Charges and indexation
Off playbookThe Supplier may increase the Charges once in each Contract Year by the annual increase in the Consumer Prices Index plus 3%. The Supplier shall give not less than 30 days’ written notice of any increase.
Document intelligence for law firms
The workspace where legal documents explain themselves.
Pratikando reads every document on a matter, compares versions clause by clause, surfaces the obligations, dates and risks that decide the outcome, and drafts the first version — with every statement traceable to the paragraph it came from.
Lindqvist Ferrante / NGL-2411 / Supplier redline v4.2
Search clauses, matters, know-how⌘KAF
Document analysis
Master Services Agreement — supplier redline
Analysed51 pages
Active matters
14%8
Documents analysed
41%155
Drafts generated
27%46
Clauses outside the playbook
6 of 12Service levels and service credits
Off playbookWhere the Supplier fails to meet a Service Level, Service Credits shall accrue at 0.5% of the monthly Charges per affected Service Level, up to a maximum of 5% of the monthly Charges. Service Credits are the Customer’s sole and exclusive remedy for Service Level failures.
Term and renewal
Off playbookThis Agreement commences on the Effective Date and continues for three years, after which it renews automatically for successive periods of 24 months unless either party gives not less than nine months’ written notice before the end of the then-current term.
Data-breach indemnity brought inside the liability cap
Deleting the clause 12.2 carve-out means regulatory fines and breach response costs are recoverable only up to 100% of annual charges — roughly £1.4m against a plausible seven-figure exposure for a group-wide incident.
Recommendation
Reinstate the clause 11.5 carve-out. If the supplier resists, accept a super-cap at 300% of annual charges for data protection claims and require cyber cover of at least £5m named in clause 21.
Version comparison
Clause 11.3Risk to our clientEach party’s total liability shall not exceed 150% 100% of the Charges paid or payable in that Contract Year.
Every extracted field keeps a page and paragraph offset back to the clause it came from.
A named fee earner accepts, amends or rejects each output before it leaves the firm.
- 12
- clause categories extracted per contract
- 41s
- median time to analyse a 50-page agreement
- 100%
- of AI output reviewed by a named fee earner
Built on AWSAmazon BedrockEU data residencyPilot cohort — 5 firms
The problem
Small and mid-sized firms lose margin in the reading, not in the advice.
Every firm we have spoken to describes the same four failures. None of them is a technology problem on its own; together they are where the hours go.
01
Document overload with no shape to it
A mid-sized matter arrives as 40 files across email, a document management system and a client portal. Nobody can say what the set contains without reading it, so the reading happens twice: once by the associate, again by the partner.
3.5hper matter re-reading documents the firm has already read
02
Version chaos at exactly the wrong moment
Redlines land the day before a deadline. Comparing them is manual, and the question is never “what words changed” but “which of these changes actually moves risk”. Track changes cannot answer that.
1 in 6negotiations where a material change is spotted only after signature
03
Repetitive drafting that still needs a lawyer
Risk memos, response letters, reports on title, amendment letters. The structure is identical every time and the content is 70% derivable from the documents already on the matter — but it is still typed from scratch.
~40%of drafting time spent on structure rather than judgement
04
Knowledge that leaves when people do
The firm has already decided how it treats indexation caps, break conditions and leaver provisions. That position lives in a partner’s memory and in a folder nobody searches, so each matter re-litigates it.
0searchable record of the firm’s own negotiated positions
Figures are from structured interviews with eight commercial and corporate practices in Italy and the UK (4–60 fee earners), conducted while defining the pilot scope. They are directional, not a published benchmark, and we share the underlying notes with design partners.
How it works
Four moves, in the order a matter actually runs.
Pratikando is not a chat window bolted onto a document store. It is a working surface for the specific tasks that repeat on every matter.
01
A workspace per matter, not a folder per file
Every document, extraction, draft and decision sits inside the matter it belongs to, with its own access ring. A client viewer sees one matter; a paralegal sees the three they are staffed on; nothing is workspace-wide by default.
NGL-2411 · Northgate Logistics Group
Master services agreement renegotiation
Alessandra Ferrante, Priya Nandakumar, Harriet Cole
ActiveCommercial contracts14 documentsCounter-position due in 3 days
Documents
14
Open risks
6
Diarised dates
4
02
Document intelligence that returns structure, not prose
Each document is normalised, read for layout, then extracted into a fixed schema: clauses by category, parties, obligations with owners and triggers, dates, and risk flags scored against the firm’s own playbook. Every field keeps an offset back to the source paragraph.
