Data & AI

How We Handle Your Data & AI

Effective date: 8 September 2026 Last updated: 13 September 2026 Version: 2.3

This is a plain-English explainer. It sits alongside the Privacy Policy, the Terms of Use, and the Data Processing Addendum, which are the governing documents. Where this explainer and those documents differ, those documents govern. We have written this page so that a subcontractor, a security reviewer, or one of the platforms you connect can understand exactly what YourCA does with your data in a few minutes.

#The one-line version

We read your sources and cite them back. We do not train on your data. We do not sign or serve a document for you.

YourCA is one product, built for the subcontractor. Every module sits inside it: Contract, Financials, Procurement, Handover, Prestart, Schedule, Directory and the rest. It reads your own contract and project records, and the systems you already use, and turns them into checked, cited, ready-to-sign work.

#What we do, and what we never do

We do:

Offer connections to the systems you already use. Where we offer one and you connect it, we use the OAuth permissions you grant, and only those, for as long as it stays connected.
Extract, chunk and index your uploads so we can find the relevant passage when you ask.
Retrieve that passage and cite it back to you at the moment you ask, inside your own account.
Draft the real, complete document for you to check, edit and sign yourself.

We never:

Use your data, or any output, embedding or index built from it, to train, retrain, fine-tune or benchmark any AI model. See Our never-train commitment, and exactly what backs it for what that commitment is and is not.
Mix one account's data into another account's results.
Sign a document for you, or serve one on a head contractor for you.
Connect to, sign into, poll or watch a mailbox, a calendar or a consumer file store. YourCA offers no connection of that kind at all, so nothing those platforms hold ever reaches it.
Write anything back into a connected system unless you run the action that does it. Nothing in the Service writes to a connected system on its own initiative.
Give legal, financial or other professional advice. YourCA flags commercial risk for you and your own adviser to check. It is not advice and it does not replace your solicitor. See clause 14.5 of the Terms of Use.

#Retrieval, not training

There are two very different ways an AI product can use your data, and the difference is the whole point.

Training means feeding your data into a model so the model's weights change and it keeps your data permanently, for everyone. We do not do this.
Retrieval at inference means that, at the moment you ask, the system looks up the relevant passages in your indexed data and supplies them to the model as context for that one answer, then cites them. This is how YourCA works. It is commonly called retrieval-augmented generation.

Your data is context for your request, not material for our models. Your index exists only to serve your own account, and disconnecting a source stops its use.

#Our never-train commitment, and exactly what backs it

We want to be precise about the nature of this commitment, because an overstated technical guarantee would be the worst thing on this page.

What it is. A contractual undertaking. We do not train, fine-tune or customise any model on your data, and we do not run any training or model-customisation code at all. Our AI provider's commercial terms also prohibit training on our inputs and outputs. That contract, plus our own, is what backs the commitment.

What it is not. It is not a technical control that YourCA itself enforces on each request. We do not set a zero-retention header, a no-training flag or a per-request opt-out on the model API, because we do not build or run models of our own to opt out of.

What is architectural fact. Your data reaches a model only as retrieved context attached to one request. Nothing about your account is written into a model. Your embeddings are computed in Sydney and stored in your own account's index.

#What we read from each source, and how we use it

There are two connections in the whole product, and they are both accounting software. This table is the complete list. It describes what a connection does where you connect it, and for as long as it stays connected. A connection is authorised by you, uses only the OAuth scopes you grant, and can be revoked by you at any time. None of it is ever used to train a model. There is no connection to a mailbox, a calendar or a file store, and there is no setting, plan or entitlement that adds one.

Source	What we read, where you connect it and authorise it	How we use it	Used to train AI?
Xero	Your contact list, your account codes and your tax rates, then the invoices and credit notes YourCA itself created and the payments applied to them	Fill the builder and the account in once, and show you on the claim when you were paid	Never
MYOB	Your customer list, your accounts and your tax codes, then the sales orders and invoices YourCA itself created and the payments applied to them	Fill the builder and the account in once, and show you on the claim when you were paid	Never

Xero and MYOB are accounting connections, and they read no documents. Neither one is a source of drawings, contracts or correspondence: nothing in the Service opens a file held in Xero or MYOB, and nothing either of them holds is indexed, searched or cited as a document. You connect one of the two, never both, because a payment claim belongs in one set of books.

#Connections are not read-only, and here is the truth about that

An earlier version of this page said every connection was read-only. That was wrong, and we have cut it rather than soften it.

It also described writing into a mailbox or a file store: sending, drafting or deleting mail, setting up a mail rule, creating or cancelling calendar events, and uploading, moving, renaming, sharing or deleting files. None of that exists. It went when the mail and file connections went, and there is nothing left in the Service that can do it.

