The demo never fails. The deposition file fails, and the contract archive fails, three weeks later on an ordinary Tuesday. Law firms come to an AI shortlist carrying two pressures at once: material a protective order already governs, and a volume of paper no partner can staff. Those two pull toward different products, which is why the answer to what is the best AI for a law firm is a split rather than a name.
Best AI for Law Firms
Best AI Tools for Legal Work:
8 Options and Where Each One Stops
The shortlist a firm actually assembles, sorted by where the document has to sit — with the limit on our own product stated first and the empty cells left visibly empty.
Built from real buyer questions in our sales meetings
Direct Answer
The decision splits, and where the document has to sit is what splits it. Two legal jobs cannot reach a network at all — querying the whole case file from a courtroom with no wi-fi, and transcribing a deposition on the machine that recorded it — and an assistant running on the device is the only shape that survives either one. Extraction of clauses, dates and obligations across a firm's entire contract base runs the other way. It is the largest legal job our conversations record, it goes to a dedicated cloud legal platform, and AirgapAI loses it outright.
Where AirgapAI is not the answer. A legal document analysis use case is outside AirgapAI's scope, because the models required are bigger than a laptop can handle unless it is an Apple laptop. AirgapAI does not support the AudioNote file extension, so recordings a firm already holds in that format cannot be imported. And a three-hour deposition recording would not open or play back, and displayed an error in the app. Those are our records about our own product, and none of them is softened here.
Verify every claim on your own material, ours included. Hand each name on the list one deposition recording in the format your firm actually records in, one representative contract set at your real volume, and one question whose answer must cite the source document by name. Products that clear those three are on your shortlist. Products that need your files converted first have already answered you.
Named options and recorded limits, not scores. The eight below are named with the limit recorded against each, and none of them is scored. For more information on scoring a shortlist against criteria you set yourself, visit the selection scorecard; for the jobs legal buyers described in their own words, visit the legal use-cases page.
Who This Shortlist Is For
Legal services is one of the sectors our sales and customer conversations return to most, and it arrives with the same complaints in the same order. There is more paperwork than there are lawyers, so compliance obligations go unmet. Cost caps how many lawyers can be staffed on merger and acquisition document review. A solo practitioner works through very large case document sets with no firm behind them. Redlining drags back and forth over minor points while outside counsel bills by the hour.
The tooling complaints are just as consistent. General-purpose assistants cannot be relied on for agreements, contracts and policies. Playbook redlining is garbage in, garbage out, only as good as the playbook behind it, and one buyer named scale as the biggest challenge with the legal AI tools already in the building. Another put it flatly: AI is not living up to expectations in the legal department.
Then comes the constraint that outranks every capability question. Case documents sit under confidentiality and protective orders. Master service agreements with hundreds of partners forbid uploading anything to an AI product that could be exposed or learned from, and some firms may work with dummy data only. Buyers cited rulings that material entered into an AI tool is discoverable and does not meet the standard for attorney-client privilege. One firm — rather than the sector — called waiver risk an immediate roadblock, and answered it by keeping the environment on premises so the material is never made public. Permission is settled before any of the eight options is.
The Eight Options, and the One Limit Recorded Against Each
Options earn a row by how often our conversations name them, never by how well they compare. One qualifier belongs on the table before you read it: every property below is something a buyer told us in one of our own sales and customer meetings rather than the result of an independent benchmark, so treat each limit as a claim to test.
