Legal AI  is easier to buy and harder to evaluate than it was two years ago. Options now span research and drafting applications, AI built directly into document systems, and general AI platforms connected into legal work.

The model is only part of the decision now. What it can reach, what controls follow the information, whether lawyers keep using it after the first month, and whether any of it produces measurable value increasingly determine whether the technology works in practice.

1. What legal AI means in 2026

Legal AI is artificial intelligence, usually built on large language models (LLMs), applied to legal work under the constraints that make legal work different: privilege, confidentiality, matter context, citation accuracy, and professional responsibility rules.

Capable, secure models are no longer scarce, and that’s true across general-purpose and legal-specific tools alike. What none of them have on their own is knowledge of your firm or organisation: which precedent your practice group actually uses, which matters carry ethical walls, which version of a document is current, or what a client’s guidelines permit. That knowledge lives in the systems you’ve spent years building, and the AI has to reach it somehow.

Legal AI now arrives in three forms. There’s AI embedded in the secure systems where legal work and content already live. There are specialised applications built around particular legal workflows. And there are general AI platforms connected into legal systems and content. Most organizations will run some combination, which makes the useful question less about which category wins and more about what data each one can reach, what controls follow the information once it gets there, and what work improves as a result.

2. Where legal AI is creating value

Four categories account for much of the reported value, though how they rank depends on practice area and organisation size, as well as the data to which they have access.

Drafting and review: First drafts of contracts, memos, and correspondence, plus review of incoming documents against your playbooks or a client’s paper. Most widely adopted in the AI usage surveys, largely because it maps directly onto work already being tracked and billed. Its main dependency is precedent quality, since AI drafting from a disorganized or outdated template library produces polished versions of the wrong content.

Knowledge and precedent retrieval: Surfacing prior work product from your own repository rather than the open web. Results are bounded by what the AI can see and easily surface. A model working across a partial collection will still answer confidently, which is a harder failure to catch than an outright refusal.

Multi-document extraction: Pulling structured information out of a set of documents at once, for due diligence, contract portfolio review, or building a chronology. This is where AI tends to replace hours rather than minutes, and where source citation matters most, since the output is a table someone will act on without opening every underlying file.

Matter and workflow intelligence: Summarizing files, extracting dates and obligations, and answering questions about the state of active work. Interest is growing quickly and it’s the least settled of the four, since understanding a matter is a harder problem than producing a document.

Adoption and spending are moving faster than measurement. Thomson Reuters found 41% of law firms reporting generative AI use in 2026, up from 28% the prior year, and firm technology budgets grew nearly 10% over the same period. Axiom’s 2026 in-house research found that only about 17% of legal teams using AI have established metrics and track return regularly, while every team surveyed expected its AI budget to grow. Plenty of organizations can describe what their AI does. Fewer can describe what changed because of it and the value derived from it. Across all four, the quality and accessibility of the underlying content can matter as much as the model working with it.

3. What actually matters when evaluating legal AI

Most evaluations over-index on the demo, which is the part of the process a vendor controls most tightly. Five things determine whether a platform holds up afterward. They don’t carry equal weight anymore. Capability is close to table stakes, context and control are where platforms genuinely differ, and adoption and value should determine whether any of it survives the first year.

(1) Capability

Can it reliably do the work, and will users be able to tell when it hasn’t?

Fabricated citations are the failure mode with the clearest public record. A public database of court decisions maintained by legal researcher Damien Charlotin held roughly 200 cases in mid-2025 and passed 1,590 by June 2026, with penalties now reaching six-figure sanctions and bar suspensions. Legal-specific tools appear in that data alongside general chatbots, and a Stanford RegLab study found that even paid legal research platforms built on retrieval produced unsupported answers on a meaningful share of queries. Courts have generally been harder on lawyers who couldn’t say what they’d checked than on the use of AI itself.

No credible system removes the need to verify, so verification is the thing to evaluate. How quickly can someone see which source produced a given sentence or citation, open it, and confirm it says what the AI claims? A tool that makes checking take seconds gets checked. A tool that makes it take minutes gets checked until the first real deadline pressure hits.

(2) Context

What can the AI actually reach, and how does it decide what’s relevant?

Access alone doesn’t produce useful context. AI with access to a curated subset of your content performs well in a demo and disappoints in practice. AI with access to everything, including superseded drafts and abandoned matters, surfaces confident answers built on documents nobody should rely on. Relevance, permissions, and recency all have to keep holding as your team adds, supersedes, and reorganises content. Failures here are often harder for users to recognise than failures in model quality. Ask how relevance gets determined, and what happens when the right answer isn’t in scope.

(3) Control

Do your existing permissions, ethical walls, audit requirements, retention policies, and client restrictions continue to apply, or does the AI introduce a second set to keep in sync with the first?

