In a meeting with the owner of a mid-sized private clinic, the first thing the administrative manager says is: "we've already done it. We've been using ChatGPT for a year". The clinic has three Enterprise licences, distributed between owner, administrative manager and marketing manager. Listening to them, it sounds like the case is closed. It is enough to look at what actually happens to understand that the work has only just begun.
The administrative manager has ChatGPT open every morning. He uses it to write emails, reformulate communications, summarise documents. At the end of the day he reckons he has saved "perhaps half an hour". The owner uses it less, essentially for faster online searches. The marketing manager uses it to generate drafts of social posts and articles for the clinic blog. Every time she writes a new post, she has to explain from scratch who the clinic is, what it offers, how it speaks to patients, what topics it has already covered.
Three people, three twenty-euro-a-month licences each, a fading perception of usefulness. The problem is not ChatGPT. The problem is that ChatGPT without infrastructure is a first-day collaborator, every time it opens a new conversation. It does not know who the clinic is. It does not know what was done the week before. It does not remember the style set in previous articles. It does not know who the patients are, what they ask most often, which services generate the most margin. Every interaction starts from zero, and the briefing required eats the time you wanted to save.
The difference between individual and company use
An individual ChatGPT subscription solves an individual problem. The person using it improves their productivity on personal tasks. They reformulate an email faster, produce drafts more quickly, access a second cognitive opinion. This use has value, but it has an intrinsic limit: the context lives in the user's head and is retransmitted manually at every conversation.
A company has a different problem. The relevant information is not in the head of a single person; it is distributed across people, documents, systems and company history. When a company says "we use ChatGPT", what actually happens is that each employee uses a separate instance, with their own mental briefing, often imprecise or incomplete. The result is fragmented by definition.
If the marketing manager produces twenty articles a year, each requires her to re-explain the clinic's tone, the list of main services, the forbidden words, the regulatory references of Canton Ticino, the patient target. Twenty times. The second article is written with an approximation of the first briefing, the tenth with an approximation of the fourth. The tone drifts slowly out of alignment. To an outside observer it looks as if the clinic has several voices writing, and in a sense it does.
The same pattern repeats with the administrative manager, who each time explains the type of client he is replying to. And with the owner, who each time sets the strategic context before being able to use AI as a sparring partner.
Why the context infrastructure is the missing step
Context infrastructure is the base of information a company makes available to an AI tool so that it can work knowing the company. It is not a single document. It is an organised corpus containing:
The operational registry (who you are, what you do, for whom, at what prices, in which geographical areas, with which differentiators). It is not the brochure. It is material written in a way a machine can consult and use in response, without ambiguity.
The recurring procedures (how to reply to a patient asking for an estimate, how to welcome a new client, how to handle a complaint). Not written in vague natural language, but structured enough to be executable.
The tone and lexicon. Not "we speak in a professional yet warm way", which is not enough. Concrete examples of well-done emails, paragraphs from past articles, sentences representing the style, words and constructions never to be used.
The work history. What has been published, said, written, sent. Not archived as a passive file, but indexed as consultable memory.
When these four components exist, an AI tool operates with the same cognition as a collaborator with five years in the company. Without them, it operates with the cognition of a first-day collaborator. The difference in output is not marginal: it is the whole difference between a useful contribution and a forced one.
The step many companies skip is believing this infrastructure arrives by buying the right product. There are platforms that present themselves as "ChatGPT for your company", promising to build the context from your existing documents. Sometimes it works. More often, after a few months you discover that the uploaded documents were disordered, incomplete, contradictory, and the context extracted reflects that disorder.
Context infrastructure is not installed. It is built. It is work of organisation and writing, done by someone who knows the company and someone who knows how an AI system consumes information. It is the exact opposite of a software configuration. It is an act of thinking about your company.
This thinking has an interesting secondary outcome. In building the context, it often emerges that the company itself had not made certain things explicit. The tone of voice was not defined, it was the marketing manager's intuition. The patient reply procedures were not written, they were habits transmitted by osmosis between secretaries. The differentiators compared to other area clinics were not clear even to the owner. The structuring exercise forces a clarification that matters beyond AI.
Another illustrative case: a Swiss tax consultancy with twelve collaborators introduced ChatGPT Enterprise for the whole team. After four months, the internal perception was that the tool had produced little. Looking at the conversations collaborators had actually held, a pattern emerged: three quarters of the initial briefings repeated information the practice already had in internal documents, but which was not structured in a consultable way. Each collaborator rebuilt the base from scratch based on personal memory. When the practice dedicated three weeks to structuring four key areas (enriched client records, practice technical glossary, reply procedures for recurring queries, writing style toward Italian and Swiss clients), production speed doubled with the same tool. The tool was the same; what had changed was the environment feeding it.
What changes after
Back to the clinic cited at the opening: what happens when context is structured?
The marketing manager stops re-explaining each time. She produces an article in forty minutes instead of two hours. The articles have a coherent tone with each other, and coherent with the clinic's historical articles, because the tool has access to the history.
The administrative manager starts delegating email reply tasks to a system that knows how to recognise the type of request (booking, generic clinical question, estimate request, complaint) and replies appropriately, with escalation to the right team when needed.
The owner stops using ChatGPT as generic search and begins to use it as a strategic assistant who knows the clinic's numbers, because the context includes internal reporting too. The questions change. Instead of "what are the best practices for patient loyalty", he asks "given our specific situation, what do we see in the patients who have returned less in the last year".
The same base tool, three radically different uses. The difference is not the tool: it is the context feeding it.
An owner who has given three collaborators a ChatGPT Enterprise subscription has taken an understandable step. He has met the tool before meeting the problem. It is an honest starting point, not a mistake. The mistake is stopping there and thinking the work is done.
The next step is a strategic decision, not a technical one. It concerns whether the company wants to accumulate structured knowledge about itself and make it operational, or whether it is content with individual tools that contribute but do not compound. Both choices are legitimate. The first produces compounding over time, because the context grows; the second produces linear, repeated help.
To understand where you are, an exercise lasting less than an hour: ask the collaborator who uses ChatGPT most often to show you a typical conversation. Look at the first message he writes. If it contains briefing about the company, the context, the style, you are using the tool as an individual tool. If it does not, and the tool still replies consistently with the company, you have already built the context. If it does not, and the tool replies generically, you are looking at the bottleneck.
The move from individual ChatGPT to business cognitive infrastructure is often described as a technological upgrade. It is an upgrade of thinking, more than of technology. Those who make it find themselves writing fewer prompts and deciding more, because the context previously rebuilt by hand is now the shared heritage of the practice.
If you recognise your situation in this description, the first step is not to change supplier. It is to diagnose the existing context: what is structured, what is implicit, what is missing. It is a concrete forty-five-minute conversation that produces a useful document even if you do not proceed. From there on, tool choices become consequences.