Two measures stay separate, because merging them would flatter the page. How widely our conversations name an option is one thing; what stands behind the single limit recorded against it is another. A product named constantly can still carry a limit one buyer described, and the reach of the name never transfers to the limit.
| Option | What it is, and what it is best for | How our conversations name it | Where inference happens | Cost shape | The one limit recorded against it |
|---|---|---|---|---|---|
| AirgapAI Ours |
Iternal's fully local assistant. Best for the two legal jobs that must work with no network: querying the whole case file from a courtroom with no wi-fi, and transcribing a deposition on the device that recorded it. | Named by legal buyers as their choice for material that will not go to the cloud. | On the device. | One-time, per device. | A legal document analysis use case is outside its scope, because the models required are bigger than a laptop can handle unless it is an Apple laptop. One firm working that use case described it. |
| ChatGPT | General drafting and summarizing, in work where no client matter material is involved. | The product our conversations name most often, across every sector. | The provider's cloud. | Free tier; paid tiers per seat, monthly. | Left empty on purpose. No mechanism-level weakness for it is recorded in our conversations — see what our conversations do not tell you. |
| Microsoft Copilot | Firms already licensed for the Office suite that want drafting inside the documents they work in. | Named nearly as often, as the suite a firm already pays for. | Microsoft's cloud. | Per seat, monthly. | Widely rolled out, and seen as too expensive to extend beyond a fraction of the workforce. One buyer described their own rollout that way. |
| Harvey | Contract review as a dedicated legal platform, including extraction across the whole contract base. | Named by legal buyers specifically, once they are looking at legal-specific tooling. | The provider's cloud. | Not recorded in our conversations. | It does not specialize in connecting to any system across the organization. A stated mechanism, described by one buyer. |
| NotebookLM | Feeding a bounded set of case documents and asking questions across only those documents. | Named occasionally, as the thing buyers compare document ingestion to. | Cloud hosted. | Not recorded in our conversations. | A cloud-hosted model, so files still transmit to their servers. Described by one buyer. |
| AudioNote | What a firm records depositions and takes timestamped notes in today. It jumps straight to the audio behind each typed note. | Named as the incumbent inside the firm that runs it. | Not recorded. Our conversations describe a recording and note-taking tool rather than an AI service. | Not recorded in our conversations. | Left empty on purpose. No weakness for it is recorded. What our conversations hold instead is our own limit, on the AirgapAI row. |
| Filevine | Case and matter management a firm keeps running alongside whatever else it adds. | Named rarely, and only as a system a firm intends to keep. | Not recorded. | Not recorded in our conversations. | Left empty on purpose. Recorded only as a competing case management system Iternal will not stop firms from using and integrating with, and no identifiable firm stands behind that record. |
| Court reporters The human incumbent |
Certified transcripts a court will accept. | Named as the incumbent the firm pays today. | None. A person transcribes. | Per transcript, with a further premium for real time. | The paid alternative: expensive for certified transcripts, and outrageous for real time. One firm described it that way. |
No price appears anywhere in that table, ours included. The cost column carries the shape of the charge — per device, per seat, per transcript — because a figure quoted outside its original scope is worse than no figure. Where our conversations record no shape at all, the cell says so.
Legal AI Software Compared: Category, Deployment and Cost Shape
Legal AI software falls into four categories: dedicated legal platforms, drafting add-ins
that live inside Word, practice and matter management, and case intelligence. Cloud
products cover breadth; an assistant running on the device covers material that cannot
leave the firm. Match the category to the job before comparing individual products.
This table extends the eight above rather than replacing them. The eight are the shortlist a firm assembles in our own sales and customer conversations; the six below are the legal-specific software a firm meets while searching, described from what each company publishes about its own product. Harvey appears in both, because it is the one dedicated legal platform that both lists reach.
