Why Compliance Cockpit
Built to be checked, not taken on faith.
Compliance work is judged twice: once when you do it, and again — sometimes years later — when somebody asks why. Almost every tool a practice uses is built for the first occasion only. This page is the argument for building for the second.
It ends with the practices this is not for, because naming a non-fit is the only part of a page like this you have any reason to believe.
The problem
Three problems, and they are not the same problem.
Practices rarely describe compliance as one difficulty. They describe three, and a tool that solves only the first makes the other two slightly worse by adding a place to look.
The work is spread across everything
A deadline tracker in a spreadsheet. Companies House open in four tabs. A shared drive of Word minutes. A single-point tool for one job, and a diary in somebody’s head for the rest.
Each is defensible on its own. Together they lose the answer to the only question a Monday actually asks — what needs a person today, and who is carrying too much of it.
Ownership is the part nobody can prove
The share register you inherit is a PDF, a spreadsheet and somebody’s recollection. The public filings that would prove it are spread across twenty years of incorporation documents, allotment returns, transfers and confirmation statements.
They do not always agree with each other. Reconciling them by hand is a day’s work per company, so it is done when something forces it — which is usually the worst moment to find out.
AI arrives, and the profession does not trust it
That distrust is rational. A figure with no working behind it cannot be defended to a client, an auditor or a court, and confidence expressed as a percentage with no arithmetic under it is a mood, not a measure.
A profession that remains answerable for its output cannot adopt a tool whose output it is unable to check. The problem is not that the software is clever. It is that it will not show you how it got there.
The argument
Four decisions. Everything else follows from them.
These are not preferences about interface design. Each one closes off a way the software could have been faster, and each one is the reason the output survives being questioned.
01
Explainable, rather than a black box
The canonical ownership engine is deterministic. Filings are read by parsers and folded by arithmetic; no language model decides what a document said, so the same filings produce the same answer today and in two years.
Confidence is scored the same way. The score starts at 100 and only ever deducts, and every deduction is named and priced: −40 where a register allocates more shares than were issued, −25 where the share totals contradict the evidence, −18 where filings disagree with each other, −15 where issued shares are attributed to nobody. Ninety and above is high, 70 to 89 medium, below that low. The explanation is assembled from the same facts that produced the number, so the reasoning and the score cannot drift apart.
02
Every figure carries its evidence
A holding is not a number in a cell. It is traced back through the filings that produced it — what changed, on what date, from which document, and the running total after it — with the evidence snippet the parser read and a link to the filing itself.
The same rule applies outside ownership. A statutory deadline quotes the section it comes from rather than being typed in. A register change writes its dated history entry in the same database transaction as the change, so a register that moved without a trail is not possible by construction. Where a document would have to assert a legal fact the records cannot support, it prints a bracketed instruction to complete before signing instead of asserting it.
03
The engine stops rather than guessing
Four conditions block approval outright: there is no reconstructed register to conclude from; a share class allocates more shares than were issued; a capital movement contradicts the recorded balance; or the live register has diverged from the reconstruction that was applied to it.
Where a class is over-allocated, the percentages are self-contradictory, so no percentage is shown at all. That is the decision the whole product rests on. A system that quietly picks the more likely of two contradictory filings is not saving you the work — it is moving the error somewhere you will not find it.
04
Prepared to the boundary, and no further
Compliance Cockpit prepares filings for twelve Companies House forms: the whole field set, marked field by field as one your records supply or one only a person can, the statutory deadline, the supporting documents, and a sealed content-hashed package. Then it stops, and somebody at your practice submits it.
There is no code path in this build that transmits a filing to Companies House. Every method on the transport client throws and opens no socket, which is a boundary you can check in about a minute rather than a promise you have to accept. It is a design decision, not a feature waiting to be finished.
0
filings transmitted to Companies House
Every method on the filing transport client throws and opens no socket. Checkable in a minute: search the directory for a network call.
100
where a confidence score starts
It only ever deducts, and only for doubt or contradiction — never for a look-through still running or a filing not yet read.
4
conditions that block approval
No register to conclude from; an over-allocated class; a capital movement that contradicts the balance; a live register that has diverged.
