Series A LegalTech Built an AI Trademark Drafting Platform

Brandiligence is an AI platform that automates trademark similarity and confusion analysis for legal professionals. It compares wordmarks, logos and classifications, then produces clearance opinions that used to take a lawyer hours. It was built for IP attorneys and brand consultants, and shaped throughout by a practising trademark attorney who does this work every day.

A trademark clearance opinion is not a piece of writing. It is a fixed argument: the same five sections, in the same order, applying the same legal test, in the jurisdiction that governs the matter, citing authorities that exist. A lawyer produces it from a blank page every time, and most of the hours go into structure and sourcing rather than into judgement.

That looks like an obvious automation target, which is exactly the trap. Out of the box, a general model gives a firm the opposite of what it needs. Every answer comes out structured differently, so nothing is client-ready. It invents or misattributes citations. It drifts between jurisdictions. It slides into an informal register no firm would put its name on.

And in legal diligence the cost of a mistake is not rework. One fabricated citation can sink a deal and take the firm's credibility with it. A tool that is right 95% of the time is not 95% useful here — it is unusable, because the attorney has to check all of it to find the 5%, which is the work they were trying to avoid.

So the requirement was not accuracy in the usual sense. It was provability: a system whose constraints a firm could inspect and rely on, rather than a system whose output they had to trust.

• Hold a fixed five-part structure on every opinion, with no drift between matters

• Cite only approved authorities, with no route to an invented or misattributed case

• Write in a register an attorney would sign, not a chatbot's

• Be measurable, so quality is something demonstrated rather than asserted

Rather than bolt a chatbot onto the product, we rebuilt the drafting step. Base models fine-tuned on attorney-written opinions, wrapped in retrieval over a curated library and a validation layer that checks the output before anyone sees it.

The design principle is easy to state and was the hard part to build: the model is constrained by the structure it must fill, the sources it is allowed to see, and the validators that check its work. It is not asked to be trustworthy. It is placed somewhere untrustworthiness cannot do damage.

The citation constraint is the clearest case. The model has no access to legal authorities beyond the firm's approved index — not a discouragement in a prompt, an absence of any path. A draft citing something that does not resolve is rejected and retried against retrieved evidence. That converts "does it hallucinate citations?" from a question about model behaviour, which nobody can answer confidently, into a question about system architecture, which a firm's own technical reviewer can verify in an afternoon.

Design decision: the attorney stays the legal authority. The system's job is to arrive with a complete, structured, source-checked draft — not to decide anything. Every control exists to keep that line intact.

The matter is normalised

Mark type, jurisdiction and NICE class are established up front, because they determine which authorities are even applicable.

Applicable authorities are retrieved

From curated case law, statutes and registry data. The retrieval step defines the universe of what the draft can cite, before a word is written.

The draft is written into a fixed shape

Five sections, in order, in the firm's register — not a free-form answer that happens to contain the right parts.

The system checks its own work

Structure, jurisdiction tags and citation formats are validated. A missing section, a malformed citation or a source that does not resolve triggers a retry drawing only on retrieved evidence. Invented cases do not reach the attorney.

The attorney reviews and signs

The lawyer edits and signs rather than drafting from a blank page. They remain the decision-maker on every matter.

Who the authority is

Not the model. The approved corpus, the jurisdiction rules, the five-part schema, the citation validator and the reviewing attorney between them define what can reach a final opinion. An unsupported source cannot pass through as a citation — a malformed or unresolved one causes rejection and a controlled retry.

What runs, in what order

Normalise the matter, retrieve the applicable authorities, draft the five sections, validate structure and citations, retry anything invalid, present a reviewable draft. Fine-tuning supplies structure and register. Retrieval supplies evidence. Schemas and validators control the output. The attorney closes the loop.

What you can see afterwards

A reviewer can inspect the authorities retrieved, the sections generated, how each citation resolved, which validations failed, what was retried, and what the attorney changed. Alongside that runs a 0–10 legal-integrity harness scoring structure, reasoning depth, citation accuracy, jurisdiction discipline, edge-case handling and tone, with production logging that compares base against fine-tuned models so a regression shows up as a number rather than as a complaint.

Trademark drafting moved from something a lawyer did by hand to something a lawyer reviews and signs.

The behaviour change is the result worth reporting. A practising attorney stopped treating the system as a novelty and started treating it as a junior associate whose work they could trust — which is a judgement about reliability that no internal metric substitutes for. The client confirmed that attorney tested it extensively and approved its direction and quality.

It has been in production since 2023, which for a first GenAI build is its own kind of evidence: the constraints held up outside the conditions they were designed in.

• A citation outside the approved index is not possible, because the model has no path to one and a validator rejects the draft if one appears

• The five-part structure holds on every draft, enforced by schema-constrained decoding rather than requested in a prompt

• Verified by the practising attorney it was built for, and continuously in production since 2023

This page deliberately publishes fewer numbers than it could, and it is worth saying why.

Our internal records from the build carry three figures: structural adherence rising from roughly 40% to 98%, hallucinated citations falling by around 90%, and attorney review getting 30–50% faster. They come from our own evaluation during development. They were never independently measured, and our own evidence standard says a number that has not been measured properly does not get published as a result. So they are not headline claims here.

What we will stand behind is stronger anyway, because you can check it. The citation constraint is architectural — have your own engineer look at it. The five-part structure is enforced by the decoder, not requested of the model. The attorney sign-off is a client-confirmed fact. And the system has run in production since 2023.

For your own estimate: matters per month × (current drafting time − assisted review time) × loaded professional cost. The second term is the one to measure honestly on your own matters — a draft that needs restructuring saves far less than one that only needs judgement applied.