What a connection writes is one accounting document, and nothing else. When you serve a payment claim, YourCA writes one document into your own accounting system and nothing else: a draft invoice in Xero, or a sales order in MYOB. If the builder later schedules less than you claimed, one more press writes a draft credit note in Xero, or a credit in MYOB, for the shortfall. That is the whole of what is written. Nothing is approved, nothing is posted to your ledger and nothing is issued: a Xero draft cannot be emailed by Xero at all, a MYOB sales order has no email on any layout, and turning either of them into an invoice your builder receives is your own press, in your own accounting system. YourCA never changes a document you edited yourself and never deletes one, except where a document it has just created does not add up to your claim to the cent, in which case it voids or deletes that document rather than leave a wrong figure in your books. It reads back only the documents it created and the payments applied to them.

What is true of any connection you have, and what you can rely on:

Every connection is authorised by you, with the scopes you grant, and revocable by you at any time.
While a source is connected, an action that writes to it runs when you run it. Nothing writes without your press. YourCA once described a capability that chased a supplier for you by sending from your own connected mailbox: that capability has been removed and there is no mailbox to send from. What Procurement does instead is what the screen says: the reminder arrives with the chase already written, and you send it from your own mail client.
Some actions we consider hard to undo carry a warning before you run them. That warning is not applied to every destructive action, so do not treat its absence as a sign that an action is reversible.
If you do not want write access at all, do not connect that source, or restrict what YourCA may do from that platform's own permission settings and revoke the connection there.

#Where your data lives

In Australia. Your Customer Data rests in Sydney. The database and your search index are Neon Postgres in ap-southeast-2. Your uploaded files sit in Amazon S3 in ap-southeast-2. Scanned documents are read by Amazon Textract in ap-southeast-2. Embeddings are computed by Amazon Bedrock Titan in ap-southeast-2. Three of those four are guarded in code: for object storage, OCR and embeddings, if the region is not ap-southeast-2 the component throws an error instead of running offshore. The guards differ in when they bite. The embedding client throws as soon as the application loads. Object storage throws the first time it is used, and only where a bucket is configured. OCR throws the first time it is used, and only where OCR is switched on. The database region is fixed by the Neon endpoint we connect to, not by a check in our code. Application hosting (Replit) and the job queue (Upstash) are deployed to Sydney by configuration and are not region-checked at all, so treat those two as a commitment rather than an enforced control.

What leaves Australia. Three things, and you should read them plainly rather than assume everything stays onshore.

AI text inference runs in the United States. Chat, extraction and contract review are served by the Anthropic API in the US, under Anthropic's commercial terms, which prohibit training on inputs and outputs and provide only limited, short-term retention for safety and abuse monitoring. Where this is a cross-border disclosure, we apply the safeguards in the Privacy Policy and, where the GDPR or UK GDPR applies, the transfer mechanisms described in the DPA.
Telemetry goes to the European Union, from our servers and from your browser alike. Usage and error events go to a PostHog project in the EU. Both paths carry the same region check: a PostHog United States host is refused, a host we cannot read is refused, and where the host is refused the telemetry is switched off entirely rather than sent offshore. If no host is configured for the browser, your browser sends no telemetry at all. Two limits worth knowing. Events raised by the pages you load are not put through our scrub and carry the network address the request comes from, so your IP address reaches the European host. And the check stops PostHog's United States cloud specifically; it does not prove that whatever host we do configure sits in Australia or the EU.
An email you forward to a job rests offshore before we pull it into Sydney. A job can be given a capture address that you forward the builder's email to. Our email provider, Resend, receives that message on our behalf and holds it until we fetch it. Resend has no Australian region. It offers North Virginia, Ireland, Sao Paulo and Tokyo, so the message rests outside Australia from the moment your mail server delivers it. What rests there is the whole of what you forwarded: the sender, recipient and subject headers, the message body, and every file attached to it, which in practice means directions, payment schedules, RFI answers and drawings. We copy it into Sydney as soon as we are told it arrived, and from then on our copy lives in ap-southeast-2 with everything else. We cannot delete their copy. Resend gives us no delete for a received message, so how long it stays offshore is their retention period and not ours. This applies only to what you forward. It is off unless you turn it on, one job at a time, by creating an address and using it, and a document you upload yourself never leaves Australia.

Payment processing (Stripe) and email (Resend) are global services and are not pinned to a region. Resend carries email in both directions: the messages we send you, and the messages you forward to a job's capture address, as set out in item 3 above. Card details are entered directly with Stripe and raw card data never reaches YourCA. The full sub-processor list, which names every provider we engage and is complete, is in the DPA, Annex 3, and clause 6.1 of the Privacy Policy. There is no separate error-tracking provider: the analytics provider is also the only error-capture channel we operate.