| Product | Category | Deployment model | Where the material sits | Pricing model | What it is best at |
|---|---|---|---|---|---|
| AirgapAI Ours |
On-device assistant. Chat and retrieval over files held on the machine itself. | Installed on the laptop or workstation. Inference runs locally, with no network call needed to answer. | The document does not have to leave the firm to be read, which is why it carries the courtroom and deposition jobs on this page. | One-time, per device. | The on-premise pick: case-file questions with no connectivity, and deposition transcription on the machine that recorded it. |
| Harvey | Dedicated legal platform. Research, drafting and review built around legal work product. | The company's cloud. | Matter material is processed in the cloud under the terms of the agreement. Get retention and training-use commitments in writing first. | Subscription, quoted by the company. | The strongest fit in this table for extraction across an entire contract base — the job AirgapAI does not cover. |
| Spellbook | Drafting and review add-in that runs inside Microsoft Word. | Word add-in; drafts are processed in the company's cloud. | Draft language is transmitted for review. Get retention and training-use commitments in writing first. | Subscription, per seat. | Transactional lawyers who already spend the day in Word and want redlining where the draft already lives. |
| Clio | Practice and matter management — matters, calendaring, intake and billing — with AI features layered on the records the firm already keeps there. | Cloud practice management. | The matter record already sits in the platform, so the AI question is about what the existing system is permitted to do with it. | Subscription, per user, by plan tier. | Small and midsize firms whose system of record is already the place new AI features arrive. |
| Darrow | Case intelligence. Public court, agency and news data scanned for potential claims. | The company's cloud. | Works from public records rather than from the firm’s own privileged files, which puts it on a different footing from the rest of this table. | Subscription, quoted by the company. | Plaintiff firms sourcing matters, at the top of the funnel rather than inside documents already under a protective order. |
| Eve | Plaintiff-side case platform: intake, records review and demand preparation in one workflow. | The company's cloud. | Client material is uploaded to the platform to be worked. Get retention and training-use commitments in writing first. | Subscription, quoted by the company. | Plaintiff practices that want one workflow from intake through demand rather than a tool per step. |
No currency amount appears here either, for the same reason it does not appear above: a figure quoted outside its original scope is worse than no figure, and every one of these products is quoted against a firm's size and plan. The column that actually sorts a legal shortlist is where the material sits. What a firm is permitted to place in a cloud platform decides which products it may even pilot, and it decides that before any feature list is opened.
Score where client material travels in your firm today, and what a privilege-safe
deployment would require, with the
law firm AI confidentiality assessment.
Best Legal AI Tools by Job
Pick by job, not by brand. Contract-base extraction goes to a dedicated legal platform.
Drafting inside Word goes to a Word add-in. Matter management stays with the practice
system already in place. Case-file questions and deposition transcription with no network
go to an assistant that runs on the device.
- Reading the whole contract base for clauses, dates and obligations. A dedicated legal platform, Harvey being the one this page's conversations name. This is the largest legal job in our record and the one AirgapAI loses outright, for the reason stated in the direct answer above.
- Drafting and redlining an agreement. A Word add-in such as Spellbook keeps the work where the draft already is, which is usually the difference between an adopted tool and a licensed one.
- Running the matter, the calendar and the bill. Practice management such as Clio. The AI question here is what the system of record is already permitted to do with the file, not which assistant to add beside it.
- Sourcing matters before they are filed. Case intelligence such as Darrow, working public data rather than the firm's own privileged material.
- Plaintiff intake through demand. A plaintiff-side platform such as Eve, built as one workflow rather than a tool for each step.
- Questions of the case file with no connectivity, and deposition transcription on the machine that recorded it. An assistant running on the device — AirgapAI for the case file, and AirgapAI Transcribe for the recording, subject to the file-format limit recorded above.
Most firms end up holding two of these at once rather than one, because the cloud platform that wins the contract base is not permitted to touch the material the protective order governs. For more information on how a private deployment is put together inside a firm, visit the AI for law firms page; for getting attorneys and staff productive on whichever tools clear the shortlist, visit the AI training for legal teams page.
Where AirgapAI Is Not the Answer for a Law Firm
The largest legal job in our record is the one we lose. Reading clauses, dates and obligations across an entire contract base is a legal document analysis use case, and Iternal records it as outside AirgapAI's scope, because the models required are bigger than a laptop can handle unless it is an Apple laptop. Iternal records the narrower version just as plainly: the chat product may not be able to do what a contract-review buyer needs. Two further records agree. Laptop-local deployment is not powerful enough for legal work without a mini desktop supercomputer, and local AI, while usable, still cannot do complex knowledge work — with legal and proposals named as the two biggest gaps.
