Coheso Team
Coheso Team
Legal AI is becoming easier to access. The harder question is whether legal departments have the systems, context, and measurement needed to make it work at scale.
Legal departments spent 2024 writing policies to restrict AI. Much of 2026 has been spent working out what to do with it.
Association of Corporate Counsel research tracks the reversal: US in-house use of generative AI roughly doubled in a single year, while the share of organizations with policies against AI fell from close to three in ten to under one in ten. Thomson Reuters’ Future of Professionals Report 2026 found that 74% of professionals now use AI tools several times a week and 44% use them multiple times a day.
The harder question is what that adoption has produced.
AI for legal means applying language models and agentic workflows to the work a legal department actually does: capturing and triaging requests, reviewing contracts, answering routine questions, drafting first passes, and producing the data that shows where legal time goes.
Whether its deployment is successful depends on what sits underneath it.
The measurement gap
Access is no longer the main constraint. Execution is.
Axiom’s 2026 In-House Legal AI Report found that only 7% of teams have scaled AI across the organization. Eighty-three percent cannot measure whether their AI spending is working, and every respondent said they intend to increase their AI budget.
The tools reached individuals before they reached the function. A lawyer tries something on a Tuesday, saves an hour, and keeps using it. Multiply that across a team and you get twelve versions of the company’s position on limitation of liability, none written down, each one of them potentially leaving when someone resigns.
The problem is not simply that teams have too many tools. It is that the context, standards, and decisions that make those tools useful are not consistently captured.
Start with intake
Before anything gets reviewed, it has to arrive. And most of it arrives badly.
A forwarded thread with no context. A Slack message that says “quick question.” A form filled in by someone guessing at what legal needs.
Capturing requests consistently is often a lower-risk place to start because the model is primarily interpreting and routing the request, rather than making a substantive legal judgment. More importantly, it creates a structured record of what the department was asked to do, by whom, and in what category.
That record is the beginning of a system of record for legal work.
A legal front door converts scattered inbound work into that record. The mechanics are covered separately in this guide to modern legal intake.
Then triage
Classification and routing come next.
Getting a request to the right lawyer with the right context attached requires accurate reading of what was asked, not legal judgment. Where positions are documented, routine questions can also be answered at the point of request.
This is where the load first visibly changes. Some requests never needed a lawyer in the first place. Others need legal judgment and should reach the right person with the relevant context already attached. The goal is to make sure routine questions can be resolved through approved guidance, while requests that require legal judgment reach the legal team quickly.
Intake and triage do more than move work. They establish the structure that makes the rest of the legal AI stack more useful.
Then the rest
Contract review has the strongest evidence behind it. The documented gains concentrate on high-volume paper where the same positions recur across hundreds of documents: NDAs, standard vendor agreements, and routine DPAs.
Teams that sequence by volume, starting with their most standardized contract type, tend to see results within weeks. Heavily negotiated and novel agreements benefit as well, with the value showing up in speed to a reviewed first pass rather than in volume handled.
Research and drafting sit behind that. ACC research puts efficiency among the most commonly cited benefits of generative AI among in-house teams, with drafting and legal research among the leading tasks.
Summarizing a long agreement or a messy thread for a business stakeholder is a small, constant win that rarely appears in a vendor demo, but lawyers mention it unprompted when asked what they actually use.
The point is not that intake replaces these use cases. It is that a consistent record of requests, context, and outcomes gives them a stronger foundation.
Why rollouts stall
A pilot succeeds because one motivated lawyer runs one tool on documents they know well, supplies the missing context from their own head, and notices when the output is wrong.
The pilot measures them, not the system.
Rollouts stall for four reasons:
- No shared standard. Quality tracks whoever ran the review.
- No captured context. Every review starts cold and returns something generic that a lawyer then has to redo.
- No record. Nobody can reconstruct what the AI proposed and what the lawyer changed.
- No measurement. A department that cannot count what it received has no denominator to prove anything against.
Thomson Reuters’ Future of Professionals Report 2026 puts a number on what structure is worth. Where a clear AI strategy is in place, 66% of professionals say AI meets or exceeds expectations. Where there is none, the figure is 22%. The tools in both groups are broadly the same.
Governance is part of the infrastructure
Legal AI governance cannot stop at a policy stating that employees should use approved tools and verify outputs.
Those principles matter, but they need to be translated into the workflow itself.
A legal team needs to know what AI can handle independently, what requires lawyer review, which information sources it can use, when a response must be escalated, and how decisions and sources are recorded.
Courts and regulators have repeatedly made clear that using AI does not remove a lawyer’s professional responsibility to check the result. The fix is not simply a better policy. It is infrastructure that makes the approved way of working easier to follow.
What to ask before the demo ends
Four questions are worth asking before evaluating a legal AI platform:
- How does the system learn the department’s positions, and does that knowledge belong to the department or to whoever configured it?
- Can guardrails be set per category of work, given that risk tolerance on an NDA might differ from tolerance on an important MSA?
- Can a decision be reconstructed six months later, and can the audit view be shown rather than described?
- Does the platform produce the numbers the CFO will ask for?
These are not just product questions. They are questions about whether the department is building a repeatable operating model for AI.
Proving the return
Value shows up in time saved, reduced external spend, and an audit trail that satisfies a regulator or an insurer.
Among the Axiom respondents who measure returns, 67% track error reduction, 56% internal labor cost reductions, 51% hours saved per task, and 46% reduced outside counsel spend.
But none of it is provable without structured intake.
That is why so many legal AI programs end in an unresolved argument about whether they worked. The savings were probably real. The evidence was never collected.
Coheso’s ROI calculator is a starting point.
The next phase of legal AI
The teams pulling ahead do not necessarily have better models.
They have a function that agreed on a standard, wrote its positions somewhere the tools can read them, decided in advance what AI is allowed to finish, and kept a record.
They have built the infrastructure around the model.
That starts at intake.
If your legal team is already experimenting with AI, the next question is not simply which tool to add. It is whether you have a consistent way to capture work, preserve context, apply standards, and measure outcomes.
That is where an AI-native legal intake and work management system becomes useful.
To learn more and see how intake lays the groundwork for AI success, book a demo with Coheso.
Frequently asked questions
What is AI for legal?
AI for legal is the use of language models and agentic workflows to support the core work of a legal department: intake and triage of incoming requests, contract review and redlining, research and drafting, self-service answers to routine questions, and analytics on legal demand.
Which AI use case should a legal team start with?
Intake and triage are a strong place to start. Capturing and classifying requests is often lower risk than substantive legal work, and it produces the structured data that other use cases depend on.
Contract review has the clearest time savings, but works best once a consistent record exists to apply it against.
Why do legal AI pilots stall after the pilot?
A pilot measures one lawyer supplying the missing context themselves. A rollout measures a system.
When intake is unstructured, each lawyer configures the tool differently, institutional knowledge does not accumulate, and nobody can prove what the investment returned.
The difference is not just the model. It is the context, standards, records, and measurement around it.