Liam Killingback
8 August 2026
AI document extraction turns an uploaded file into structured matter data: parties, dates, amounts and obligations read out of the text and written into fields a person can check. MatterFirst, a legal practice management platform for Australian law firms, shows each extracted value with the snippet it was read from, so a fee earner verifies rather than trusts.
Extraction is not the same thing as generation
Three different jobs get sold under the word AI, and they carry different risks. Extraction is useful to a law firm in proportion to how quickly someone can confirm it.
Extraction reads a document you already hold and returns values that are in it. The settlement date on page four is either 14 October or it is not. The failure mode is a wrong or missing value, checkable against the source in seconds.
Generation produces text that was not there before: a letter, a clause, a summary. The failure mode is confident invention, and checking it requires knowing the answer already.
Retrieval and chat answer questions across a matter, grounded in your files but phrased by the model.
The distinction matters legally, not only technically. The Supreme Court of New South Wales Practice Note SC Gen 23, which commenced on 3 February 2025, restricts generative AI in affidavits, witness statements and expert reports, while leaving routine preparation such as summaries, indexes and chronologies open provided citations are checked by hand. Reading a contract you were sent and filing its dates is a long way from generating evidentiary content, and a firm that treats the two as one policy will either ban something safe or permit something it should not.
Queensland Law Society Guidance Statement No. 37, published 31 May 2024, puts the obligation plainly: "Each piece of discrete work produced using an AI system (or any other production method) should be checked." Extraction is the AI task that makes that duty cheapest to discharge, because the evidence sits on the same screen.
The five steps between an upload and a matter field
Vendors compress this into one animation. It is five steps, and each fails differently.
| Step | What happens | How it fails | What to test in a trial |
|---|---|---|---|
| 1. Capture and read | OCR and layout detection turn a PDF or scan into text, tables and form fields | Poor scans, handwriting, stamps over text, multi-column layouts | Upload your worst scanned rates notice, not a clean sample |
| 2. Classify | The system decides what kind of document this is | A variation is filed as the wrong type, so the wrong rules run | Upload three versions of the same document type from different sources |
| 3. Extract | Values are pulled against a schema of expected fields | A field is missed, or a plausible wrong value is returned | Check a document where the answer appears twice with different meanings |
| 4. Confidence and provenance | Each value is returned with its source location, and a confidence score where the model gives one | The value is shown with no way to see where it came from | Click a value. If you cannot reach the source text, the check is manual |
| 5. Write and trigger | Values populate matter fields, key dates, contacts and workflow steps | Silent overwriting of data a person entered | Extract onto a matter that already has data and see what it does |
Step four is where products separate. Every vendor can produce a list of extracted terms, and far fewer show the sentence each term came from. That one feature decides whether checking a ten page contract takes two minutes or twenty. Step five decides whether you save anything at all: extraction that produces a report is a research tool, while extraction that fills the matter and sets the key date is what removes the re-keying.
What extracts reliably, and what still needs a person
Extraction is strongest where a document type recurs with stable structure, which is why property and commercial practice see the benefit first.
| Document | Usually extracts cleanly | Still a judgement call |
|---|---|---|
| Contract of sale | Parties, purchase price, deposit, settlement date, property address, lot and plan | Whether a special condition is unusual for this transaction |
| Rates and water notices | Period, amount, assessment number, due date | Which adjustments apply and to whom |
| Lease | Term, commencement, rent, review dates, option periods | Whether the review mechanism operates as the parties intended |
| Loan and mortgage documents | Lender, facility amount, borrower, key dates | Whether a condition precedent is satisfied |
| Correspondence | Sender, date, matter reference | The concession buried in paragraph three |
The pattern: extraction is reliable on values that exist in the document and weak on conclusions that require knowing the transaction. A vendor claiming the second should be asked to demonstrate it on your documents rather than theirs.