The deposition job carries its own boundary, and it is a file boundary before it is an intelligence one. AirgapAI does not support the AudioNote file extension, so recordings held in that format cannot be imported, and Iternal makes no promise that AudioNote files will ever be importable. Separately, a firm adding a large deposition audio file to the app hit a maximum file size error, and a three-hour recording would not open or play back and displayed an error in the app.
Those two records are not one record, and merging them would lose what each says. The statement that no import promise exists carries no identifiable firm behind it in our evidence; the file-size failure was recorded with an identifiable firm behind it. Kept apart they say something a buyer can use: the format gap is a stated position, and the size failure happened to somebody.
Set beside the adoption record, the account reconciles. Law firms in our conversations chose AirgapAI for their most sensitive information precisely because it does not go to the cloud, and Iternal ships a separate on-device transcription app in the same product line aimed at this work. Both hold at once: the disconnected jobs are won on the device, and the heavy document-analysis job is not.
Pin it down: questions for your evaluation
-
Will you transcribe one of our real deposition recordings, in the format we record in today, before we sign anything?Whether the file-format and file-size limits recorded against AirgapAI apply to the recordings your firm actually holds.
-
What hardware does contract analysis at our volume require, and what happens on the laptops our attorneys already carry?The boundary between what runs on a device and what needs server-class infrastructure, in writing.
-
Which of the jobs we have described does the product cover today, and which sit on a roadmap?Coverage now against coverage promised, before a roadmap becomes a purchase argument.
-
If our case documents are under a protective order, what would this deployment place outside our control?Whether the product may be used on matter material at all, which decides the shortlist before capability does.
What Our Conversations Do Not Tell You
Three cells in the limit column are empty, and the emptiness is the accurate result rather than a gap somebody forgot to fill. An absence of recorded weakness is an absence of evidence, never evidence of absence, and no blank cell here is an endorsement. Each is named below with the reason and the test that would fill it:
- ChatGPT. The one candidate weakness our conversations hold against it names our own product. Strike our product's name from the sentence and nothing about ChatGPT remains, which makes it our differentiation claim rather than a property of the tool, so it is not a limit. The test that would fill the cell: put your retention, training-use and discoverability questions in writing before any client matter goes near it.
- AudioNote. No weakness for it is recorded at all. What our conversations hold instead is our own limit — AirgapAI cannot import its file extension. The test that would fill the cell: export one recording and see which of your other tools can open what comes out.
- Filevine. It appears only as a competing case management system Iternal will not stop firms from using and integrating with, and no identifiable firm stands behind that record. The test that would fill the cell: ask for its integration surface, and what leaves the firm when a matter is synced.
One further exclusion, stated so the standard is visible. A recorded opinion — that somebody judged a product weak, or expects it to fail — names no mechanism a buyer can test, so records of that kind stay out of the table entirely. A limit earns its cell by describing how a product falls short.
How to Test This on Your Own Material
A feature grid sorts nothing in legal work, because the failures happen at the file layer and at volume rather than in the capability list. Three requirements find both, using material your firm already has:
- One deposition recording, in the format you record in today. Not a converted file, not a supplied sample. The format is the test, and it is where our own product has a recorded gap.
- One representative contract set, at your real volume. A ten-document demonstration proves nothing about a portfolio. Ask for the volume you hold, and watch what the answer costs in time and in hardware.
- One question whose answer must name its source document. A cited answer can be checked by a paralegal in a minute. An uncited one has to be redone by a lawyer.
Send those three unchanged to every name in the table, ours included. The replies sort the list faster than any comparison chart — and a supplier that asks you to convert your files first has answered the format question without meaning to.
