AI legal automation
Adoption stopped being the question in 2026. Building something a professional will sign is still the hard part: privilege, citations that must be real, and reasoning a partner can defend. We build legal AI products for founders and internal tools for firms.
What is actually happening in legal AI right now
What is left is an execution problem, and the numbers show where firms have got to.
Industry surveys through 2026 report that AI use in corporate legal departments nearly doubled year over year, reaching 87% of general counsel from 44%. Firm-level adoption climbed from 26% in 2024 to 42% in 2026, and roughly 70% of attorneys now use AI weekly.
On the market side, legal AI took about $4.3 billion of venture funding across 180+ deals in 2025, and 2026 has already added $1.4 billion across 31 deals. Harvey's near-$11 billion valuation settled whether the category is real.
The market split in two: a few platform players, and hundreds of point solutions attacking one workflow each. Most are built by people who practised the workflow and need an engineering team to make it real.
Where AI is actually working in legal
Not every legal task is equally suited to AI, and the adoption data is unusually clear about which ones are.
Contract work leads by a distance. Around 64% of legal departments using AI apply it to contract drafting, review and analysis, with 63% using it to identify contract clauses. This is the beachhead, and it is where most new products start.
Summarization is close to universal. About 83% are using or experimenting with it, which makes it the single most common generative AI task in legal.
First-pass document review, around 37%, and growing, because it maps cleanly to how the work is already structured.
Transcription at 53% and foreign-language analysis at 40%, both quietly useful and rarely discussed.
Legal research and e-discovery carry the highest stated comfort levels, though at litigation scale the incumbents are entrenched.
The pattern underneath all of it is the same. AI does the first pass; a lawyer verifies every flagged item before it counts. ABA Formal Opinion 512 keeps a named lawyer accountable for anything that leaves the firm, so the products that work are the ones designed around a human checkpoint rather than around removing one. Any product that tries to take the lawyer out of the loop is solving the wrong problem.
The billable hour problem nobody builds for
Most software teams have never heard of it.
Under hourly billing, ethics guidance requires a lawyer to bill the time actually spent. So when AI turns a sixteen-hour task into minutes, as Harvard's Center on the Legal Profession documented, the firm bills less. The billable hour turns efficiency into lost revenue.
Firms on fixed fees have the opposite experience: faster work becomes higher margin. Which is why around 67% of legal departments and 55% of firms expect AI to change how hours are billed, and roughly 71% of clients now prefer a flat fee for an entire matter.
This is a product design constraint. The products that stick either sell to firms already on fixed or capacity-based pricing, or prove value in capacity terms, more matters handled by the same team, not hours saved.
We build with that in mind, and we will ask which side of it you are on before we write a line of code.
The four constraints that make legal AI different
Most AI projects fail on engineering, not on the model. Legal AI fails on four constraints specific to the domain.
What may leave your environment, what may reach a third-party model, and what must never be trained on. An architecture decision made on day one, not a policy document written at the end. Private deployment, self-hosted models, redaction layers and zero-retention configurations exist for exactly this, and which one you need depends on your data, not on fashion.
In most products a hallucination is a bug. In legal it ends careers and draws sanctions. Systems that cite must retrieve from a verified corpus, link every assertion to a source a human can open, and refuse rather than invent when the source does not exist. Refusal is a feature here, and it has to be engineered deliberately.
A lawyer cannot tell a client or a judge that the model said so. Every material output needs a visible chain: what was retrieved, what was reasoned, what a human approved. That shapes the interface as much as the backend.
Where client data lives, how long it persists, who touched it, and whether you can prove all three later. Firms get asked this in security reviews. Founders get asked it by their first enterprise customer.
Build for these four at the start and the product scales. Retrofit them and you rebuild.
Who we build for
You know the workflow because you have lived it, and you need someone to build it properly. Intake, drafting, review, billing, e-discovery, matter management. We build the first production version and hand you a codebase your future engineering hire can take over, with full IP from day one.
You have a workflow eating hours that no vendor tool fits, and buying another SaaS seat is not the answer. We build the internal tool around how your firm actually works, deployed where your data is allowed to live.
What we build
Generation grounded in your templates and precedents, clause extraction, deviation from standard, obligation and risk surfacing, with a human decision point before anything is relied upon.