One thing we are not going to oversell
The extraction architecture supports a second, independent reader — Google Document AI — validating a filing in the background as a cross-check that never overrides your record. It is off by default, it has no screen in the product today, and it is switched on by whoever runs your deployment rather than by a setting in an interface.
That is architecture, and we would rather describe it as architecture than let you buy it as a feature and discover the difference afterwards.
The exception
Told to you by us, rather than found by you.
The Companies House boundary is in the code, and that is the strongest form it can take. Making Tax Digital for VAT is not the same guarantee, and it would be dishonest to let the first sentence cover the second.
A real HMRC VAT submission client exists. Whether it can reach HMRC at all is decided by your deployment: it needs production recognition, a stored authorisation and complete fraud prevention headers, and without them the connection resolves to not connected and nothing can be sent. In sandbox, sample data is generated locally and the screens label it as such. Which mode you are in is named on the screen you are looking at.
Pilot deployments ship without HMRC credentials at all. We put this on the page that argues for trusting us, because a caveat you have to go looking for is worth nothing.
In practice
What the argument is worth on a Tuesday afternoon.
Principles are cheap to write. These are the six moments where this way of building the software produces a different afternoon from the alternative.
A client asks why they hold 40 per cent
You open the derivation rather than the correspondence file. Each movement is listed in date order with the filing behind it, aliases merged so one person is counted once, and the running total after every step.
You take on a company whose register has no history
Record the opening position: the register as it stood on the day you took the company on, with the evidence it rests on named and an approver against it. Nothing about the live rows changes. What changes is that “as at” works from that date forward.
Two Companies House filings disagree
The difference is flagged, dated and explained in filing order, and it holds the workflow until somebody at your practice decides what is correct. A genuine disagreement between two filings made on different dates reads as an explanation, not a bare warning.
A year later, somebody asks what the client actually agreed to
The approval captured a hash of the exact documents the client was shown at the moment it was requested. Regenerate the pack and you either get the same content or proof that it changed — rather than an email saying “yes, go ahead”.
A member of staff must not see a particular client
The restriction is a query fragment composed into every read — the client list, companies, deadlines, search, the company secretarial worklist, the document packs — rather than rows fetched and then hidden. A restricted list says why it is restricted.
The partner asks what is actually at risk this week
Deadlines are computed from the record rather than maintained by hand, priority is derived from the date, and every count on the screen is the same query as the list it opens. The number and the page behind it cannot disagree.
Fit
Who this is for, and who it is not.
Compliance Cockpit is opinionated, and an opinionated product is wrong for some practices by design. Here is how to tell which one you are before you spend an hour on a demonstration.
A good fit if
- You look after UK-registered companies for clients, and the statutory record is your responsibility rather than theirs.
- More than one person touches the same company, and you need to be able to say who did what and when.
- You have inherited registers that are correct and have nothing behind them explaining how they got that way.
- Ownership questions — reconstructing a share register, reconciling a PSC position, seeing through a chain of holding companies — cost you real hours.
- You would rather be shown a disagreement between two filings than handed a tidy answer that hides it.
- You expect to be asked, one day, where a figure came from.
Not for you if
- You want submission to happen without a person in the loop. A member of your practice takes the prepared package to Companies House, every time. That is deliberate and it is not on a roadmap; if it is a dealbreaker for you, it will stay one.
- Your book has no UK-registered companies in it. Deadlines here quote Companies Act 2006 sections and ownership is rebuilt from Companies House filings, so a practice of sole traders and unincorporated businesses would be paying for machinery it never uses.
- You want the software to be the authority. A reconstruction from public filings is a proposal to check, never the answer — Companies House publishes filings, not a complete current register of members. If the appeal of AI is that something else decides so you do not have to, this is the wrong product and we would rather say so now than in month three.
- You are a single company looking after your own affairs. This is built around a practice holding a portfolio: practice roles, worklists across every client, portfolio deadlines and a client-visibility wall. One company would be buying a portfolio system to hold one record.
If two of those four describe your practice, please do not book a demonstration. We would rather lose the hour than sell you the wrong thing.
Test the argument against your own portfolio.
Half an hour, no slides. Bring a company whose register you have never been able to prove, and we will show you the parts that do not work yet as well as the parts that do.