Three questions to settle before you upload a client file
Where does the inference run? Storage location and processing location are different questions, and the second is more often left unanswered. QLS Guidance Statement No. 37 lists "Where is the data being processed?" among the questions a practitioner should ask. Clio's own help centre states that "Depending on where your firm is located, AI features in Clio may process your queries on servers located outside your home jurisdiction. After your request is processed, any resulting data remains securely stored within your region." That is the kind of statement to look for from every vendor, including the one you already use. Actionstep's security page says firms "have the option to store their firm's data in any one of our four AWS data centers in Australia, Europe, Singapore, or the United States", which answers storage without answering inference.
Is the output traceable? If a value cannot be traced to the text it came from, supervision turns back into re-reading the document.
Who checks, and is the check recorded? Practice Note SC Gen 23 and Guidance Statement No. 37 both put the responsibility on the practitioner, not the vendor. A workflow that routes extraction results to a named person, and records that they accepted them, is what turns a compliance position into evidence. The OAIC's guidance on privacy and the use of commercially available AI products is the companion reading on the privacy side.
What it should cost, and whether you can bill it
Two models exist. Some vendors bundle AI into the subscription, some meter it. Bundling is simpler to budget and gives you no per matter record. Metering takes more attention and produces exactly the record a costs discussion needs.
This is not an abstract preference. Guidance Statement No. 37 says that AI licence fees "charged on a 'per user', lump sum or annual license basis are a practice overhead that cannot be billed to a client", while "an AI license fee (or other software) charged on a 'per matter' basis can fit the usual definition of 'disbursement' more easily". Whether a particular charge is recoverable depends on your costs agreement and your client. A firm that wants the option needs a system that records AI cost against the matter in the first place.
MatterFirst publishes its rates on the AI pricing page: 60c to open a document plus 9c per page, and 15c per document per question for bulk review, ex GST and fixed until 1 August 2028. A one page rates notice costs 69c and a ten page contract costs A$1.50. Chat, drafting and summarising the document in front of you draw nothing from the balance. Above A$25.00 of work the system quotes the total before it runs.
LEAP takes the other approach, stating on its Matter AI page that "Matter AI is included in every LEAP subscription at no extra cost". Smokeball's AI page describes Archie as "available as an add-on to both Grow and Prosper" without publishing a rate. Neither approach is wrong, and the disbursement question is what decides which one suits your firm.
How MatterFirst handles this
MatterFirst is a legal practice management platform for Australian law firms, built by North Cape Technology in Melbourne. Document AI accepts PDF, DOCX and scanned images, and extracts parties, dates, obligations and financial terms into structured fields on the matter. Each key term is shown with the text it was read from, and its confidence where the model returned one.
Firms define what gets extracted. The Key Term Extractor lets a firm set its own terms per document category, such as property address, trust account number or a licence reference, and map each to a matter field, a contact role or a key date. Extraction is configurable per matter type rather than being a single fixed schema.
Extracted data drives the matter rather than sitting in a report. A party found in a document is matched against your contacts, created if new, and assigned to the matter with the right role. A contract deadline can fill an existing key date. A document upload can trigger an automation scoped by matter type. Issue flagging assigns a severity of info, warning or error and is enabled per document category, and the risk summary is never shown in the client portal.
For volume work, bulk review projects run the same questions across hundreds of files and return an exportable grid. On where processing runs, MatterFirst publishes the split: reading the document runs on Azure AI Document Intelligence in Azure's Australia East region, and analysing what was read runs on AWS Bedrock in the AWS region your firm chose, Sydney by default. Both are in Australia. Every AI charge records its matter, and that record exports as a statement formatted for a disbursements ledger.
Plans start at $199 per month per workspace with users included, and every paid plan carries a monthly AI balance in Australian dollars: A$100 on Solo, A$280 on Practice, A$600 on Firm. See pricing for the current plan detail.
