March 3, 2026·4 min read·AI, Transformation, Strategy

Companies transitioning to AI need a data concierge before any copilot

Most Brazilian companies that committed to 'doing AI' in 2026 aren't ready yet. Here's why a data concierge is the step between the decision and the first AI feature that actually works.

Gabriel Fernandes

Data Wizard

Ler em português

The most common phrase in Brazilian boardrooms in 2026 is some variant of "let's do AI". It's past the "this is hype, let's wait" phase and into "everyone is doing it, so should we". The budget is approved. The committee is formed. The board has asked for a plan by next quarter.

And that's where the problem starts. Because "doing AI" isn't a project. It's five or six simultaneous projects, three of which the company didn't know existed until they tried. And none of them is the copilot the board imagined.

What the conference deck didn't tell you

The deck that sold AI to your board showed a chat answering CFO questions in real time. What the deck didn't show was:

The three different definitions of "revenue" in your warehouse that the model will discover and, without guidance, pick the wrong one.
The fact that your ERP doesn't have consistent foreign keys between customer and contract , so the copilot might say "X closed the deal" about someone who never closed anything.
The total absence of catalog, lineage, or governance policy deciding what the model can see. (Spoiler: by default, it sees everything, including the salary table.)
The vendor of the AI platform isn't going to fix this for you. They sell the tool. The foundation is your problem.

I wrote about this dynamic in more depth in fix your data before adopting generative AI, but the main point is: nearly every company transitioning to AI discovers, three to six months in, that the real problem was never the model.

Three things AI transition demands beyond the model

1. Canonical definitions for the metrics the model will quote

If "revenue", "active customer", and "margin" are not defined exactly once in your warehouse, any copilot you plug in inherits the ambiguity, and amplifies it. The canonical definition exercise is usually the first thing I run with clients entering AI transition. The method is detailed in duplicated metrics.

2. Lineage traceable to source

When the model quotes a number, someone needs to be able to walk, in under five minutes , from the number back to the raw row in the source system that produced it. If your architecture can't do that today, you'll spend more time auditing hallucinations than benefiting from the copilot.

3. Governance that decides what the model can see

Not a wishlist in Confluence. Enforced policy at the warehouse layer: row-level security, sensitive-field masking, allow-list of tables the model is permitted to query. In Brazil this gains extra weight because LGPD treats personal data exposed through a copilot as seriously as data exposed through a classic breach.

Where the data concierge fits in this transition

The concierge is the bridge between executive ambition ("we want AI") and technical reality ("our foundation isn't ready"). The role isn't to build the copilot, usually a vendor does that. The role is to make sure that when the copilot is plugged in, it has something trustworthy to read.

Concretely, in an AI-transition engagement, I usually sequence three phases:

Readiness diagnostic (2 weeks). Audit the current foundation against real AI requirements, not the marketing ones. Map metrics that disagree, points where lineage breaks, sensitive data accessible without controls.
Foundation build (8–12 weeks). Model the warehouse with clear layers, define canonical metrics, deploy a catalog, enforce governance at the warehouse level. Don't build AI yet. Build the base that makes AI viable.
Pilot accompaniment (retainer). When the AI vendor brings in the copilot, the concierge stays available for the questions that show up, governance questions the vendor can't answer, metrics the model is misinterpreting, scope decisions that need foundation context.

If your company has committed to AI but hasn't started, or if you're already in and the pilot is losing momentum, a 30-minute call is worth it. You walk me through the current state; I tell you whether the foundation supports the copilot you're trying to plug in, and what needs to happen first.

Assess your AI readiness

Why this is especially urgent in 2026

The AI hype cycle has moved past the optimistic phase. Brazilian and international boards have authorised budget. CEOs have announced in interviews. The political cost of walking back is high. What's left is delivering, and delivering well is harder than delivering. Companies that can distinguish between "ship an AI feature" and "transition to AI" will pull ahead. The others will spend six to twelve months on stalled pilots before accepting the correct order of operations.