Where to Go Next
Three questions decide the rest of a legal AI purchase, and each is owned elsewhere. For more information on what an AI deployment may touch when the material is privileged or sits under a protective order, visit the records and privilege page. For more information on what a private AI assistant costs, visit the cost page. For more information on the hardware a document-analysis workload needs, visit the sizing page.
Answered elsewhere
- The jobs legal buyers described in their own words, sector by sector — see the legal use-cases page.
- Scoring a shortlist against criteria you set yourself — see the selection scorecard.
- Reading a contract portfolio nobody in the building has time to read — see the contract portfolio page.
- Where a locally run assistant earns its place beside the productivity suite a firm already licenses — see the existing-suite page.
FAQ
FAQ: Choosing AI for Legal Work
No single product wins, and the split is the answer. Where the document has to sit decides it: an assistant running on the device wins the two jobs that cannot reach a network — querying the whole case file from a courtroom with no wi-fi, and transcribing a deposition on the machine that recorded it — while extraction across a firm's whole contract base goes to a dedicated cloud legal platform. Sort your own jobs by that one question and the shortlist sorts itself.
A legal document analysis use case sits outside AirgapAI's scope, because the models required are bigger than a laptop can handle unless it is an Apple laptop. Iternal records three related boundaries in the same voice: the chat product may not be able to do what a contract-review buyer needs, laptop-local deployment is not powerful enough for legal work without a mini desktop supercomputer, and local AI remains unable to do complex knowledge work, with legal named as one of the two biggest gaps.
Not when they carry the AudioNote file extension — those recordings cannot be imported, and Iternal makes no promise that they ever will be. Two further failures are recorded separately, and they stay separate: a large deposition audio file added to the app returned a maximum file size error, and a three-hour recording would not open or play back and displayed an error in the app. Hand over one real recording, in your real format, before anything is signed.
Because our conversations record no mechanism-level weakness for them. ChatGPT, AudioNote and Filevine carry empty cells, and an empty cell is the accurate outcome rather than an oversight: inventing a weakness for every option would be the worse failure by far. An absence of recorded weakness is an absence of evidence and never evidence of absence, so the test a firm would run to fill each cell is named alongside it.
No. Filevine appears in our conversations only as a case management system a firm intends to keep, and Iternal will not stop firms from continuing to use it and integrate with it. No identifiable firm stands behind that record, so treat it as a starting point rather than a finding: ask for the integration surface, and for exactly what leaves the firm when a matter is synced.
Send three requirements unchanged to every name on the list, ours included: one deposition recording in the format you record in today, one representative contract set at your real volume, and one question whose answer must name its source document. The first two are where products fail quietly, and the third is what makes an answer checkable by a paralegal in a minute instead of redone by a lawyer in an hour.
Legal AI software is built around legal work product — contracts, matters, filings and depositions — rather than around open-ended chat. It divides into four categories: dedicated legal platforms, drafting add-ins that run inside Word, practice and matter management systems, and case intelligence. A general assistant drafts and summarises well; it does not hold the matter record, and it does not settle where a privileged file is permitted to sit.
Start from the job rather than the brand. Harvey is the dedicated legal platform on this page and the strongest fit for extraction across a whole contract base. Spellbook works inside Microsoft Word for drafting and redlining. Clio runs the practice and the matter record. Darrow works case intelligence from public data. Eve is built for plaintiff-side workflows end to end. AirgapAI is the on-premise pick for material that cannot leave the firm.
AI for lawyers divides on where inference happens. A cloud platform transmits the document to be read, which the firm’s confidentiality obligations have to permit before any pilot starts. An assistant that runs on the device answers without transmitting anything, which is why an on-device product carries the courtroom case-file question and the deposition recording here, and why the contract base still goes to a cloud legal platform.
Test the Two Files That Decide It
One deposition recording and one representative contract set will tell your firm more in an afternoon than a quarter of comparison charts. Run them against every name on this list, ours included, and let the failures sort it. We have written ours down already.