Most evaluations start with a shortlist and work backwards to the criteria.
That is how a firm ends up buying the best demo.
Decide what your practice actually requires first.
Then put every tool through the same test, in the same order, on the same evidence.
This guide is that test. Ours is scored on it too.
The first question is not what a tool can do, it is who it was drawn up for. A platform aimed at one practice will be deeper in it. A platform aimed at every practice will be broader across all of them. Both are legitimate businesses, and only one of them matches your docket.
Built for everyone? Then built for no one. Find the tool built for the work your firm actually does.
Placement is our reading of each product's public positioning and documented feature set, not a benchmark score. The research suites sit low and right on purpose: they run a long way into a matter, they are simply aimed at every practice while they do it. Focus is a choice, not a grade.
Every tool here can produce text about your matter. The question is whether the output arrives in the shape the court expects, with the caption right, the authority checked and the local rules honored, or whether you are rebuilding it from a draft before anything gets filed.
| Tool | Caption validated Validates the caption against the court's local form. | Citations verified Confirms each cited case exists, the quotation is accurate and the pin cite is right. | Local rules checked Checks the filing against the court's local rules. | Judge-aware drafting Drafts to the assigned judge's documented preferences on structure, authority and length. | AI declaration certificate Produces the certification of AI use that a growing number of standing orders now require. | Certificate of service checked Checks that a certificate of service is present and complete. | Word-count certificate checked Computes the word count and checks the certificate against the court's limit. |
|---|---|---|---|---|---|---|---|
| OurFirm.ai | Handled by the tool, first-party | Handled by the tool, first-party | Handled by the tool, first-party | Handled by the tool, first-party | Handled by the tool, first-party | Handled by the tool, first-party | Handled by the tool, first-party |
| CoCounsel | Partial, needs review, unverified reading | Partial, needs review, unverified reading | Left to you, unverified reading | Left to you, unverified reading | Left to you, unverified reading | Left to you, unverified reading | Left to you, unverified reading |
| Protégé | Partial, needs review, unverified reading | Partial, needs review, unverified reading | Left to you, unverified reading | Left to you, unverified reading | Left to you, unverified reading | Left to you, unverified reading | Left to you, unverified reading |
| Harvey | Partial, needs review, unverified reading | Partial, needs review, unverified reading | Left to you, unverified reading | Left to you, unverified reading | Left to you, unverified reading | Left to you, unverified reading | Left to you, unverified reading |
| Legora | Partial, needs review, unverified reading | Partial, needs review, unverified reading | Left to you, unverified reading | Left to you, unverified reading | Left to you, unverified reading | Left to you, unverified reading | Left to you, unverified reading |
| ChatGPT | Left to you, unverified reading | Left to you, unverified reading | Left to you, unverified reading | Left to you, unverified reading | Left to you, unverified reading | Left to you, unverified reading | Left to you, unverified reading |
| Claude | Left to you, unverified reading | Left to you, unverified reading | Left to you, unverified reading | Left to you, unverified reading | Left to you, unverified reading | Left to you, unverified reading | Left to you, unverified reading |
Local rules coverage is federal district courts and New York State e-filing (NYSCEF). Other state courts are not covered.
Underlying AI model, and multi-agent execution
01
Every product in this guide rents its intelligence from the same small set of frontier labs, on the same release schedule. On the underlying model row, four of the six competing tools score exactly what OurFirm.ai scores. Whatever one vendor can license this quarter, the rest can license next quarter.
A supply layer at the top feeds all seven products in the guide, drawn identically because on this criterion they are alike. OurFirm.ai scores full here, and so do four of the six competing tools.
02
Take out what everyone already has and the model stops being a differentiator. It flattens into a floor: the same input, available to anyone willing to pay for it. The question worth your diligence is what a product is built to do with that input once it has it.
The supply layer flattens into a single thin band labelled table stakes. All seven products still sit on it, receiving the same input. The band stays in place for the rest of the sequence.
03
This is where the products stop looking alike. OurFirm.ai runs specialized agents in parallel across one matter: research, drafting, citation verification and formatting, each working at the same time on the same record. A general legal AI platform runs a single thread, one prompt and one answer at a time.
Below the band the paths separate. One path fans into four agents running in parallel: research, drafting, citation check and formatting. The other continues as a single thread. Paths are labelled by category, not by product name. The OurFirm.ai path is first-party. The path for a general legal AI platform is our own reading and is not a sourced finding.
04
The parallel agents converge on a reconciliation step, and what comes out is a filing, with the caption, the authority and the local rules already checked against each other. The single thread emits a draft with the unresolved items still in it, and the reconciling left to you.