#How YourCA reads a document and cites it

Extract. Your upload is read. Scanned pages and images go through OCR in Sydney.
Chunk. The text is split into passages, keeping clause structure so a citation can point at a real clause rather than a page.
Embed. Each passage is turned into a numeric index entry in Sydney, stored against your account only.
Retrieve. When you ask something, the passages that actually bear on your question are pulled and handed to the model as context for that one request.
Cite, or say nothing. A value we derive from your contract carries the clause reference and a verbatim quote from that clause. If a value cannot be cited, it is left blank rather than estimated, never shown as a plausible-looking figure. This is enforced in code, not left to the model: an uncited value is stripped before you ever see it.

A blank you have to fill in yourself is honest. A confident wrong number on a document you are about to sign is not.

#Statutory documents: we draft, you sign and serve

Payment claims, statutory declarations, notices and compliance certificates are drafted by YourCA and exported to you. Five rules govern every one of them:

You get the real, complete document, filled from your own contract and project records. Not a worksheet, not a checklist.
Every field stays editable until you sign it. You are the one signing, so you must be able to correct anything on it.
The signature block is left blank, for your wet or applied signature. YourCA never signs. Where a witness is required, the witness block is blank too and the document says so on its face.
We export, and stop. You download it, check it, sign it and serve it yourself through whatever channel you already use.
Where a form is prescribed by law, you get that exact form, filled, or nothing at all. We will not print a redrawn lookalike of a prescribed form.

YourCA cannot transmit one of these documents to a head contractor. Serving on your behalf is not merely switched off, it is not built, and there is no environment setting, account flag or operator override that turns it on. Recording that you served a claim is a diary entry you make afterwards. It moves no document anywhere.

Every generated document carries this statement on its face: "YourCA prepared this draft from your own contract and project records. It is not legal advice. Check every figure and every statement on it, then sign and send it yourself."

#What we secure, and what we do not claim

YourCA is run by one person. We would rather you know exactly what that means than discover it in a security review.

In place: TLS 1.2 or higher in transit; encryption at rest through our managed hosting providers; your integration credentials encrypted with AES-256-GCM before storage; role-based access and tenant isolation so one account's data never reaches another; optional multi-factor authentication; rate limiting, security headers and server-side input validation; audit logging of administrative activity; and automated typecheck, test and build checks on every change proposed for, or merged into, the main branch of our codebase before release.

Not in place, and we will not pretend otherwise: no SOC 2, ISO 27001 or IRAP certification; no third-party penetration test; no security team, security operations centre or continuous monitoring; no intrusion detection or web application firewall of our own; no automated vulnerability or dependency scanning; no malware or antivirus scanning of any file that reaches us, whether you uploaded it or forwarded it in on an email; no documented and tested incident response or disaster recovery plan; and no customer-managed encryption keys. The full list is in the DPA, Annex 2.

What we do instead on every file that reaches us, stated so you can judge it. We accept only the formats we can read, we cap the size and the number of files, and we then read the leading bytes of each file and refuse it if they are not the type it claims to be. That stops a program, an archive or a web page being stored under a document name. It does not stop a genuine PDF or Office file that carries something harmful inside it, because nothing here opens a file looking for that. What we store is served back to you as a download, under the type its extension promises rather than the type the file claimed, and only images and PDFs open in place.

The same checks, and the same limit, apply to a file attached to an email you forward to a job. An attachment gets the identical type and leading-byte verification, is refused on the identical grounds, and is never scanned for malware either. Two things are worth knowing about that path specifically. The sender shown on an arrival is what the message claimed about itself, and forwarding weakens the checks that would ordinarily corroborate it, so it is not proof of who sent anything. And nothing that arrives that way is applied to a claim, a variation or any other register on its own: YourCA suggests where it belongs and a person presses. Treat a document that reaches you through YourCA from a party you do not know exactly as you would treat it in your email, because on the forwarding path it did reach you through your email.

We do not hold professional indemnity insurance. If your own obligations require a control or a cover we do not have, tell us before you upload anything, and do not send us data that depends on it.

#What you can recover from us if something goes wrong

We would rather you read this now than find it in clause 15 after a bad day. YourCA is one person, trading as a sole trader, with no company between a claim and that individual and no professional indemnity or cyber insurance behind it. We have set our liability as low as the law allows, and the whole of it is capped at the greater of A$1,000 and the fees you have actually paid us in the 12 months before the problem arose. There is a minimum of A$1,000, so the cap is never nil, even if you have paid us nothing because you were on a free pilot or a code that reduced every fee to zero. A$1,000 is a small amount and we are not presenting it as a real remedy for a large loss: it is there so the limit is never nothing. If a single A$249 contract review is all you have paid, A$1,000 is the most you can recover, however large your own loss. Every payment we could owe you draws on that same amount, including any refund and anything we pay in connection with an intellectual property claim, and we no longer give the uncapped intellectual property indemnity an earlier draft of the Terms contained.

