This is the second article in our AI adoption series. Read the first, Why Legal AI Adoption Stalls Before It Starts, for how organizations move from scattered AI experimentation to a structured approach.

A better way to find the legal problems worth solving with AI

Ask a room of lawyers, legal technologists, consultants, or vendors for AI use cases and you will not leave with a shortage of ideas.

Contract review. Matter summaries. Legal research. Drafting. Due diligence. Knowledge search. Billing review. Chronologies. Client intake. An AI tool can generate another 50 before the meeting is over.

The hard part is no longer coming up with things AI could do. It is figuring out which problems are actually worth solving.

Use cases still matter. They show what is possible, spark ideas, and help organizations learn from what others are doing. The mistake is treating the list as discovery.

Start with “Where can we use AI?” and almost every manual process can become an AI opportunity. Identify work that creates friction, consumes time, or leaves important information out of reach, and the conversation changes.

Look first at work people should no longer have to do. Then look for information, risk, knowledge, or relationships the organization should be able to see but cannot reasonably see today. Then ask what context, access, and authority a solution needs.

That is a much tougher test than generating use cases, and it is more likely to lead to something worth building.

Start with the work that should disappear

Some of the best opportunities are hiding in work people have stopped noticing. You see it when someone:

• Searches several systems because they do not know where the information lives
• Asks a colleague for something that technically exists somewhere else
• Copies information from one system into another
• Checks two sources because neither is trusted completely
• Recreates a document because finding the original will take too long
• Maintains a personal spreadsheet because the official system does not provide the view they need

None of these activities sounds particularly significant on their own. That is partly why they persist.

Look for the recurring behaviors: searching, moving, checking, reconciling, recreating, asking, chasing. Then ask why the person has to do them.

If a lawyer repeatedly struggles to find prior work, the answer might be better search, metadata, filing practices, or knowledge management. If staff repeatedly re-enters the same information, two systems may simply need to exchange data. If someone is manually assembling information from several sources because understanding it requires comparison, synthesis, or judgment, AI may have a meaningful role.

The useful question is not “Can AI do this?” It is “Why does someone have to do this at all?”

Be willing to decide that AI is the wrong answer

A good AI discovery process should eliminate use cases as well as create them. Some problems are better solved through configuration. Others need automation, integration, process redesign, better training, or technology the organization already owns.

Take a legal department manually assembling information from several systems for a regular status report. It looks like an obvious AI opportunity.

Maybe it is.

But if the work consists mostly of moving structured data into one place, integration or reporting may solve the problem more reliably. If someone has to read narrative updates, resolve conflicting information, identify what changed, and explain what matters, AI may add something different.

The distinction sounds simple. In practice, it is easy to skip because once an organization starts an AI initiative, there is pressure to find AI projects.

That is backwards.

The problem should earn the technology.

Pay attention to the boundaries that already exist

There is another useful signal when evaluating an opportunity: where does the organization start to become uncomfortable?

Many legal organizations are already more comfortable with some AI activities than others, even if they have never written down exactly why. Summarizing a document a lawyer has already opened feels different from searching across thousands of matters. Creating an internal draft feels different from communicating with a client. Suggesting a next step feels different from taking it. Reading information from a system feels different from changing the information inside it.

Those differences tell you something. The concern may be the sensitivity of the information, the audience, the consequences of an error, the ability to see what happened afterward, whether an action can be reversed, or how much authority the organization delegates.

Those are not reasons to stop connecting AI to legal work. They are design inputs.

From context to consequence

Context creates value. Access creates capability. Authority creates consequence. This distinction becomes more important as legal AI moves beyond a tool that waits for a prompt.

Model Context Protocol (MCP), APIs, integrations, and agentic systems can allow AI to work with information across the legal technology environment and, increasingly, do something with it.

Context comes first. What does the AI need in order to understand the work correctly? A contract without the client’s standards, a document without its matter, or an isolated piece of work product without the surrounding history may be available to the system and still lack what it needs to be useful.

Then there is access. Which systems and information sources does the solution actually need to reach? Connectivity can remove a lot of unnecessary human work. A lawyer should not have to repeatedly download governed information, move it into another application, perform the work, and then reconstruct the context simply because the systems do not connect.

