How it works
How a trial balance becomes your monthly MIS, in six stages
The interesting part of this product is not that it produces a workbook. It is where the numbers come from, what the model is allowed to touch, and why the second month costs so much less than the first.
Free to create. No subscription, and credits never expire.
Watch it once
The whole thing, end to end, before reading the detail below.
- 01
Upload your raw data
You add the raw data from your accounting system, in any format. They are encrypted as they arrive under that company's own key, used only for the runs you pay for, and kept for you until you delete them.
Before you pay for anything, the screen shows file names, sizes, sheet counts and row counts. Not what is in them: no sheet recognition, no mapping preview, no data-quality findings. Those are results, and results appear inside a paid action.
- 02
Ledgers are matched to a canonical schema
Each ledger is matched to a standard MIS head for you, and the mapping is kept for this company so later months reuse it.
Matching is by header name, never by column position, so a reordered file does not break it. Where a name is unambiguous it is matched deterministically without any model involved; only genuinely ambiguous ledgers go to the AI, and anything it cannot place is shown as Unmapped rather than guessed.
- 03
The engine computes, then the model writes
Every figure in the workbook comes from a deterministic engine. The model writes the commentary around those figures and never produces a number itself.
Commentary is generated with placeholders where figures belong, and the engine fills them. That is why a sentence cannot disagree with the table above it — the sentence does not contain a number until the engine puts one there.
- 04
The workbook is validated before you see it
A series of checks runs against the output: the trial balance ties, the statements agree, the ratios reconcile to their inputs.
A check that fails because of the data — a trial balance that does not balance, a month missing, a balance no head takes — is never hidden: the workbook is delivered with the warning on its Checks sheet and on screen, saying exactly what to look at. A check that fails because of us — a figure lost between the file and the report, a formula that does not reproduce the engine — stops the job, and nothing is charged for it.
- 05
You chat the dashboard into shape, and present from it
The first dashboard is built for you. After that you say what the board should show — a comparison with last year, a trend, a ratio of your own — and it appears beside the chat, saved for that company.
The chat describes a box or writes a formula; it never writes a figure. The engine computes every number, including your own formulas, in exact arithmetic, and each one opens to its source. Present puts the dashboard full screen for the meeting, and Save as PDF keeps a copy stamped with when its figures were computed. There is no export to slides, on purpose: a deck is out of date the moment it is made, and a number on a slide cannot be questioned.
- 06
Next month reuses everything
Load the new trial balance. If the ledger structure has not changed, the refresh makes no AI calls at all.
New or renamed ledgers are flagged for you to place; everything else carries forward. This is the property the whole design is arranged around, and it is what makes a monthly refresh cost a fraction of the first setup.
The AI never outputs a number, and everything else follows from that
The AI never outputs a number. Every figure in every workbook, dashboard, commentary line and chat answer is computed by the deterministic engine and inserted through a placeholder.
That constraint is why the commentary can be trusted at the same level as the table: they are the same figures, from the same computation, and there is no path by which a model-written digit reaches a report. It is also why Magic MIS can show the lineage of any number on demand — the number has a provenance because it was computed rather than written.
What people ask about how a trial balance becomes the report
What does the AI actually do?
It proposes ledger mappings where a name is ambiguous, and it writes commentary around figures the engine computed. It never computes, adjusts or writes a number, and it never sees your raw files — only redacted structural profiles and aggregates the server sends for that specific action.
Why does a monthly refresh cost less than the first setup?
Because it does less. The expensive part of the first month is establishing the mapping. On a refresh where the ledger structure is unchanged, that work is reused and no AI call is made — so the price reflects computation and validation rather than model usage.
Can I correct a mapping the product got wrong?
Yes, and your correction is the record from then on. The mapping is versioned per company, so a later month reuses your decision rather than re-proposing the original one.
What happens if a check fails?
It depends whose fault it is. If the data is the cause — a trial balance that does not balance, a missing month, a ledger nothing maps to — the workbook is still delivered, with the failed check named on screen and on its Checks sheet, because a report with a stated caveat is more use mid-close than none. If we are the cause — a balance lost between your file and the report, or a formula that does not reproduce the engine's figure — the job stops, nothing is delivered and its credits go back.
Do I see the price before I am charged?
Yes. Every button that costs credits shows its price before you press it, and pressing it is what holds them. A job that would need more than its standard price stops and shows you a quote first, and nothing is charged unless you accept. There is no path where analysis, mapping results or output appear without a charge being captured or held first.
Read next
- What you get each month: a checked MIS, a dashboard and commentaryWhat lands each month: the movements that matter, explained, inside a checked Excel workbook with live formulas, a dashboard and commentary. Every figure traces to its ledger.
- Security and data handlingWhere your accounting files go: encrypted under each company's own key, redacted before the AI sees any of them, and opened by no one on our side. Stated exactly.
- Credit packs: prepaid, no subscription, credits never expireSix credit packs from Starter to Scale. Buy once and spend credits when you run something: no subscription, no per-seat fee, no minimum, and credits that never expire.
See it on your own month
Create an account, load last month's trial balance, and get the report — paid per action.
Prepaid credits, no subscription. Credits never expire.