Four things that cap does not touch. A wilful breach of confidence by us, which includes a deliberate unauthorised disclosure of your data, is outside the cap, and the DPA says the same for your personal information rather than pulling it back inside. An infringement by us of your intellectual property is outside it as well, and so is an infringement by you of ours: those carve-outs run both ways, as does every other item on the list in clause 15.3. Your consumer guarantees under the Australian Consumer Law cannot be excluded, restricted or modified, and we do not try to: most business customers buying at our prices are "consumers" for that purpose, because the amount paid does not exceed A$100,000. For a breach of one of those guarantees, the law lets us limit our liability to supplying the Service again or paying the cost of having it supplied again, which is what we do, and a court can refuse to let us rely on that limit if it decides doing so is not fair and reasonable. And nothing we write excludes liability for fraud or for death or personal injury.

Two more things we would rather you read here than find later. We give no warranty that the Service matches our Documentation. The only warranties in the Terms are that we use reasonable care and skill and do not knowingly introduce malicious code; the Documentation says how the Service is meant to work, not what we promise it does (clauses 14.2 and 14.3, and your consumer guarantees still apply on top). And your own liability to us is capped as well, at the greater of A$10,000 and five times the fees you have paid us in the preceding 12 months, with only the fees you owe us sitting outside that (clause 16.2). If we agree a pilot with you, our cap for that pilot is the greater of A$1,000 and the fees you paid for the pilot, so an unpaid pilot carries an A$1,000 cap rather than nil, and that measure is narrower than the 12-month figure above (clause 8).

Decide before you buy whether a remedy that size is enough for what you intend to put through YourCA. If it is not, use your own advisers and your own cover for that work. The full position is clause 15 of the Terms of Use.

#Your controls

You choose the scopes. Where you connect a source, that connection can only reach what the OAuth permissions you granted allow.
Disconnect any time. Removing a connection stops further reading from that source.
You run the writes. While a source is connected, any action that changes something in it runs when you run it. There is no capability left that writes or sends on its own. Some actions we consider hard to undo carry a warning before you run them, but not every destructive action does.
AI runs when you run it. AI processing happens when you start a feature that uses it, so you control it by choosing what to run. No scheduled job reads anybody's mail. YourCA once described one that would read a connected mailbox overnight to prepare a brief and draft a chase; it has been removed, along with the connection it needed. There is no account-wide off switch. If you need AI processing restricted for your account, email admin@yourca.ai.
You sign, and you serve. No document is signed or sent for you.
Export and delete. You can export Customer Data while your account is active and request deletion as set out in the Terms and DPA.

#We complement your tools; we don't compete with them

YourCA reads across the systems you already use rather than replacing any of them. Keep your inbox and your accounting exactly as they are. We do not use a connected source's data to build a competing replica of that source's features. Our value is contract intelligence and the statutory layer on top of the records you already hold, not a wrapper around any single platform.

#How this maps to each platform's rules

These are undertakings we give for each connection, and they apply where and when we offer that connection and for as long as it stays connected. They are how we will handle the data, not a claim that we hold any particular platform's approval, verification or certification today. Where a platform requires something of us before a connection may be offered, we will have it before we offer that connection, not after.

Xero: where we offer a Xero connection, we comply with the Xero Developer Platform terms.
MYOB: where we offer a MYOB connection, we comply with MYOB's developer terms for its Business API.
The mail and file platforms: there is nothing to undertake, because there is no connection. YourCA does not connect to, sign into, poll or watch a mailbox, a calendar or a consumer file store, so no data from one of those platforms reaches us and none of their developer programmes applies to us. Earlier versions of this page gave undertakings for those platforms, including the annual independent security assessment a mail provider requires before an app may read mail in production. We removed the connections instead of obtaining the assessment, so those undertakings have gone with them.

Our own governing documents put the same commitment in contract: see the Privacy Policy, the Terms of Use, and the DPA, Annex 2 (AI-specific measures). What YourCA costs is set out on our home page and in your Order.

#Questions

Privacy and your data rights: admin@yourca.ai
Security and vulnerability reports: admin@yourca.ai
Commercial and contract questions: admin@yourca.ai

Michael Dewick trading as YourCA (ABN 84 390 063 197) · Sydney NSW, Australia