The four parallel agents converge on a reconciliation step which emits a filing. The single thread emits a draft that still carries unresolved items. Both paths reach their output at the same point in the diagram: it shows sequence, not elapsed time. The OurFirm.ai path is first-party. The path for a general legal AI platform is our own reading and is not a sourced finding.
There is a difference between reporting statistics about a judge and modeling how that judge has handled issues like yours. Grant rates and timing are useful context. A profile built from the judge's own orders and transcripts is a different instrument entirely.
| Tool | Coverage |
|---|---|
| OurFirm.aiJudge behavioral profiles | Full. First-party |
| CoCounselLitigation Analytics dashboards | Partial. Our own reading, not a sourced finding |
| ProtégéLex Machina analytics | Partial. Our own reading, not a sourced finding |
| HarveyNot offered | None. Our own reading, not a sourced finding |
| LegoraNot offered | None. Our own reading, not a sourced finding |
| ChatGPTNot offered | None. Our own reading, not a sourced finding |
| ClaudeNot offered | None. Our own reading, not a sourced finding |
Scoring reflects each vendor's publicly documented capabilities at the revision date. Every OurFirm.ai claim is first-party. Competitor cells are our own reading and are marked as such until a public source is recorded against them.
Judge profile
Illustrative. Not a real judge.
Hon. Jane Q. Example
United States District Court
This card is a specimen, not a real judge, and is here to show the shape of a profile rather than the contents of one. Each line is labelled by the body of material it comes from rather than cited to a document. Updated September 2026.
Judge AI
Analytics report what a judge decided. A profile reads how they think. Each cluster below is one dimension of a judge, built from their own orders, transcripts and docket entries; open one and it tells you what the bench does, in the language it does it in, and what that means for the way you file.
Illustrative. Not a real judge.
Select a dimension to read it.
This row is not a preference you weigh against price. Filing a fabricated citation is a sanctions problem and an ethics problem, so verification is required, functionally and professionally. What separates the tools is what they actually confirm: that the case exists, that the quote is accurate, that the pin cite is right, and that the authority supports the proposition it is cited for. Four things, on every cite, on screen in the draft you are reading. Ask any other tool how you check a cite and the answer is a hyperlink out to another platform and a control-F, one cite at a time, done by you.
Scoring reflects each vendor's publicly documented capabilities at the revision date. Every OurFirm.ai claim is first-party. Full labeled cells for all fourteen criteria are in the scorecard at the end and on the compare pages.
What the platform knows
Every vendor rents the same frontier models, and those models improve for everyone on the same schedule. Whatever intelligence one of them can buy, the rest can buy too. The durable difference is the proprietary data underneath and whether you can open the source document behind any answer it gives you.
How to actually test it
Demos run on clean fact patterns chosen by the people who built the tool. Insist on a live matter instead: your record, your jurisdiction, a real filing deadline, a document you would put your name on. Judge the output the way the judge will, not the way the sales engineer does.
Depth versus coverage
One tool that covers every practice area will be shallower in each of them, and that is a reasonable trade if your work is genuinely broad. Decide which you are before you shortlist. A firm that mostly litigates and buys for breadth ends up paying for departments it does not have.
Where the risk actually sits
“We cite our sources” and “we confirm the case exists, the quote is accurate, the pin cite is right, and the authority supports the proposition, inside your draft” are different products at the same price point. Ask which one you are buying, then ask to watch it catch a bad cite on screen, without leaving the document.
Do you confirm the case exists, the quote is accurate, the pin cite is right, and the authority supports the proposition? Or treatment only?
Is your no-training commitment contractual or a policy you can change? Show me the clause. Every competitor above is policy-based or plan-dependent.
Do you publish a judicial behavioral dataset, or do you report statistics like grant rates and timing?
Can I get a published rate and clear billing terms, or is this a negotiated enterprise contract with pricing behind a sales process?
