The Command Center is a fast analyst, not a system of record. This page lists every way it lies to you and the specific habit that catches each lie before it reaches a board deck or your bank.
Everything here assumes you have already wired data in and built a few dashboards. The point of this page is to make the thing trustworthy enough that you actually run real numbers on it.
A model that is right 95% of the time and silent about the other 5% is more dangerous than one that is right 80% of the time and tells you which 20% to check. The job is turning silent errors into loud ones.
I give you the failure mode, why it happens, and the guardrail you design in. No hand-waving.
A reasoning model is a brilliant analyst who occasionally fat-fingers a number and never volunteers that he did. Your system of record stays Shopify, Stripe, your bank, and your warehouse. The Command Center sits on top and interrogates them. It is never the source of truth itself.
Write one sentence for your team and pin it: these numbers are a fast analyst, not the books. Spot-check monthly. Never wire a number straight to the bank or a board deck without the verify step.
If you internalize nothing else, internalize that verification is not optional polish. It is the product. A Command Center without verify clauses is a confident guess generator.
This is the big one. A model can add a column wrong on a large CSV, transpose a figure, or invent a column that was never in the file and compute against the phantom. It does all of this in fluent prose. Nothing about the output looks wrong.
Two separate problems hide under one symptom. Arithmetic error (the model does the math wrong) and fabrication (the model references data that does not exist). Both surface as a number that reconciles to nothing.
Why it happens: when the model sums a 200,000-row CSV in context, it is doing token-by-token reasoning, not running a deterministic sum. The bigger the file, the more room to drift. Fabrication happens when a prompt asks for a column the export does not contain and the model infers one rather than telling you it is missing.
Make every prompt verify itself a second way. This is the single highest-leverage habit. Every dashboard prompt ends with a clause that forces the model to recompute a key total by an independent method and report both numbers. If they disagree, you see it. Paste this at the end of any prompt that sums money:
VERIFY: re-sum total spend a second way (sum each channel column
independently, then add those subtotals) and confirm it matches your
first total. Report BOTH numbers. If they disagree, stop and tell me
rather than picking one. If any column you used isn't actually in the
file, name it and stop.
Push the arithmetic to a deterministic engine. The most reliable fix is structural, not promptable. The cleanest version of this build is not "model adds up a CSV" at all. It is scheduled ingestion landing raw data into a warehouse (BigQuery is the recommended default for most at this size), a conformed daily model where MER, blended CAC, and contribution margin are computed once in SQL, and Claude connected to one warehouse MCP to reason over already-computed numbers. The model writes the query; the warehouse does the counting. Source for the warehouse path and the official BigQuery MCP: https://docs.cloud.google.com/bigquery/docs/use-bigquery-mcp (captured June 2026, verify). A model adding up a CSV by hand is the least trustworthy configuration on the spectrum; a model querying a warehouse is the most. Where you sit decides how hard you verify.
There is a second-order win here most operators miss. When you compute blended metrics in the warehouse, you also kill the triple-count problem: Meta, Google, and Klaviyo each claim the same order, so summing platform-reported revenue inflates the top line. Define revenue truth as Shopify/Stripe at the daily grain, never the ad platform, and report MER (spend against actual total revenue) as the headline. Per-platform ROAS is a directional hint. (Modeling guidance from the build architecture, captured June 2026, verify against your own stack.)
Spot-check against Shopify monthly. Once a month, pull one number you already trust from Shopify or your existing dashboard and eyeball it against the Command Center's version of the same number. If they match all year, your trust is earned. The first time they don't, you have found a schema or definition problem before it cost you a decision.