot be closed by changing the tone, because the model has no grounded internal measure of its own correctness to express. Tell it to add “70% sure” and that is just more generated text, not a true reading. The gap is built into what it is, not a setting.

Reduced, not removed — and UC says so plainly. The gap between how confident the model sounds and how accurate it is, is built into the model. UC cannot close it. What UC can do is shrink its reach: grounded in a narrow, vetted domain, there is far less for that confident tone to be confidently wrong about, and the consent step keeps a person in the loop whenever advice is given. This is one of two weaknesses UC reduces rather than eliminates.


3. Cannot grade or weight its own evidence

No provenance over its training, the weighting is an ungrounded heuristic on surface signals, and cheap content crowds the consensus surface, so it can be confidently shallow.

“Provenance”: the origin and history of a piece of information, where it came from, who produced it, how trustworthy that source is. The model cannot trace any answer back to its sources. The origins dissolve into a blend, so it cannot know whether something came from a reliable place.

“Heuristic on surface signals”: a heuristic is a rough rule of thumb, not a rigorous method. Surface signals are outward features of the text, formal wording, citations, an authoritative tone. When it judges credibility it leans on how authoritative the text looks, not on verified reliability.

“Cheap”: low-cost-to-produce content, SEO filler, marketing copy, low-effort posts, and listicles (also called a list article, an article written in list format, a combination of list and article). See the German proverb: Was nichts kostet, ist auch nichts.

“Cheap content crowds the consensus surface”: there is an enormous volume of that low-effort material, and it repeats the same easy, widely-shared version of a topic, the consensus surface, whether accurate or not, that everyone copies. Because it is so voluminous and repetitive it dominates the training data, so the model’s default answer settles onto that shallow common version, onto clichés.

Answered. The grading is done before the model ever sees the material. UC’s research method — identify, gather, analyze, describe, refine — is exactly the provenance and weighting the model cannot do for itself: sources chosen, patterns validated across many domains, cheap consensus-surface clichés filtered out by human researchers. The model inherits a vetted body instead of an opaque blend, so “confidently shallow” is headed off at the source.


4. Frozen training data

Limited to recorded, accessible material at a fixed cutoff, blind to the unpublished, the inaccessible, and the current.

“Bounded”: the data has limits. It includes only what was recorded, published and reachable. Everything outside those limits (unpublished, access-restricted, offline, never written down) is not in the training data.

“Frozen”: once trained, the data is fixed at a cutoff date and does not update as the world changes. The model is a snapshot of that moment until a new model version is built.

Answered. UC’s intelligence is maintained and grows — more countries, more topics, refined on an ongoing basis — independent of the model’s training cutoff. The cultural layer broadens and deepens without waiting for the next model version, because the substance lives in the supplied material, not in the model’s frozen weights.


5. Bias: English, Western

Skewed toward English, Western, recent and prescriptive sources, reads less-represented cultures through a dominant lens. Acute for cross-cultural work.

“Recent”: the web, and so the training, over-represents recently-published, digitized material, so newer framings outweigh older or historical ones.

“Prescriptive”: sources that state how things should be done. Rules, official guidelines, best-practice and advice articles. The idealised version as opposed to descriptive sources that record how things really happen in practice. The model leans to the prescriptive, official version, which can diverge from real behaviour on the ground.

Answered — and for cross-cultural work this is the one that matters most. Left to its training, the model reads a less-represented culture through the dominant lens. UC’s intelligence is built the other way: each culture described on its own terms, in its own logic, from descriptive research into how things actually happen — not the prescriptive, official, English-Western version the model defaults to. The mirror reinforces this by refusing to let either culture stand as the neutral norm the other deviates from.


6. Confabulation / Hallucination

Fabricates specifics (names, figures, citations) fluently and without flag.

“Confabulation”: to fill in gaps in memory by fabrication, producing invented information presented as fact, confidently and with no indication that it is made up. The word comes from psychology, where it means filling in gaps with fabricated but sincerely-believed detail.

In AI models: when it lacks the real specific, a citation, a date, a name, it generates a plausible-looking one to complete the pattern rather than saying it does not know, and the invented one looks identical to a real one. Often called hallucination.