Search and question answering across matters, contracts and knowledge, every answer linked to its source document.
Structured capture, conflict checks, routing and qualification, so the first hour of a matter is not spent retyping.
The most-adopted use case in legal, built so a reader can always open the underlying passage.
Time capture, matter workflows, and the internal systems that never make it onto a vendor roadmap.
Evaluation harnesses before ship, tracing in production, human-in-the-loop where being wrong is expensive, and the ability to swap models without rewriting the product.
Proof
Legal AI stands on confidential data and professional reliance, and that is the exact ground we build on. Here is where we have shipped it:
AutoIQ. An assistant for mechanics that turns symptoms into ranked, explainable fixes. The closest analogue to legal AI we have shipped: a domain expert relies on the output, so the system has to show its reasoning rather than assert an answer.
Read the case study →Volopay and FLIN. Regulated fintech. Volopay was the first production version of a YC S20 company that has since raised $31M+, operating across 6 countries. FLIN ships regulated financial workflows in Indonesian and Philippine markets. Financial data, regulatory constraint, audit expectations.
Read the case study →Healthcare products handling patient data. Confidentiality as an architectural requirement rather than a policy line.
CampaignHQ. Our own SaaS, built, run and operated by us on AWS. We know what it costs to keep a product alive after launch, because we do it every day.
Read the case study →What it costs
The short answer
A first legal AI build at KUMO runs $20K to $50K over 4 to 16 weeks: one workflow, shipped to production with evaluation and human review built in. Multi-workflow platforms run $50K to $100K. The proof above is the software we have shipped where confidential data and professional reliance are the engineering problem.
$20K to $50K
4 to 16 weeks
One legal workflow shipped to production, with evaluation and human review built in from the first release.
$50K to $100K
16 to 24 weeks
A multi-workflow platform: retrieval over your corpus, several workflows, and the audit and review layer around them.
$5K to $10K per month
Ongoing, cancel with 30 days notice
The engineering team on retainer once it is live: new workflows, model updates, evals, and the security reviews your clients will run.
Every engagement starts with the four constraints above, mapped against your data and your risk position. If the honest answer is that an existing tool already does this, we will say so. For a deeper breakdown of what drives the numbers, see our guide to what it costs to build an AI agent.
Technologies and platforms
FAQ
Contract work leads: roughly 64% of legal departments using AI apply it to drafting, review and analysis, and 63% to clause identification. Summarization is the most common single task at around 83% using or experimenting. First-pass document review, transcription, foreign-language analysis and legal research follow. The consistent pattern is AI doing a first pass with a lawyer verifying every flagged item.
Most expect so: around 67% of legal departments and 55% of firms anticipate changes to how hours are billed, and roughly 71% of clients prefer flat fees for a matter. The underlying tension is that hourly billing requires billing time actually spent, so efficiency reduces revenue, while fixed-fee work turns the same efficiency into margin.
We build AI products where a professional relies on the output and the reasoning has to be visible, most directly AutoIQ for mechanics, plus regulated software in fintech and healthcare where confidential data and audit expectations are the engineering problem. Ask us what maps to your problem and we will tell you straight.
That is your call and your bar's, not ours, and it drives the architecture. We build for whichever answer you give: third-party APIs with zero-retention terms, private cloud deployment, or self-hosted open-weight models on infrastructure you control. Decide it before the build, not after.
By not letting the model answer from memory. Answers are retrieved from a verified corpus, every assertion links to a source a human can open, and the system refuses when no source exists. Refusal is designed in, and evaluation runs before every release to catch regressions.
A first workflow shipped to production runs $20K to $50K over 4 to 16 weeks. A multi-workflow platform runs $50K to $100K over 16 to 24 weeks. Ongoing engineering is $5K to $10K per month.
Yes. A firm usually needs an internal tool deployed where its data is allowed to live; a founder usually needs a product they can sell. The four constraints are the same.
You do, from day one. Source, infrastructure as code, documentation, in your GitHub organisation. Full IP transfer, no licensed components.
No. We build software. You and your counsel decide what is permitted; we engineer to that decision and document what we built so your reviewers can check it.
30 minutes, no deck. Describe the workflow and the data it touches. We will tell you what is buildable, what the constraints force, and roughly what it costs. If the honest answer is that you should buy something off the shelf, you will get that too.
