The sponsor’s questions arrive on a schedule. The answers should too.
You close several entities against a cadence somebody else set. Covenants get tested whether or not the reporting is ready, add-ons arrive with their own chart of accounts, diligence lands with two weeks’ notice, and the audit committee has started asking what the AI touched and who approved it.
The modules that matter when the entities are subsidiaries
Each one runs a real workflow, in the order it runs. Every step that reaches the ledger stops at a named person.
The close stops being a workbook one person carries and becomes a process with a visible state.
You watch headroom tighten in week two, not on the compliance certificate.
The weekly cash question is answered from the ledger rather than reassembled every Monday.
The model the sponsor reads and the books it describes start from the same numbers.
The next add-on gets onboarded in the same shape as the last one, in days rather than a quarter.
The PBC list stops being an email thread with forty attachments and a spreadsheet nobody owns.
A departure, a vacation or an interim controller stops being a reconstruction project.
Modules you do not need are switched off rather than shipped as clutter. What each one does in your close — your entities, calendar, chart of accounts, materiality thresholds and approvers — is configured during the build, against your books.
One rule holds across every workflow above
The AI retrieves, analyses, drafts and proposes. Deterministic code calculates and checks. An authorised person approves anything that reaches the ledger. There is no unsupervised posting, and no setting that turns it on.
Four-tier action gate
Every action is classified into exactly one side-effect tier. Ledger and customer-facing writes are proposal-first, human approval only. An unrecognised action fails closed into the most-gated tier.
Preparer and approver
Review policy set per entity, with role floors and per-check materiality overrides, and separate preparer and approver roles wherever the control requires segregation.
Readable proposals
Each proposal carries before and after, its source data, risk level, the role required to approve it, a content fingerprint, an expiry, and its execution record.
One audit chain
The prompt, the run, each tool call, the proposal, the approval or rejection, the sync — recorded as one chain you can walk backwards.
Circuit breaker
When a task’s accuracy degrades, the system revokes that task’s autonomy and alerts. It can only demote. It can never promote itself.
Grounded answers
Answers cite period, basis, source reference and sync timestamp. Prose claims are verified against facts captured at the tool boundary; anything unverified is marked.
Four rungs. You can stop after any of them.
The Run is monthly; nothing else here is. Nothing renews against your will, nothing commits you to the next rung, and each one ends with something written down that you keep — the last of them the source itself.
The Close Teardown
Not a demo and not a discovery call. You bring your close — the checklist, the calendar, the entity you argue with most — and we go through it together.
The Pilot
Pick the process that costs you the most every month — the intercompany reconciliation, the flux commentary, the board pack, the covenant certificate. We run it. Not a report about running it: the thing itself, on your calendar, against your books, delivered in the month you would have delivered it anyway.
The Run
We keep running it, then the second process, then the third. An evidence pack every close — what ran, what it proposed, who approved it, what changed. The second process costs less than the first, because the control layer it needs is already standing.
The Handover
After a year of it running, your team takes it in‑house — the source, the training, and support while they settle into it. This is the plan from the first day, not the thing the contract is written to avoid.
If you would rather own it outright from the start, we build it that way instead — scoped against what your close actually is, so it carries no published price. Most people do not start there. The teardown is where you find out whether you should.
Find the layer that is least ready to carry it.
Twelve questions across six control layers. About six minutes, and the result stays on your screen — sharing it is optional.