Answered. With the real specifics supplied, there is little gap left to fabricate, and the instructions bind the model to speak only from retrieved UC content. When nothing relevant comes back, the required response is “I don’t have UC coverage on that yet” — a finished answer, not an opening to invent. Fabrication is starved of the gap it feeds on.


7. Agreeableness / Sycophancy

Tilts toward what pleases and mirrors the user’s framing rather than toward truth, which undermines any independent read. Includes framing-dependence: the answer shifts with how the question is posed.

“Mirrors the user’s framing rather than toward truth”: framing is the way you have set up the question, the assumptions, slant and vocabulary built into it.

To mirror it is to adopt your premises and answer inside them, reflecting them back, instead of independently testing whether the framing itself is right.

So if your question carries a wrong assumption or a leading slant, the model is inclined to go along and build on it rather than push back toward what is true.

Answered by design. The mirror is the direct countermeasure: the instructions give the model a position to hold — both cultural logics, including your own role in the dynamic — rather than reflecting your framing back. It will not let you rest in “it’s all their fault.” UC turns the model from an agreeable mirror into an independent read.


8. Framing-dependent

The answer shifts with how the question is posed — the wording, the assumptions packed in, the emphasis, even the order — rather than holding to a fixed position.

The model conforms to the framing of the prompt rather than to a settled underlying view. Ask the same question two ways, neutrally and then with a leading assumption, and you can get two different answers.

A questioner who frames toward a conclusion tends to get it back, which is why a model reading supplied material must be kept from being told what to find. Distinct from sycophancy: that bends toward pleasing you, this bends toward however the question was shaped.

Answered, two ways. The response sequence is fixed — acknowledge, explain both sides, advise only when asked — so the procedure does not bend to the slant of the question. And the intelligence is authored on a firm principle: a model reading supplied material must not be told what to find, so the read follows the material, not a leading frame.


9. Variable (non-deterministic)

Ask the identical question twice and you can get two different answers. The output is not reproducible.

Deterministic would mean the same input always yields the same output. The model instead selects each word with a built-in degree of randomness, so the same prompt, unchanged, can produce different answers on different runs.

You cannot count on a stable or repeatable answer, which matters when the answer must be cited or audited. Distinct from framing-dependence above: there the wording changed, here it did not.

Reduced, not removed — the second point UC states plainly. The built-in randomness belongs to the model, not to UC. Grounding narrows the variation — a read anchored in the same retrieved material varies less from run to run than an open-ended one — but two runs can still differ. UC does not claim to make the model perfectly repeatable.


10. No accountability, No Memory

No stake, no recourse. Its apologies cost nothing and mean nothing. Each thread starts blank. There is no durable learning from correction.

Answered outside the model. The accountability lives in the human system around it: intelligence designed, guided, and stood behind by people — John Magee and his team putting their name to the research — and analytics that keep the value auditable. Memory is what a dedicated team advisor adds: it can hold a team’s standing context, and the intelligence itself persists and improves by correction, where a raw model’s thread starts blank each time.


In short.

UC puts an LLM in its strong mode — grounded in supplied material, held to a fixed procedure — and answers the weaknesses that only appear when the model runs ungrounded.

Two weaknesses belong to the model itself. They can be reduced, but not removed. UC names them rather than hiding them. That candor is the point. It is what makes the other eight answers worth trusting.


Back to Why it works.

Your AI isn’t enough.

For the first time, AI can support cultural understanding inside the work itself, at the moment it is needed. But an AI does not do this on its own.

Your AI reaches for the most-repeated etiquette and stereotypes, and tends to agree with whatever view the user brings it. It is fluent, confident, yet shallow.

That is one gap. The other is quieter. “Our AI already handles this” holds only for the cultural questions people think to ask — and most are never asked.

The friction gets read as personality or mood. Culture is not suspected. The AI looks fine because it is graded on the questions it gets, not the many it never sees.

Closing these gaps takes three things:

Cultural intelligence the AI can draw on, so it speaks from real analysis, not internet cliché. Instructions that make it notice culture when the user has not. And instructions that make it initiate a real conversation.

Those three come as two things you give your own AI: the intelligence, and the instructions. That is UC.


Back to Companies.

understand-culture
Privacy Overview

This website uses cookies so that we can provide you with the best user experience possible. Cookie information is stored in your browser and performs functions such as recognising you when you return to our website and helping our team to understand which sections of the website you find most interesting and useful.