How the platforms compare on document extraction
Verified on 8 August 2026 from each vendor's own public pages. "Not published" means the vendor's page checked does not state it, not that the product cannot do it.
| MatterFirst | LEAP | Clio | Smokeball | Actionstep | |
|---|---|---|---|---|---|
| Extracted values written into matter fields | Yes, into fields, contacts and key dates | Not published | Dates extracted from documents can be turned into scheduled events | Not published as a general feature | Not published |
| Firm-defined extraction fields | Yes, Key Term Extractor, per document category | Not published | Not published | Not published | Not published |
| Each value shown with its source text | Yes, with confidence where returned | Answers include references | Summaries include inline linked citations | Not published | Not published |
| Same questions across many documents at once | Yes, bulk review, exportable grid | Matter AI works across matters of ten to more than 10,000 documents | Document Analysis accepts multiple files or a folder upload | Not published | Not published |
| Where AI inference runs | Azure Australia East for reading, chosen AWS region for analysis | Not published | May process queries outside your jurisdiction; results stored in region | Not published | Not published |
| AI pricing published | Yes, per document and per page | Included at no extra cost | Not published | Archie is an add-on, rate not published | Not published |
Sources: MatterFirst Document AI and security; LEAP Matter AI; Clio help centre on AI features and document summarisation; Smokeball AI; Actionstep AI and security.
Who this suits, and who it does not
Document extraction pays back fastest where the same document shapes recur: conveyancing, property, commercial and leasing work, and any firm processing volumes of third party paper such as rates notices, title searches and contracts of sale. If your firm opens twenty conveyancing files a month, the re-keying you remove shows up in your own time records within a fortnight.
It is worth less to a practice whose documents are mostly bespoke correspondence and advice, where there is no stable schema to extract against, and to a firm that will not move extraction to the point of upload, because running it later as an extra step is where the saving disappears.
MatterFirst suits Australian firms of roughly two to twenty fee earners that want onshore processing, state based trust accounting and document heavy workflows in one system. Firms needing an integration that is not Xero, Stripe, Microsoft 365 or Google Workspace should check the position first: those four are the connectors available today, and the rest of the catalogue is listed as coming soon. The evaluation checklist collects the questions worth putting to every vendor on your shortlist, including MatterFirst.
FAQ
Does AI document extraction actually save time, or just move the work? It moves the work from re-keying to checking, and checking is faster only if each value is traceable to its source. Ask any vendor to click through from an extracted date to the sentence it came from.
Can I use AI to read a client's contract without breaching confidentiality? Queensland Law Society Guidance Statement No. 37 says practitioners should take reasonable steps to ensure an AI tool "does not misuse access to confidential data or unduly expose it to dissemination", and lists where the data is processed as one of the questions to ask. Get the processing location, the training position and the retention terms in writing before you upload anything.
Does MatterFirst process documents in Australia? Yes. Reading the document runs on Azure AI Document Intelligence in Azure's Australia East region, and the analysis runs on AWS Bedrock in the AWS region the firm chose, Sydney by default for Australian firms. Both regions are in Australia, and both are stated on the MatterFirst security page.
Can I charge AI costs to the client? Guidance Statement No. 37 treats per user, lump sum and annual AI licence fees as practice overhead that cannot be billed, and says a per matter charge fits the definition of a disbursement more easily. Whether a particular charge is recoverable is between you, your costs agreement and your client. What software can do is record the cost per matter, which MatterFirst exports as a statement formatted for a disbursements ledger.
Can I define which fields get extracted from my own documents? In MatterFirst, yes: the Key Term Extractor lets a firm define terms per document category and map them to matter fields, contact roles or key dates. For other platforms, ask whether the schema is fixed or firm configurable, because the answer is usually not on the marketing page.
What does document extraction cost in MatterFirst? 60c to open a document plus 9c per page, so a one page rates notice is 69c and a ten page contract is A$1.50, and bulk review is 15c per document per question. Those draw on a monthly balance in Australian dollars that is included with every paid plan.
The short version
Extraction reads what is in a document. Generation writes what is not. Buy them separately in your head, even when they arrive in one product.
Test four things: your worst scan, whether every value traces back to its source text, whether extraction writes into the matter or only into a report, and where the inference runs. The practice notes and the ethics guidance already tell you the check is yours to perform. The remaining question for a vendor is how fast they let you perform it.