And the worst part is you think you're making informed decisions. AI just pulled all of it back out. What you do with it decides the next ten years of your company.
By Eddie Arenas, CEO of Quarks Data
Monday, 8:40 a.m.
Commercial committee. Someone asks the question worth millions: which of our large accounts is about to stop buying from us?
Silence. Sales pulls a report. Finance pulls another. The numbers don't match. Someone says they'll check and get back to everyone. Nobody gets back to anyone. Three weeks later the account is gone, and the next meeting is about why it left.
It isn't a hard question. To answer it, sales, finance, and logistics would have to be talking about the same customer. In sales it's a client. In finance it's a tax ID. In logistics it's an account number. Three systems, three names, zero connection.
You aren't short on data. You're short on a place where your business exists as one thing.
Forty years of deciding with 10%
For decades the industry sold you one answer: structure it. More ERP, more warehouse, more dashboards. It worked as far as it could work. What nobody told you was how much of your business was being left outside.
McKinsey put it in uncomfortable numbers: structured data (SKUs, product specs, transactions, balances) is roughly 10% of the information available. The other 90% is unstructured. QuantumBlack, McKinsey's AI arm, estimates that 80% to 90% of enterprise data sits trapped in that format.
Trapped is the right word. Think about what's in there. The WhatsApp thread where your distributor told you five months ago he was buying less because a competitor gave him better credit terms. The shelf photo your rep took that never left his camera roll. The physician's note explaining why a patient abandoned treatment. The email where you negotiated the pricing exception. None of it fits in a JOIN. None of it has a schema.
So you built your intelligence on the 10% that fit in a table and called it the truth about the business. The other 90% kept existing, kept moving money, and kept living in exactly one place: your channel manager's head.
That's your best predictive model. And it walks out the door the day he resigns.
What AI actually changed
The popular reading is that AI gave us better answers. That's a weak reading. What AI changed is what counts as data.
A contract in PDF becomes a set of entities, obligations, and dates. The shelf photo becomes a facing count and a projected stock-out. The WhatsApp thread becomes a purchase intent with a date, an amount, and a counterparty. Extraction, classification, vector embeddings, graph structuring: what used to be noise becomes raw material you can query by meaning instead of by keyword.
That isn't prettier BI. It's a domain expansion. Your company just multiplied by ten the surface of itself it can model.
And that's where almost everyone crashes.
Readable is not governable
The numbers in McKinsey's State of AI 2025 read like a mirror: around 88% of organizations already use AI in at least one function, but only 39% attribute any EBIT impact to it, and most of those report under 5%. Roughly 5.5% reported meaningful value. With agents it's worse. Nearly two-thirds of companies worldwide have already experimented with them and fewer than 10% have scaled them to tangible value. Eight in ten name data limitations as the blocker.
In Latin America the gap is sharper. The World Economic Forum report with McKinsey finds that only 23% of the region's organizations generate any economic value from AI and just 6% report significant value creation. Only 10% connect their AI to business strategy. The opportunity on the table: between $1.1 and $1.7 trillion a year.
You already know what failure looks like, because you probably lived it. You signed, you watched a dazzling demo, you deployed. By month three the system contradicts the official report, nobody can trace where the number came from, and the pilot dies quietly with everyone being polite about it.
AI didn't fail. A layer was missing.
The layer where meaning lives
Between raw data and AI applications there is a semantic layer: the place where data becomes knowledge. In practice it's implemented with ontologies and knowledge graphs. The Ontology defines what things are, how they relate, and what rules govern them. The graph connects the real data from all your systems into a network of entities.
Without that shared foundation, agents act on contradictory interpretations of the same data and the error rate grows with scale. Translated: the more you automate, the faster you're wrong.
What makes an Ontology powerful isn't that it's a dictionary. It's that it's executable and versioned. Concepts like important client, critical SKU, or acceptable exposure aren't universal. They're yours, specific to your industry and your regulatory context. Codifying them explicitly, with an owner, a version, and a change history, is what separates an agent that executes from an agent that hallucinates.
And it changes the nature of your software. A business rule stops being a paragraph in a manual or a condition buried in a stored procedure, and becomes a first-class artifact: queryable, testable, auditable, and modifiable without rewriting the application.
That's the whole thesis. Your entire business can live in that layer. Not a copy of your business. The business: the operating model, the decision logic, the metrics, the permissions, the traceability. And on top of it, as consumers of one shared model, your dashboards, your agents, your APIs, and your applications.