The one row we lose is row 2, breadth beyond litigation. We do litigation and nothing else. A low score elsewhere is not a bad product either: the other six are research libraries and general assistants aimed at every practice at once, so a litigation row sits outside what they set out to do. Breadth is a choice, not a failure, and it is the one we declined to make.
| Criterion | OurFirm.ai | CoCounsel | Protégé | Harvey | Legora | ChatGPT | Claude |
|---|---|---|---|---|---|---|---|
| 01Purpose-built for litigators | full. Litigation only | none. Research + general legal AI | none. Research + general legal AI | none. General legal AI | none. General legal AI | none. General-purpose | none. General-purpose |
| 02Breadth beyond litigation | none. Litigation scope only | full. Broad legal suite | full. Broad legal suite | full. Broad practice mix | full. Broad practice types | full. General-purpose | full. General-purpose |
| 03Underlying AI model | full. Multi-model + legal fine-tuning | full. CoCounsel (multi-model) | full. Protégé (multi-model) | full. Multi-model (GPT/Claude/Gemini) | full. Proprietary + frontier LLMs | partial. General frontier models | partial. General frontier model |
| 04Court-ready documents | full. Court-ready output | partial. Drafting assistance | partial. Drafting assistance | partial. General drafting | partial. General drafting | partial. Unreviewed drafts | partial. Unreviewed drafts |
| 05Judge AI, judicial modeling | full. Judge behavioral profiles | partial. Litigation Analytics dashboards | partial. Lex Machina analytics | none. Not offered | none. Not offered | none. Not offered | none. Not offered |
| 06Multi-agent execution | full. Multi-agent swarm | partial. CoCounsel skills / agents | partial. Protégé agentic workflows | partial. Workflow Builder + agents | partial. Agentic research workflows | partial. General agents | partial. General agents / tools |
| 07Legal research quality | full. Nationwide caselaw + semantic search | full. Westlaw research (deep) | full. Lexis research (deep) | full. Deep Research, 200+ sources | full. Agentic research + Tabular Review | partial. Not litigation-tuned | partial. Not litigation-tuned |
| 08Semantic docket search | full. PACER + NYSCEF, direct pull | partial. Docket tools, keyword | partial. CourtLink dockets, keyword | partial. Limited docket | none. Not offered | none. No docket access | none. No docket access |
| 09Discovery and doc review | full. Matrix document review | partial. CoCounsel review | partial. Review tooling | full. Vault doc analysis | full. Tabular Review grid | none. No structured review | none. No structured review |
| 10Deposition and trial prep | full. Outlines + objection prep | partial. Transcript tools | partial. Transcript tools | partial. General litigation workflows | partial. Research surfaces precedent | none. No structured support | none. No structured support |
| 11Citation verification | full. Every cite verified | partial. KeyCite citator | partial. Shepard's citator | none. No cite-verify pipeline | none. No US cite-check | none. Hallucination risk | none. No cite verification |
| 12Enterprise-grade security | full. Enterprise-grade security | partial. Enterprise controls | partial. Enterprise controls | partial. Enterprise controls | partial. General controls | none. Not security-designed | none. Not security-designed |
| 13No training on client data | full. Contractual no-train | partial. Enterprise data terms | partial. Enterprise data terms | partial. Policy-based | partial. Policy-based | partial. Varies by plan / API | partial. Varies by plan / API |
| 14Pricing and access model | full. Platform fee + compute costs | none. Enterprise / premium, not public | none. Enterprise / premium, not public | none. Enterprise / not public | none. Custom / not public | partial. Low per-seat cost | partial. Low per-seat cost |
| Full of 14 | 13 | 3 | 3 | 4 | 4 | 1 | 1 |
Scoring reflects each vendor's publicly documented capabilities at the revision date. Every OurFirm.ai claim is first-party. General-purpose rather than trial-specific scores partial, and a capability outside a product's stated purpose scores none, because a research suite aimed at every practice is not failing at litigation, it is not attempting it. Each cell's factual label is on the compare pages.
We handle the most complex litigation in the country. OurFirm is our second chair.
You know your practice better than any vendor pitching you, so do not take our scoring on faith, or anyone else’s. Name a judge you appear before and we will build the profile from that judge’s own opinions and transcripts, with the citations behind every line. Nothing of yours changes hands, and you can tell us whether it reads true.
No commitment, and nothing confidential required. Name the bench and read the profile the way you would read an adversary’s brief.
How we scored this. Revised September 2026, next review March 2027. Scoring reflects publicly documented capabilities as of the revision date; each cell's factual label is shown in the matrix above and on the compare pages. Claims about competitors are OurFirm.ai's reading of their public materials, not statements from those companies. General-purpose rather than trial-specific scores partial. A capability outside a product's stated purpose scores none, because a contract tool is not failing at litigation, it is simply not attempting it. Every OurFirm.ai claim is first-party and we stand behind it. If you believe we have mischaracterized a product, write to us and we will correct it.
Security. Contractual no-train with our AI subprocessors, per-firm and per-matter isolation, TLS 1.3 and AES-256, SOC 2 Type 1 and Type 2 complete, HIPAA-aligned BAA and GDPR and CCPA-aligned DPA under NDA.