Circo Studio · Knowledge Management
Essay · Brand piece
Circo Studio Idea Lab
ES / EN
Open essay Knowledge Management
A look at the work that doesn't show

AI is a
mirror. Barely.

A language model, however sophisticated, can only read what is there. If the data is incomplete, disordered, or stale, AI repeats it with confidence — and that confidence is dangerous. This essay is about knowledge management: the continuous work that decides whether the AI we deploy will produce real value or just plausible answers over a broken context.

Over the last few years we've learned something that repeats with every client we work with: AI's value does not come from the model. The model is a commodity — in two years, the next one will be better, cheaper, and bundled into something else. What is not a commodity is the context we give it. Clean data. Complete records. The organizational memory the system knows how to retrieve when needed. That work — silent, continuous, almost never urgent — is what separates a useful AI from one that hallucinates with style.

The evidence is recurrent: systems where the same status coexists with three different spellings, where critical fields are empty in most records, where business knowledge — the why behind a decision, the cause of a loss, the context of a relationship — is buried in email threads or in people's heads. All of that does exist, but not in a place the AI can find.

This is no one's fault. It's the normal physics of any living operation: work pushes, data accumulates, priorities shift. But AI does not forgive that physics. If the data isn't there when the model looks for it, it simply does not exist. And then the summary sounds good but omits what matters. The suggestion is plausible but ignores the context. The report is tidy but lies.

Without continuous curation, AI does not scale: it degrades. And it does so with good handwriting.

Curating data is a job, not a project

There is an important difference between implementing a system and maintaining a system that is useful for AI. The first happens once. The second happens every month. It involves capture policies, inconsistency review, enrichment of critical fields, retirement of obsolete information, decisions about what to archive and what not to. It is librarian's work: boring, invisible, indispensable.

A single inconsistency, in a single record, alters nothing. Multiplied by thousands, over years, it produces what every data team knows: a system no one fully trusts, because everyone knows something is off somewhere. And if the human team doesn't trust it, the AI shouldn't either.

Working metaphor

Curating data is like cleaning a professional kitchen. No one remembers the person who cleaned well yesterday. Everyone remembers the one who didn't clean when today's food comes out wrong. The work doesn't go away: if no one does it on purpose, the next incident does it — and then it hurts.

What it means in practice

The AI capabilities deployed over an operation — summarizing threads, monitoring markets, suggesting next steps, extracting entities, classifying automatically — are all consumers of context. Each is only as good as the data it finds. Implementing AI well is inseparable from incorporating knowledge management habits the team has to turn into routine:

Practices of continuous curation

  • Immediate capture, not deferred. Critical fields get filled in at the moment of the event, not afterward. Recent memory is infinitely more accurate than later reconstruction.
  • Periodic enrichment of the archive. Inherited data without structure is matter without form. That enrichment — assisted by AI, validated by humans — is what turns archive into intelligence.
  • Regular inconsistency audit. Fields that contradict each other, statuses that don't reconcile, impossible dates. AI can detect them; the team has to decide what to do with them.
  • Formal retirement of stale information. A record that has been inactive for years is not "archive": it is noise. Telling history apart from obsolescence is part of the craft.
  • Ontology review. Categories (statuses, types, reasons) get adjusted with experience. If "Other" holds 40% of the cases, the list came up short.

An honest way to understand the value

There's a tempting myth worth dismantling: that you can throw anything at the AI and it figures it out. That all you have to do is upload the emails, the loose Excel files, the old documents, and the WhatsApp screenshots, and the model "takes care of it." In practice it doesn't work like that, for two reasons that reinforce each other.

The first is about quality: the more noise in the context we feed the model, the more confidently it hallucinates. A system with three versions of the same status, records with contradictory fields, and unstructured text produces plausible but wrong answers — and produces them with the same certainty as the good ones. The user has no way to tell them apart.

The second reason is economic. Each model call is paid for in tokens, and input tokens cost too. If to answer a question we have to send the model the entire archive — emails, histories, manuals, attachments — we are paying for noise. At the scale of a real operation, that noise is not marginal: it multiplies the monthly bill without improving quality.

Scenario A · disordered context

"I throw it all at the model and let it sort itself out"

Input tokens / query~6,500
Real relevance~60%
Hallucination riskHigh
Relative monthly cost1.0×
Scenario B · curated context

"I select what's relevant and structure it"

Input tokens / query~2,100
Real relevance~95%
Hallucination riskLow
Relative monthly cost0.21×

* Typical reference values for medium-sized operations with general-purpose models. The cost difference holds at around 4–5× regardless of the specific model chosen.

The more ordered the context, the lower the noise, and the better — and cheaper — the AI.

That's why not everything is a chat

The third design point that emerges when you do the math: not every interaction with data is solved by a chatbot. A chat is the right interface when the question is open, exploratory, or conversational. But when the data structure is known — an identifier, a status, a date, an amount — a form with validation is infinitely more precise, cheaper, and more reliable than a chat with a model.

The rule is not new. Well-designed systems always knew this: capture with structured fields, validate states with closed lists, link records by identifier. You don't need an AI for that — you need a well-thought-out form. AI comes in where the form falls short: interpreting free text, summarizing threads, detecting patterns, proposing connections. That's where the model adds value no other tool can give.

The right tool for each situation

The data is known and discrete
Form with validation. Closed list, mask, strong type. Zero model calls, zero tokens, zero hallucination.
We need to see the aggregated state
Dashboard / reporting. Metrics calculated over structured data. Instant response, no per-query cost.
There is free text to condense
Generative AI over curated context. Summary, extraction, assisted classification. Human validates if the action is sensitive.
The question is exploratory and open
Chat with the model. Free conversation over structured data the chat knows how to query. Measured use, not by default.

Our discipline at Circo is simple: if we can be precise, we are precise. A validated field before a prompt. A dashboard before a summary. A structured summary before an open chat. Generative AI is the last resort, not the first — and that is precisely why it works so well when we use it: it arrives clean at the point where it actually adds value.

A mutual commitment

The ROI of AI does not depend on the model we choose. It depends on how much order we give to the context it consumes. The invisible work of knowledge management is where value gets created; the model is merely where it becomes visible. That's why we put as much effort into data habits as into technical implementation: because without the first, the second is theater.

This essay is an invitation to think about AI in the coming years not as magic, but as discipline. It has new tools, yes, but the path for those tools to return value is old, known, and tedious: take care of the data, order the context, be precise where you can. The rest — the models, the platforms, the providers — is the part that changes all the time.

Where others see old data, we see knowledge waiting to be ordered. That order is the work. That work is what makes AI, eventually, return value that can be measured.
Circo Studio Idea Lab ← Back to the Lab
Essay · Knowledge Management
Circo Studio