On October 8, 2026, Anthropic released a beta of Claude Dashboards, a tool that lets you connect your company’s data warehouse, ask a question in plain language and, a few minutes later, have an interactive dashboard with live data. Each chart even comes with the SQL query that produced it and the time of the last refresh.
The demos are genuinely impressive, and LinkedIn is already full of posts declaring the end of BI tools and their replacement by Claude dashboards. That swap, however, is less simple than it looks.
Five days after the launch, Jeremy Juventin, who leads the Data and AI team at PrestaShop, published the article Claude Dashboards is magic. Don’t get fooled. on Medium. In his view, the tool works as a great analyst and a terrible source of truth, precisely because it is too good at the wrong task.
Juventin’s warning and why it makes sense
The magic Juventin talks about is real. Speed, SQL written straight against the tables and live data make Claude Dashboards very valuable, especially for people exploring data. His warning is about the moment when that result stops being an exploration and starts being treated as the company’s official number. According to him, that is where three breaks happen.
- The end of unified metrics: In a well-governed company, revenue is defined once and every dashboard reuses that definition. A dashboard born in a chat writes a brand-new query, from scratch, for every chart. So two people can ask the same question and get different numbers, both looking professional, with no warning that they do not match.
- The end of lineage: The trail that takes data from the source system to the dashboard, through ingestion and modeling, breaks at the last step. The query lives inside the chat, outside the code repository, and when someone renames a column nobody knows which dashboards will break.
- The end of governance as a mechanism: With no code review, no version control, no certification of what is official and no owner, dashboards multiply by nature, since every new question creates a new dashboard. Juventin projects that within six months a company can pile up hundreds of them, each with its own definitions and none documented.
"The AI isn't the risk. An AI with no access to your definitions is."
Juventin makes clear he is not against the tool. At PrestaShop, his team uses nao, an open-source analytics agent that does essentially the same thing. For him, the difference is where the context lives. There, metric definitions, business rules and modeling are files that the data team builds, versions and tests before deploying. So the agent uses the company’s definition instead of inventing a new one for every question.
We at Flip agree with him, not least because that is the premise we built our tool on.
The 4 stages that come before any Flip answer
Bringing this to our world: when a distributor or a retailer starts using Flip, the first deliverable is a data layer built specifically for that company. It is assembled in stages, and each stage is tested before the next one exists. The language model, the part of the AI that talks to the user, only enters at the end of that chain and only sees what the chain produces.
- Extraction from the ERP with no transformation: Every table from the management system (customers, products, orders, invoices, inventory, purchases and prices) arrives exactly as it came, with a daily load audit. Nothing is corrected at this stage.
- Canonical layer with one definition per concept: Here the data is organized into standardized tables, each with a clearly defined level of detail, stable keys and fixed-type columns. This is where it is decided, once, what net revenue is, where cost comes from and which date counts. The data team writes, reviews and versions that code in the company’s repository.
- Summaries and predictive models: Aggregations by customer, product, branch and period, plus the stock-out, forecast, churn, elasticity and offer models, all read the same canonical layer. Revenue per customer in a summary has to reconcile with the canonical layer, and that is checked on every run.
- Data dictionary and the client’s domain: It gathers the glossary of terms, the metric catalog with formulas and caveats, the rules for resolving ambiguity (after all, is a “sale” an order or an invoice?) and the list of what each topic does and does not answer. This is what Flip consults on every question.
To give an idea of what this means in practice, at one of our clients, a fresh produce distributor, the dictionary holds 202 metrics. Each one records its definition, calculation method, source table and caveats. Gross margin, for instance, can only be calculated from invoices, because an order does not yet carry the real cost of goods. Since that rule lives in the dictionary, Flip has no other path to that number.
At another client, an auto parts distributor, data comes from 23 different sources in the origin system and goes through about 35 models before it becomes an answer. At every hop from one stage to the next, the numbers are checked against the previous stage, within a tolerance defined in advance.
How Flip answers each of these problems
Unified metrics with one definition per concept
Flip does not point the language model at raw ERP data. It works on the canonical layer, the summaries and the models, always together with the company’s dictionary. Every topic it knows how to handle, such as churn, stock-outs, ABC curve, checkout or inventory adjustments, declares which tables it may answer from and what is out of scope. That is why two people asking “what was this month’s margin?” reach the same base, the same formula and the same number. When the question is ambiguous, Flip says which interpretation it chose.
