A restaurant, a clinic, a hotel and a fuel station can run on the same books — because there is only one set of books. Eleven industry packs post into a single double-entry core, sealed by a hash chain that makes a silent edit impossible.
Two hundred screens across thirteen areas, from the journal entry to the consolidated group statements — every one of them posting into the same book.
One unbroken line: flat while the data rests, a single sharp beat as the entry posts, flat again while the book settles, then a rising staircase of periods compounding into one point of light.
Data at rest. Nothing has been posted yet.
The entry posts — balanced, and stamped into the chain.
The period closes. The book stops moving.
Periods compound, and the number becomes a decision.
Most suites give each department its own database and reconcile them afterwards. Here the point of sale, the payroll run, the depreciation schedule and the hotel's night audit all write the same journal — so the trial balance is never a merge, it is simply the truth.
Every document in every pack resolves to a balanced entry before it posts. There is no path in the system that writes a number without its opposite, which is why the trial balance reads balanced rather than reconciled.
Each posted entry stores the fingerprint of its predecessor, and the chain is verified rather than merely stored. Alter a historic entry and every entry after it stops verifying — tampering becomes arithmetic instead of suspicion.
Cost centres, projects and dimensions cut across every pack, so a report never has to ask which system a figure came from. A closed period rejects writes everywhere at once.
Accounting, restaurant, clinic, pharmacy, salon, HR, school, training, hotel, fuel and car rental. Every one is marked ready in the product registry — none is a placeholder waiting on a roadmap.
The journal is where the platform is honest. Every posting carries the fingerprint of the one before it, so history cannot be quietly rewritten — and the trial balance is computed from what actually posted, never from a cached summary.
The difference line reads balanced because the system has no path to make it otherwise, not because a job reconciled it overnight.
Previous month, previous year or a custom window, filtered by branch, cost centre, project and account level, and drillable down to the entry.
Blockers, warnings, unposted drafts and pending approvals are all counted before a period can close. Reopening is an event with a name attached, not a silent toggle.
The balance sheet, income statement and cash flow are derived from the same posted entries as everything else — which is why they agree with the trial balance without anyone reconciling them.
Balance sheet, income statement and cash flow, alongside the statement of changes in equity and OCI, and the notes that accompany them.
Deferred revenue recognised across the delivery period with its own schedule; group entities consolidated with intercompany elimination, translation reserve and minority interest.
The standard audit file of the general ledger is generated for the auditor rather than extracted for them.
Customers, suppliers, employees and shareholders are one party register with roles, not four disconnected lists — so a supplier who is also a customer nets correctly instead of appearing twice.
Buckets resolve to the invoice, the invoice to the entry, and the entry to its audit trail.
Exposure is scored per party against limits the sales path enforces, rather than a report that notices afterwards.
Sensitive identity fields live in their own table behind stricter permissions than the party record that references them.
Stock movements post to the ledger as they happen, so inventory value on the balance sheet is the same number the warehouse screen shows — not a figure someone transcribed at month end.
Movement-based valuation with landed cost, so margin is computed against what the item actually cost to have rather than its purchase price.
Batch and expiry tracking with FEFO where a pack needs it, and a stocktake that posts its own adjustment entry.
Ranked by contribution rather than volume, computed from the cost centres the entries already carry.
Invoices, returns, vouchers and batch payments are not a layer above the accounts — each one resolves to a journal entry the moment it posts, tagged with the pack, branch and cost centre it came from.
Orders, invoices, landed cost and supplier settlement, with batch payments generating a bank file rather than a list to retype.
Returns computed from posted entries at the statutory rate, in the ZATCA e-invoicing format.
Each browser call carries its product scope, so the database tags the document and its entry at write time — not the interface.
The hard problem with AI over accounting is not fluency — it is that a confident wrong number is worse than no number. Forty-three AI tools ship across the platform and twenty-two are in scope on the accounting core, fourteen of them deterministic: no model call, no API key, and no AI credit consumed. Three mechanisms exist for exactly this problem, and the engineering notes in those files name them as moats.
Every figure the AI states is stamped with the ledger's hash-chain head at the moment it was computed — head, sequence, entry count and a chain-integrity flag.
If the books move afterwards the answer marks itself stale instead of continuing to read as live, and a hallucinated number fails the stamp mechanically.
It does not wait for a human to notice. A broken chain means the figure is not counted at all.
Approval limits, closed periods and role scope are enforced on the write path in the database, not requested of the model in a prompt.
Hiding a button is not isolation — a direct call still lands. Enforcement has to sit where the write does.
As the source file states it: the AI is not required not to violate — it cannot.
Anomaly detection is calibrated against your own history rather than a fixed threshold, using the median and median absolute deviation instead of mean and standard deviation.
Ledger data is skewed and outliers poison an average, so a fixed rule floods a large business with false alarms and stays silent for a small one.
Below eight observations it declines to judge at all, and every result carries how much was examined against how much exists.
The rule these engines share, stated in the source: no checks and no signals means no score — declaring “100% ready” when nothing was examined attests to a void.
These properties hold because of how the data is written, not because of a policy asking people to be careful.
The hash chain is verified rather than merely stored, so a broken chain is detected and reported instead of sitting silently in a column.
Approval limits, closed periods and role scope are enforced where the write lands, mirroring the guard that runs before any data loads.
Product isolation is an explicit field in the tool registry. Matching on URL prefixes breaks silently at the first rename; a declared scope breaks the build.
Tool, cost, scope and outcome are written to an operations ledger that reconciles, so AI spend is auditable like any other cost.
Described from the product's own implementation. No compliance certification is claimed.
Web ERP is the accounting core of the Web Mister AI platform — the book every industry pack posts into. That shared core is exactly why a clinic's appointments can tell a pharmacy what to buy, and why a hotel's night audit lands in the same statements as a fuel station's pump shift.
Web ERPRestaurant & CaféClinicPharmacySalon & SpaHR & PayrollSchool & KGTraining CentreHotel & ApartmentsFuel StationsCar RentalEvery pack listed is marked ready in the platform's product registry. Web ERP is the core described on this page.
The fastest way to judge Web ERP is to follow a single document — raised in a pack, posted to the journal, stamped into the chain, and read back in the statements.
