Strata is the AI analyst for businesses that run on QuickBooks-class systems and decide on spreadsheets. Ask in plain English. Every number shows its work. A human approves everything it learns. Useful the day you connect it.
Read-only by construction · Your own AI key · Your data never trains anyone's model
Pull the monthly P&L by product line → compare July against June → decompose the margin change into price, cost, and mix → reconcile the drivers against the total.
SELECT month, product_line, SUM(revenue) AS revenue, ROUND(100 * SUM(revenue - cogs) / SUM(revenue), 1) AS margin_pct FROM monthly_pnl WHERE month IN ('2026-06', '2026-07') GROUP BY month, product_line ORDER BY product_line, month;
July gross margin came in at — down from June — and it isn’t a pricing problem. Freight surcharges on hardware did most of the damage at ; new discounting on two accounts cost another −0.8, and the mix shift toward hardware explains the rest. The three drivers sum to the full −3.1 — nothing left unexplained.
Every answer arrives in this order — plan → query → number → proof → your correction.
Your business generates plenty of data — it’s just trapped in the accounting system, the CRM, and nine spreadsheets, guarded by one overloaded “numbers person.” The figures below show their work.
lost to manual data wrangling across ~10 disconnected tools.
Intuit QuickBooks survey · Aug 2024of CFOs don’t fully trust their own financial data.
BlackLine CFO survey · 2024of spreadsheets contain errors — and the month runs on them.
Panko · spreadsheet-error researchof small businesses fear wrong numbers from AI — while most already use it.
Karbon State of AI in Accounting · 2025The missing ingredient isn’t a smarter model. It’s trust — and trust isn’t a feeling, it’s machinery: receipts on every number, a human gate on every lesson, and a test suite that keeps right answers right.
Ask in plain English → trust every number → and it compounds, with your sign-off. Each piece exists somewhere. The system exists nowhere else.
Plain English in; your numbers out — revenue, margin, cash, aging — along with the exact tables it read to get there. No dashboard to build, no report-request queue, no warehouse project before the first question.
AI commentary about a report you already ran, a dashboard builder and a lost weekend — or a real AI analyst locked behind a data warehouse you don’t have.
Services carries the business at margin. Installation runs , and hardware is 11.2% after freight — freight wasn’t in your old spreadsheet’s formula.
Click any figure and see the exact query, the exact cell, and the exact version of the data it came from. Numbers you can put in front of your bank, your board, or your auditor — and defend when someone asks where they came from.
A number you take on faith. Microsoft’s own documentation says Copilot “can produce incorrect answers” — and tells you to check.
Answers you approve become your answer key — a test suite the product has to pass. Every change replays it, five trials per question, and a change that breaks a right answer doesn’t ship. And when the data can’t answer, Strata says so and names what’s missing — it never invents a number.
Nothing stands between a model update — theirs, or their AI vendor’s — and your Monday report. An answer that was right yesterday can silently change.
“We exclude intercompany transfers from revenue.” Say it once — a reviewer you trust approves the lesson — and every future answer uses it. Your definitions, your fiscal calendar, your exceptions: an analyst that’s measurably better at your company every week.
The chatbot on day 400 is the same generic bot as day one. Your corrections vanish when the chat window closes — and anything it does “learn,” nobody reviewed.
You bring your own AI key, so your data runs under your agreement — and never trains anyone’s model. Access is read-only, enforced by the database itself, not by promises. Every AI call is metered on one ledger you can read.
Their AI, their terms — training and retention buried in the terms of service, AI cost bundled where you can’t see it, and no say in which model reads your financials.
Four stages, running continuously — and a human gate in the middle of them. Nothing goes live automatically.
Every answer invites correction — a thumbs-down, a fix, a note. Nothing you teach it is lost.
Corrections queue for a person you trust. Nothing goes live without their approval.
Approved lessons fan out — to your definitions, your answer key, and every future answer.
Every change replays the full answer key. Break one right answer, and it doesn’t ship.
Everyone sells answers.
Ours get provably better every week.
Because the product isn’t the model — it’s the learning system around it.
Not a BI project. Not a quarter of consulting. A read-only connection and a first question, the same day.
Secure, read-only connect in minutes. First cited answers on your own numbers the same day.
Your definitions, your account names, your fiscal calendar — each one approved by you, then used everywhere.
The answers you’ve marked correct become the test suite. The first Verified badges land.
Scheduled reports arrive on time, watches flag what moved, and week one’s questions answer themselves.
AI that writes commentary about a report you already ran. No questions, no receipts, no learning.
Microsoft’s own docs say Copilot “can produce incorrect answers.” And the data model behind it is your unpaid job.
Real chat, real lineage — behind a data warehouse a small business doesn’t have and shouldn’t need.
Locked to one ledger, upgrade-gated — and grading its own homework.
And the system is what compounds — every answer, correction, and approval makes the next one better.
10–500-employee businesses on QuickBooks, Xero, or Zoho — multi-entity, board- or lender-accountable, spreadsheet-weary — and the advisors who serve them.
“Know your numbers without becoming the numbers person.”
A cited answer on your own books, in the first session.
“Every number defensible — to the board, the bank, the auditor.”
Click a figure: the query, the data snapshot, the definition behind it.
“Serve twice the clients with the judgment you already have.”
The two-day board pack, as a scheduled report with receipts.
“Stop reconciling ten tools. Ask one question.”
One governed layer over the ledger, the CRM, and the spreadsheets.
Enforced by database permissions, not promises. It can look at your books; it can never write to them.
Bring your own key, keep your own terms. Your data never trains anyone’s model — ours included.
Spend sits on one ledger you can read. No silent model swaps — a downgrade is always labeled.
Everything it learns passes through a review queue owned by someone you trust. Nothing goes live automatically.
Every query, decision, and approval is on the record — who, what, when, and on which version of the data.
When the data can’t answer, it says so and names exactly what’s missing. It never invents a number to be helpful.
A 30-minute working session: connect read-only — or start from our reference environment — ask your hardest question, and watch the answer show its work.
Read-only connect · No obligation · Your own AI key from day one