Some of this belongs in the contract: whether your information trains the vendor’s models, where it’s processed and stored, and what happens to prompts and outputs after a session. Some providers also apply automated or human review processes to content flagged by their safety systems, which may carry different retention rules than ordinary usage. Legal teams should understand what triggers those reviews, whether customer content can be included, who can access it, and how long that information may be retained. The rest is architectural and comes down to whether the AI tool inherits permissions and walls from systems you already maintain or forces you to rebuild them inside the tool.

Professional responsibility adds a layer firms tend to hit late. Depending on the tool, how it handles client information, and the representation involved, existing ethics guidance may call for informed consent or disclosure, and general language in an engagement letter may not be sufficient on its own. Client guidelines increasingly carry their own AI requirements, and many organizations haven’t reread theirs since signing, which may have been before AI was a consideration.

Manual movement of documents creates another consideration. When users repeatedly download, upload, or copy governed content between systems, the organisation takes on additional workflow friction and potentially additional copies to manage. At scale, governance cannot depend on every user remembering where content moved or what controls should follow it.

(4) Adoption

Will people use it in six months, and will you know what they’re using it for? Raw usage counts hide more than they reveal. They don’t show what people are accomplishing, whether it connects to anything the business is trying to do, or whether three teams are quietly building the same thing because nobody coordinated.

Most evaluations treat this as a footnote, but it’s where deployments quietly fail. More than half of firms report providing no training on responsible AI use, so a good deal of current adoption is running without support or acceptable use guidance, which puts people in a less confident and much more risky place. Every additional workflow and new system adds friction and task switching, and a better performing tool can still struggle if using it means repeatedly leaving the applications where people already work and results in time-consuming manual confirmation of accuracy. Ask what usage looks like at 90 and 180 days across a vendor’s existing customers rather than at launch, and push for the actual value they are getting beyond simple usage metrics.

(5) Value

Can you show that something improved? Is that tied to the business strategy?

Few teams measure return consistently or think of their business strategy before deploying AI tools, so simply defining success before deployment puts an organisation ahead of most of the market. Pick a baseline you can observe now, such as hours on a recurring document type or turnaround on a category of request.

Value in legal work also can’t be reduced to hours saved. Ethics guidance is clear that time spent learning a tool generally isn’t billable and that an hour of work is an hour on the invoice, so under hourly billing, efficiency can create tension with revenue on exactly the drafting and research work AI compresses first. Depending on the organisation, the case may instead rest on capacity, turnaround, realisation, quality, client experience, or the ability to take on work you previously turned away. Deciding which of those you’re pursuing is part of defining value as it relates to business strategy and client satisfaction rather than a separate conversation about billing.

4. Questions to ask before you buy

Capability

• Can every output be traced to a specific source document?
• How does a user verify an answer, and how many steps does it take?
• What does the system do when it doesn’t have a good answer?

Context

• What can the AI platform access, and what’s excluded?
• How does it determine which documents are relevant?
• What happens when content is added, superseded, or reorganised?

Control

• Is our information used to train models, for us or anyone else, and will you commit to that in the contract?
• Where is our information processed and stored, and does that satisfy our clients’ residency requirements?
• Does the AI inherit our existing permissions and ethical walls, or require a separate model?
• What happens when permissions or walls change after deployment?
• What’s retained after a session, and for how long?
• How is sensitive content determined, reviewed, and retained?

Adoption

• Is this embedded in tools our people already use, or does it require a separate workflow?
• Have people been provided with adequate training and acceptable use guidance to give them confidence and reduce risk to the organisation?
• What does sustained usage look like at 90 and 180 days for your existing customers?
• How are people using the tools and what value are they deriving from them?
• Is coordination ensuring duplicative efforts aren’t occurring?

Value

• What baseline should we establish before deployment?
• What results have your customers measured beyond usage or time saved?
• How does pricing behave as usage grows?

5. Where we land on this

The through queue across all five criteria is that the model matters less than what surrounds it. Capable AI models and tools are becoming widely available and will keep improving. What your organisation knows, the controls governing who can see it, and the impact on your people’s workdays and your business strategy should outlast whichever tools you use to work with it.

That’s the reasoning behind how NetDocuments approaches legal AI: bringing AI to governed content where permissions and ethical walls already exist, while letting that content work securely with the other AI tools legal professionals choose. In practice that runs from natural-language search across the repository to structured extraction with citations back to the source.

The questions above apply to us the same as to anyone, and they’re worth asking whether or not we’re on your list. If you do one thing before your next vendor conversation, pick a recurring task and establish a baseline for how it’s performed today. You’ll have something concrete to hold the technology against.