Three things almost everyone skips
Treat ingestion as a product. Batch, streaming, structured or unstructured: everything comes in once and stays usable for everyone. A pipeline per use case is a factory for parallel truths.
Share meaning, not just data. A lake with everything in it and no common definitions isn't a foundation, it's a warehouse. Analytics, models, and agents have to understand the same thing by the same thing.
Governance travels with the data. Unstructured data is ingested, transformed, and recombined continuously. Access controls, lineage, and quality checks belong embedded in the pipeline, not in a quarterly review. When an agent assembles its context by retrieving documents dynamically, you need a record of which model accessed what, under which policy. Without that you don't have an audit trail, you have faith.
And a corollary that always gets ignored: what the agent produces is new data too. The day you accept agent output without traceability, you've started contaminating your own core.
Ontology is local, and that's your advantage
Here's what no global platform is going to solve for you.
An Ontology is the semantic model of a domain in its context. There is no universal version for banking, none for pharma. A generic layer doesn't model multiple simultaneous exchange rates, or WhatsApp as a primary and contractually relevant commercial channel, or informal distribution networks, or importation cycles, or FEFO by lot under local health authority, or the regulatory surface of SUDEBAN in banking, OFAC in cross-border, INVIMA or COFEPRIS in healthcare. Global software treats that complexity as a feature request. We treat it as the starting point.
And that's where the strategic argument comes from. If the Ontology is specific to your company and your context, then it isn't a commodity you buy: it's an asset you build. It's the only thing in your stack a competitor can't license. The more of your business lives in the core, the more impossible it becomes to operate without it, and the more expensive you become to replicate.
From cost line to revenue line
A pharmaceutical distributor moves millions of transactions a month and sells blind to its network of allied pharmacies: it has no idea what to recommend to each one. Once procurement, inventory, and demand live in one model, you can run a recommender on top that tells every partner what to order and when.
And then something happens that wasn't in the original business case: the distributor starts selling that recommender to its network. Its data stops being a cost line in IT and becomes recurring revenue.
That's the pattern. The lab that connects sell-in, sell-out, and channel behavior stops reacting late and runs real CPFR with its distributors. The clinic that connects prescription, dispensing, and follow-up catches treatment abandonment before the relapse. The bank that consolidates the customer model sees deterioration before the default. In every case, the question that used to die unanswered now has an answer and an action.
Monday, 8:40 a.m. Ten years from now.
Same committee. Same question.
Except that before anyone sat down, the core had already answered it. Three accounts flagged, each with its reason: one lowered order frequency, one mentioned a competitor's credit terms in a conversation last week, and the third has two lots nearing expiry that nobody had cross-referenced against its rotation history. The agent already reconciled which rate each transaction was booked at. The draft message to the partner is waiting for approval. The projection of which SKU runs out on Thursday is already out, by store.
The meeting isn't for putting the puzzle together. It's for deciding. Twenty minutes instead of three weeks, and the channel manager is no longer the only place in the world where that judgment lives.
That's what's coming, and it isn't science fiction: it's architecture. In ten years data won't be something your company queries. It will be the place where your company thinks. Dashboards will look like what they always were, an old photograph of a business in motion. The advantage won't sit with whoever has the biggest model, because we're all going to have access to the same models. It will sit with whoever modeled their business with enough rigor that a machine can reason over it without inventing.
There is a difference between surviving this continent and learning to design from it. The 6% capturing value from AI in Latin America today don't have better algorithms. They have better infrastructure: AI embedded in core processes, metrics with owners, and a data architecture that holds. None of that shows up in a demo. All of it is exactly what separates executing from experimenting.
The next decade won't reward whoever buys the newest tool. It will reward whoever built the core their business ended up living in: one model, governed and versioned, spanning the 10% that was always in tables and the 90% you were throwing away.
When that exists, your company stops needing someone to put the puzzle together by hand. It answers itself.
That's what we build in Essential 137, and you can explore it at Essential or in our work as an AI company in Venezuela.
References
All figures cited come from McKinsey & Company and QuantumBlack, AI by McKinsey: Latin America in the Intelligent Age (with the World Economic Forum, January 2026), The State of AI in 2025 (November 2025), Building the Foundations for Agentic AI at Scale (April 2026, based on Rewired, 2nd edition, Wiley 2026), Charting a Path to the Data- and AI-Driven Enterprise of 2030 (2024), and Unlocking Consumable Data for Generative AI (2025).