Lineage with a trail on every answer
Every Flip answer records the question, the topic identified, the query executed and how long it took. On screen, the “How I got here” button opens the execution trail and shows the path Flip took, with the tables it used, the filters it applied and the hypotheses it tested. That query is kept in the company’s audit trail, next to the layers that feed it, all versioned in the same repository. So when a column changes its name, the data team can see the impact from the ERP all the way to the final answer.
Governance with rules that work day to day
In his article, Juventin defines governance as a set of essential mechanisms, not a committee. At Flip, there are five:
- Reviewed and versioned code. Every table structure, transformation and migration goes through review and is kept in the company’s repository with its history.
- Quality checks on every load. Every time the data is refreshed, a battery of checks confirms everything is in order. At one of our clients there are 95 checks spread across the 8 stages of the process. They verify, for example, that there are no duplicate records, that no important key came in empty, that net values add up, that the row count matches the source, that every item has a cost and that no personal data shows up in the tables Flip reads. Each check has a severity level. The most critical ones stop the load and prevent the following stages from running on bad data, while the others raise alerts or are simply logged. Every result goes into a history that the responsible team follows.
- End-to-end reconciliation. At each stage, Flip checks whether the final number can be traced back to the source, within a tolerance defined in advance. Whatever has not been verified yet shows up as “not verified”, never as approved.
- Configuration validated before going to production. Each company’s rules live in a versioned file that is validated in a test environment and only then goes to production, without being rebuilt. The instructions that guide the language model are versioned and auditable too.
- Read-only. Every query is validated before it runs and is only allowed to read, never to change anything. No technical error reaches the user’s screen. Each company’s data is isolated, with role-based permissions and an audit trail.
Answers that already come with their caveats
The last item on Juventin’s list of recommendations is to train the people who read the numbers, not just those who build the dashboards. Flip contributes to that in the answer itself. It states how far the data goes and checks whether percentages add up to 100%, whether there are negative quantities where there should not be any, and whether the numbers in the text match the ones in the table. A forecast never arrives alone either. It comes with its error measured by WMAPE, an indicator that weights the error by sales volume, and compared with a simple method that just repeats the recent average. That way, the decision maker knows how much to trust the number before placing the order.
Juventin’s recommendations and how Flip responds to each
| What Juventin recommends | How Flip handles it |
|---|---|
| Separate exploration from reporting | Flip delivers the official number with the agility of a conversation. The question is free, but the definition used is always the company’s, so the answer is ready for the meeting. |
| Feed the AI your definitions | The company’s data dictionary, metric catalog and business rules are part of the context of every question. The model uses metrics; it does not invent them. |
| Promote, don’t proliferate | Recurring questions become schedules or executive summaries, with the same definition, delivered to decision makers. Saved analyses become notebooks shared with the team. |
| Restrict sources | Flip only sees the curated layer (canonical, summaries, models) through views with no personal data. The raw ERP is out of its reach. |
| Train your readers | Every answer states the base used, the time coverage, the sanity checks and, for forecasts, the error measured against the naive method. |
When to use Claude Dashboards and when to use Flip
Flip is not a free-form exploration tool over a generic database. If you need a fast analyst for questions that will never leave your screen, Claude Dashboards and similar products are excellent, and Juventin’s article offers good recommendations for using them safely.
Flip was built to deliver the number the sales director takes to the results meeting, the forecast that sets this week’s purchase and the stock-out alert that arrives before the customer complains. Those numbers need to be the same today, tomorrow and for anyone who asks. That is why every answer is born from a definition, goes through checks and reconciliation, and arrives with its path explained.
As Juventin reminds us, generating a dashboard has become easy and fast. The hard part is still deciding what each number means. Flip was built to start exactly there.
Want to see Flip working with your company’s definitions?
Tell us which problem weighs most on your business today, whether it is stock-outs, demand forecasting, customers buying less or margin. We will show you Flip working on it, with your company’s numbers and language.
Talk to us to learn more: joao.chencci@agrconsultores.com.br