- Clauses
- 12
- 6 off playbook
- Parties
- 3
- incl. guarantor
- Obligations
- 6
- 1 overdue
- Key dates
- 4
- ready to diarise
- Risk flags
- 6
- 1 critical
- Confidence
- 94%
- mean per field
Clause 8.1
24-month automatic renewal · nine-month notice window
Playbook alignment 82% · PB-02 breached
03
A comparison engine that answers “what moved”
Versions are aligned clause by clause, so renumbering is reported as renumbering. Each change is labelled with direction of risk and an impact note tied to the client’s commercial position — the question a partner actually asks.
Clause 3.4
Charges and indexation
Risk to our clientCPI, capped at 4% per Contract YearCPI + 3%, no cap
Clause 5.2
Service credits
Risk to our client1.5% per Service Level, not exclusive0.5% per Service Level, sole remedy
37 changes · 11 clauses · net risk shift +38 towards our client
04
A drafting copilot that must show its work
Templates take structured input — audience, commercial priority, negotiating posture — then retrieve the firm’s playbook and previous drafts before composing. Sections without a citation are dropped, and a named fee earner signs off before anything leaves the firm.
Proposed wording
Amendment — clause 11.5
“Nothing in this Agreement limits or excludes liability for death or personal injury caused by negligence, fraud, or the Supplier’s obligations under clause 12.2 (Data Protection Indemnity).”
Sources — every section is traceable
- Clause 11.5Supplier redline v4.2
- PB-04Playbook — liability caps and carve-outs
- Variant AClause bank — data protection super-cap
Capabilities
Six things the platform does well, rather than everything badly.
Each capability was scoped against a task a fee earner already does by hand today, and each one produces output that can be checked.
Clause, obligation and risk extraction
A schema-constrained pass over every document on the matter: clauses by category, parties and their roles, obligations with owner and trigger, dates, and risk flags scored against the firm playbook. Confidence is reported per field, and low-confidence fields are surfaced for review rather than hidden.
94%mean field
Extraction coverage — last 155 documents
- Accepted62
- Amended28
- Rejected10
Version comparison with risk direction
Clause-level alignment, word-level diff, and an impact note for every change that says which way risk moved and why.
- Clause 11.3Cap 150% → 100%
- Clause 16.1Exit charge added
Precedent and know-how search
Semantic search across the firm’s documents, clause bank, playbooks and previous drafts — scoped to what the searcher is allowed to see.
liability cap carve-outbreak notice conditiongood leaver vesting
Drafting from the firm’s own templates
Risk memos, counter-position letters, amendment letters, reports on lease terms, board minutes. Structured inputs in, cited sections out — summary, the two or three issues that matter, proposed wording, and next steps. The draft arrives ready to be argued with, which is the point.
- 6
- templates live
- 46
- drafts generated
- 82%
- accepted or amended
- 55min
- median time saved
Obligation and deadline register
Notice windows, renewal dates and reporting duties are extracted with their clause reference and pushed into a register the firm can act on. The nine-month non-renewal notice is diarised on the day the contract is executed, not the week it is missed.
Playbook enforcement and reviewer trail
The firm’s standard and fallback positions are encoded as rules, so a deviation is named rather than noticed. Every AI output records who accepted, amended or rejected it, when, and against which model version.
What Pratikando deliberately does not do: give legal advice, file anything, or send a document outside the firm. It prepares work for a lawyer to take responsibility for.
Why now
Operational efficiency stopped being optional, and the tooling finally fits the work.
Two of these forces are commercial and one is technical. Together they are why a document intelligence product aimed at small and mid-sized firms is buildable this year and was not three years ago.
I
Fee pressure has reached the mid-market
Fixed fees and capped scopes are now normal for exactly the work that used to absorb reading time. A practice that cannot compress document review compresses its own margin instead.
Buyers of legal services increasingly price by matter, not by hour.
II
Extraction became accurate enough to be defensible
Structured clause and obligation extraction crossed the threshold where a lawyer can rely on it as a first pass. That is recent, and it is the difference between a demo and a tool: the output can be checked against the paragraph it came from in seconds.
The constraint moved from model capability to product design.
III
Auditability is becoming the price of entry
Firms are being asked by their own clients, insurers and regulators how AI is used on their matters. Products that record who reviewed what, against which model, are the only ones that survive that question — and it is far easier to build that in from the start than to retrofit it.
Human-in-the-loop review is a design constraint, not a feature.
Where we are starting
Commercial contracts, corporate transactions, real estate and data protection work in firms of 4 to 60 fee earners in Italy and the UK. These practices have enough repeat document work to benefit immediately, and short enough decision chains to run a pilot in weeks rather than quarters.
