Commercial lending
Before you can assess the deal, you have to build the picture.
A commercial file arrives as forty documents describing a group nobody has ever drawn. This tool reads the pack and draws it, then tells you what credit will ask.
In development with practising commercial brokers. The demonstration on this page runs on a fictional group.
Every complex file starts the same way.
Financial statements for four entities. Two trust deeds, one varied in 2019. A PPSR search you have not read yet. Three years of tax returns. An application form that describes the structure in two lines and gets one of them wrong.
Before you can form a view on serviceability, you have to work out what the group actually is.
That work is not billable, and it is not optional.
The things that kill a commercial deal are rarely visible in a summary.
Already registered, ranking ahead of the facility you are writing.
Drawn from the company, with no complying agreement on file.
Owed to a corporate beneficiary and never paid.
None of these appear on a balance sheet as debt. All of them change the answer.
Reading the documents is not the difficult part.
Holding four different relationships in your head at once is. Each runs on its own logic, and they rarely point the same way.
Ownership
Who holds the shares and the units.
Control
Who is trustee, and who is appointor. In a discretionary trust, a different question to ownership, and the one most often skipped.
Security
Which financier holds what, over which asset, in what priority.
Guarantee
Who stands behind whom. Often the opposite direction to ownership.
A group with six entities has more relationships than most people can hold without drawing it. That is not a failure of attention. It is arithmetic.
What it does.
-
1
Reads the pack
Statements, returns, trust deeds, ASIC extracts, PPSR searches, loan documents. Including the scans.
-
2
Builds one structure
Every entity and every relationship: ownership, control, security, guarantee, cash flow.
-
3
Draws it
An interactive map. Select an entity and every exposure connected to it traces, with the numbers alongside.
-
4
Reviews it
What needs answering before the file goes to credit, graded, each point naming its source document.
It does not make the call. It makes sure you are making it with the whole picture in front of you.
The Kestrel group is not real.
A fictional commercial plumbing contractor in southern Adelaide: three entities, two directors, a super fund holding the workshop, and six findings. A reasonably healthy file where the issues are fewer and subtler. It reads in about thirty seconds and works on a phone. Every figure is invented.
Two things are worth finding when you open it.
- The rent paid to the super fund.
Above market, which is a compliance problem and an add-back at the same time. The rare case where one correction fixes both.
- The serviceability section.
In this file, both readings of the accounts agree. The review distinguishes between cases rather than applying a rule.
Fictional group, synthetic figures, no sign-up. The demonstration states this again inside.
Want to see the tool under load? The Hartley group is a six entity civil contracting file with eighteen findings, including a security registration that would stop settlement. Open the Hartley group file.
Your client data. The short version.
The examples above use invented data. If you were to run this on a real file, the first question is where that file goes. Here are the answers, including the ones that are less comfortable.
Are the demonstrations using real client data?
No. Every entity, person, and figure in both examples is invented. No document was uploaded and nothing was processed to produce them. They are pre-generated examples of what the output looks like.
The tool is not connected to this website. You cannot upload a file here, which is deliberate.
If I used this on real files, where would my data go?
That depends on how it is set up, and you choose.
Inside your own environment. If your practice already runs an enterprise AI subscription, the tool is installed inside it. Your data stays within the boundary you have already assessed and approved. Nothing new leaves.
Under your own account. If you do not have an enterprise subscription, you supply your own model provider account. Processing runs under your agreement rather than ours. We do not hold your data and we do not see it.
De-identified before it leaves. Where the file must not leave Australia in identifiable form, a de-identification layer runs on your own machine first. Names and identifiers are replaced with placeholders before anything is sent. The list that reverses them stays on your machine.
Most practices choose one of the first two. The third exists for those who need it.
How does de-identification work if the tool is meant to map named entities?
This is the right question, and it is the part that took the most work.
Identifiers are not deleted. They are swapped for consistent placeholders before the file is sent, and swapped back when the result returns. Every variation of a company name resolves to the same placeholder, so the structure stays intact while the identity does not travel with it.
The list that maps placeholders back to real names never leaves your machine. It is held encrypted, only for the length of the job, and destroyed when the job finishes.
Is de-identification absolute?
No, and anyone telling you otherwise is overselling.
The models that do this work are very good. They are not certain. The organisation that publishes the one we use describes it in its own documentation as a data minimisation aid rather than an anonymisation or compliance guarantee, and notes that it can miss uncommon identifiers. We agree with that assessment and would rather you heard it from us first.
A name in an unusual position can be missed. Financial figures in combination can be identifying even without names attached.
De-identification substantially reduces what is exposed. It does not reduce it to zero. If your obligations require certainty rather than reduction, the version that runs inside your own environment is the better answer, because nothing leaves at all.
What does the de-identification actually run on?
An open-weight model published under a permissive licence, running on your own hardware. Not a service, and not a subscription. Nothing is sent anywhere for the de-identification step itself, because the whole point of that step is that it happens before anything is sent.
We did not build the detection model. We built the layer around it that keeps entity names consistent, holds the mapping locally, and reverses it on return.
Because it is open, you or your adviser can inspect exactly what it does. Most vendors cannot offer that.
Who is responsible for privacy compliance, you or me?
You are. You hold the client relationship and the obligations that come with it.
What the tool does is make those obligations easier to meet, by keeping data inside your control or by de-identifying before disclosure. Your privacy policy and collection statements will need to cover AI processing, and that is worth checking with your own adviser rather than taking our word for it.
Is the AI making decisions about my clients?
No.
It builds the picture and flags matters for your attention. Every finding is attributed to the source document it came from, and figures that were inferred rather than read are marked as inferred, so you can see what the tool worked out as opposed to what it found.
The judgement is yours. That is not a disclaimer, it is the design. A tool that made the call for you would be useless, because you would still have to check it.
What if it gets something wrong?
It will, sometimes. Every finding is traceable to a page in a source document precisely so you can check it quickly.
The tool is built to shorten the work of assembling the picture, not to replace the professional reading it. You remain the licensed person making the assessment, and the output is structured to support that rather than substitute for it.
How much does the processing cost?
Very little. Model usage for a full document pack runs to a few dollars.
The cost of this tool is the tool, not the processing.
If your compliance officer has a question that is not answered here, ask it directly. A question we have not thought of is more useful to us than one we have.
Where this is up to.
The underlying tool is in use with a financial adviser. The commercial lending version is being built with practising brokers, because the review logic is only as good as the expertise behind it.
If you write commercial deals and want to see it run against a real file, get in touch. What I am after is your view on what matters when a file goes to credit.
The same engine has other applications, in financial advice and estate planning among them. This page is about lending.
Nothing on this page or in the demonstration is credit advice, financial advice, or legal advice. The tool surfaces what to check. It does not conclude.