Method and data note
We built the pipeline. Here is what it does.
The production system behind every service on this site is ours: built for legal deliverables rather than licensed from a vendor and pointed at legal files. Most providers are vague about this. We would rather you had the whole answer before you ask - what the pipeline does, where a lawyer takes it back, where your data sits, and what we never do with it.
The infrastructure
Six things the pipeline does that a general AI product does not.
This is the part of the business we invest in. Everything else on this site - the published prices, the 48-hour standard, the willingness to take a twelve-item job - is downstream of it.
Ingestion and OCR
Scanned records at any quality, native productions, email archives, exports from systems nobody supports any more. Nothing has to be re-keyed and there is no platform for your firm to adopt at the other end.
Classification and near-duplicate clustering
Document typing, email threading and near-duplicate grouping, so the same content is never read twice. On a large production this routinely removes a third of the population before anyone reads a page of it.
Schema-bound extraction
Every deliverable has its own extraction schema: a medical chronology pulls entirely different fields from a privilege log or a lease abstract. The schema is the deliverable's contract, and it is versioned per firm where your format differs from ours.
Source binding
Each extracted fact carries the document and page it came from. Any line in a deliverable can be followed back to the record behind it, which is what makes the output usable in a filing rather than only as a reading aid.
Confidence scoring and exception routing
Everything the pipeline produces carries a confidence score, with the threshold set per service. Below the line, the item goes to a lawyer's queue instead of into your file - and the confidence distribution ships with the work rather than staying with us.
Consistency checks and QC sampling
Document families and duplicate groups are checked so related material cannot be treated inconsistently, a statistical sample is scored on every job, and the agreement rate is reported to you rather than asserted.
Build rather than buy
Why we did not just licence someone else’s.
Buying would have been faster. It would also have made us the same as every other provider in this category, with the same cost base and the same reasons to quote by the hour.
See how much of each service it produces →- Because a licence becomes your line item
- Per-seat software that a provider rents has to be recovered from somebody, and it is recovered from the client. Owning the stack is what makes a published price possible at all, rather than a quote that has to cover a tooling bill.
- Because legal deliverables are not summaries
- General AI products summarise documents well. A demand package, a privilege log, a disclosure schedule and a fifty-state survey are not summaries - they are structured deliverables with their own rules, formats and failure modes, and those rules have to live inside the system rather than in a prompt.
- Because the improvement has to be ours to give away
- Every exception a lawyer resolves and every QC sample scored feeds back into the thresholds and the schemas. The system gets faster on our time, not on your invoice: if our process improves, that is our gain and your published price does not move.
Where the machine stops and a person starts.
Every matter runs the same sequence. The pipeline proposes; a qualified lawyer disposes. The supervision of the engaging attorney sits across all of it.
What we use it for
Our pipeline handles extraction, classification, summarisation, timeline construction, and schema mapping. This is what allows us to price per unit instead of per hour, and to return work in days instead of weeks.
What a human does
A qualified lawyer defines the scope, reviews the output, resolves every flagged exception, and signs off before delivery. We do not deliver unreviewed machine output.
What we do not do
We do not use your files, prompts, or outputs to train any model. We do not use consumer AI products for client material. We do not make legal determinations, and we do not advise your clients.
Data handling
Every row, filled in honestly.
- Models used
- Enterprise-grade large language models, deployed with model training disabled at the account level.
- Training on your data
- Never. Disabled at the account level.
- Where data sits
- India, in encrypted cloud storage.
- Retention
- 90 days after delivery, then permanent deletion.
- Deletion on request
- Within 5 business days, confirmed in writing.
- Access
- A small number of named individuals, under NDA.
- Transfer
- Encrypted in transit and at rest.
- Confidentiality
- NDA signed before any file is received.
Confidentiality in practice
The rules behind the table above.
01
Read access only, unless the matter requires more
We work from what you send us or from read access to the specific file share for the matter at hand. We do not require standing access to your case management system, and any access we do hold is scoped to the engagement, not the whole practice.
02
No client data trains any model
Matter content, documents, and communications are never used to train a model, ours or a third party's. Where a provider's default terms would otherwise permit training on submitted content, we use the deployment mode that disables it.
03
A qualified lawyer signs off before delivery
Judgment calls, close questions, and anything the pipeline marks as low confidence go to a person. We do not deliver unreviewed machine output, and we do not make legal determinations or advise your clients.
04
Confidentiality is contractual before it is cultural
We run a conflicts check before accepting any engagement, and we execute a mutual confidentiality agreement before receiving a single file. Named individuals only see client matters.
05
Every assertion is anchored to a source
Extraction and summarisation work cites the page or line it came from. That is a quality control and a confidentiality control at once: nothing enters your work product without a reviewer being able to see exactly what it is based on.
06
Deletion on request, confirmed in writing
Working copies of client materials held for the duration of an engagement are deleted on request once the matter closes, and we confirm deletion in writing.
The questions a risk committee asks
Is the AI making decisions about our matter?
No. The pipeline handles extraction, classification, summarisation, timeline construction and schema mapping. A qualified lawyer defines the scope, reviews the output, resolves every flagged exception and signs off before anything is delivered. Nothing unreviewed leaves us.
Do you train models on our files?
Never. Training on customer content is disabled at the account level, and we do not use consumer AI products for client material. Your files, prompts and outputs are not used to improve any model, ours or a third party's.
Where does our data physically sit?
In encrypted cloud storage in India, accessible only to a small number of named individuals under NDA, encrypted in transit and at rest, and deleted 90 days after delivery or within 5 business days of a deletion request, confirmed in writing.
Can you complete our security questionnaire?
Yes. Send it with your outside-counsel guidelines and we answer it ourselves rather than routing it through a sales team. Most of what a questionnaire asks is already answered on this page.
Questions about this
Ask us anything about how this works.
Send us your security questionnaire, your outside-counsel guidelines, or just a direct question. We answer them ourselves.