Trust & governance
Confidentiality is the reason firms say no. We designed for the question before it was asked.
These are the four controls a firm’s risk partner tests first. Each one is in the prototype today, not on a roadmap slide.
Role-based access, matter by matter
Five roles from workspace owner to client viewer. Access is granted per matter, not per workspace, and a client viewer sees exactly one matter with no export rights.
Workspace isolation by design
Each firm’s documents, retrieval and activity stay inside their own workspace boundary. Cross-tenant access is not a permission that exists to be misconfigured.
Human review before anything is relied on
Every extraction, comparison and draft carries a review state and a named reviewer. Nothing is marked fit for client use by the system itself.
Auditability a risk partner can read
Access, AI invocation and review decisions are recorded with actor, role, resource, outcome and model version, and exportable for a date range.
Audit log — live from the prototype
ImmutableAccepted AI comparison output
allowedAlessandra Ferrante · Workspace owner · Comparison v3.1 → v4.2
Generated draft from template
allowedAlessandra Ferrante · Workspace owner · tpl_counter_letter
Uploaded document batch
allowedSofia Renzi · Associate · doc_helvia_processor_batch
Attempted to approve AI output
deniedMarcus Bell · Paralegal · Analysis doc_vireo_csa_site2
Exported analysis to PDF
allowedDaniel Okonjo · Associate · doc_kestrel_delay_notices
Attempted to open restricted matter
deniedHarriet Cole · Paralegal · mtr_vireo_series_b
The denied entries are real behaviour, not decoration: a paralegal cannot approve AI output, and a document upload cannot reach a matter the uploader is not staffed on.
Client documents are never used to train models
Inference runs against managed models with no retention of prompt content for training.
Retention follows the matter, not the vendor
Documents follow the firm’s retention policy; AI traces are kept for 24 months for audit.
Ethical walls are enforceable
A matter can be walled so that even a workspace owner must record a reason to enter.
Built on AWS
Infrastructure for secure legal workflows.
Pratikando is designed to support secure document workflows, reliable AI-assisted review, and controlled rollout from pilot firms to broader adoption.
How the platform is grounded
Serious infrastructure underneath a refined product — built to earn trust from firms and scale without changing the model of work.
- 01
Document
Stored per matter
- 02
Intelligence
Bedrock extraction & search
- 03
Human review
Named fee earner
- 04
Adoption
Pilot → firm rollout
Document intelligence on Amazon Bedrock
Extraction, semantic retrieval and assisted drafting run on Bedrock — inside the account and region boundary, with no client content used for model training.
Secure document storage
Matter-centric storage keeps each firm’s files organised and isolated, so review and drafting stay attached to the right workspace.
Role-aware access
Authentication and matter-level permissions decide who can see, analyse or approve work — before any AI output is relied on.
Reviewable activity
Access, AI use and human review decisions are attributable, so firms can explain how the product was used on a matter.
Operational visibility
Monitoring across the application and AI path supports controlled pilots and reliable day-to-day operation as usage grows.
The same foundation serves a five-firm pilot and a broader deployment. Quotas and capacity change; the product model does not.
- Region
- EU residency
- Intelligence
- Amazon Bedrock
- Path
- Pilot → scale
Pilot programme
Five design-partner firms. Six weeks. Evidence at the end, not a testimonial.
We are onboarding the first cohort now. The pilot is deliberately short and deliberately measured, because a firm should be able to say no on data.
3 of 5 places remaining
Cohort opens next quarter
Week 1
Workspace and playbook
We create the firm workspace, load two or three live matters, and encode the firm’s standard positions on the clauses it negotiates most.
Weeks 2–3
Run it on real matters
The pilot team uses extraction, comparison and one drafting template on live work. We sit in on two review sessions a week and fix what gets in the way.
Weeks 4–6
Measure and decide
We report acceptance rate per output type, time saved per matter and every case where the output was wrong. The firm decides on that evidence.
Who this is for
- 4 to 60 fee earners
- Commercial, corporate, real estate or data protection work
- A partner who will sponsor the pilot
- Willingness to encode two playbook positions
What the firm gets
- Firm workspace with unlimited matters for the pilot period
- Playbook encoding done with you, not handed to you
- Named engineer on a shared channel
- Pilot pricing held for 12 months after conversion
Design-partner firms are not named publicly until they choose to be. The prototype on this site runs entirely on invented matters and clients so it can be shown without a single confidentiality conversation.
Bring one negotiation. We will show you what the platform sees.
The fastest way to judge Pratikando is to run it against a redline your team has already reviewed, and compare. That is how every pilot conversation starts.
Pilot enquiry