Ask what the use case requires.

Authority is different again. Retrieving information is not the same as changing a record. Recommending an action is not the same as executing it. Drafting a communication is not the same as sending it. As AI-enabled workflows gain more ability to act, human review, monitoring, and the ability to reverse an action become more important.

Give the solution the context it needs, the access required to do the job, and the authority appropriate to the consequences.

The second discovery path: what should we be able to see?

Removing unnecessary work is only half of the opportunity.

The other half is harder to spot because nobody is performing the missing task. The information simply is not visible.

Legal organizations accumulate enormous amounts of knowledge across matters, contracts, correspondence, work product, negotiations, advice, and client interactions. Yet important decisions still rely on only the portion of that information a person can find or happens to know exists.

A lawyer may remember that the firm handled a similar issue several years ago but not know which matter or who worked on it. A corporate legal team may have negotiated the same provision hundreds of times but have no practical way to see patterns across those negotiations. A practice leader may know the firm has deep experience in an emerging area without being able to quickly see the matters, people, and work that demonstrate it.

These are not inefficient workflows in the usual sense. Nobody necessarily needs a faster button. The problem is an incomplete view.

What should the organization be able to see that it cannot reasonably see today?

AI may help surface relationships, patterns, knowledge, or risk that would be difficult for a person to identify across a large body of information. The value is giving the person making the decision a better view.

What better visibility can look like

Security is one example. Existing security systems determine and enforce access, monitor activity, and apply established controls. AI should not replace those functions.

But imagine a situation where policy technically permits an action and the surrounding circumstances are still unusual. The question may no longer be simply, “Was this allowed?” It may be, “Is there enough context here that someone should take a closer look?”

That could be a useful AI-assisted visibility problem, with a person making the security judgment.

Relationship intelligence offers another example. Law firms already have CRM and relationship-intelligence technology. Calling that “AI CRM” does not create a new idea.

The more interesting question is what the organization knows that the structured record does not show. Matter history, correspondence, previous work, and other institutional content may reveal experience, expertise, or relationships that are difficult to see from contact data alone.

Legal AI creates information too

This is the part of AI design I think deserves much more attention.

We spend a lot of time thinking about what information AI consumes. We spend far less time thinking about what happens to what AI creates.

Suppose an AI-enabled workflow produces a useful matter chronology, an analysis of a negotiation, a client briefing, or a summary of prior work. Where does it go? Can another lawyer find it six months later? Does it remain connected to the matter and the source material that informed it?

Can someone tell what the AI generated, what a person reviewed, and what became final work product? Does the output become part of the organization’s institutional knowledge, or does it disappear into an individual’s AI workspace?

AI-enabled workflows are becoming another source of substantive work product. Where that work goes afterward needs to be part of the use-case design.

The DMS has a role here because legal teams already use it to connect work to the matter and preserve the context around that work. Valuable AI-generated output raises the same question: how does it stay connected to the legal work it came from?

Otherwise, an organization can use AI successfully to extract knowledge from its institutional content while creating a new body of work that nobody else can reliably find or reuse.

Do not create a new information-management problem to solve an AI problem.

Change the discovery conversation

Use-case lists are useful. Just use them later.

The discovery conversation can come down to three questions:

• What work should disappear? Look for the searching, moving, checking, reconciling, recreating, asking, and chasing that people have learned to accept as part of the job.
• What should we be able to see that we cannot see today? Look for decisions people make without the knowledge, patterns, risks, relationships, or context the organization already possesses.
• What context, access, and authority does the solution actually require? Decide what the system needs to understand, where it needs to reach, and what it may do.

To put these questions into practice, use the AI Opportunity Discovery Worksheet to identify potential opportunities, understand what is creating the friction, and determine whether AI is actually the right intervention.

Sometimes it will be. Sometimes a simpler solution will be better. Either answer is useful because the purpose of discovery is not to produce more AI projects. It is to identify problems worth solving.

That answers one question: Did we find the right problem?

Then comes the next one: Is solving it this way actually worth the investment?

In the next article, we’ll look at how to run an AI Value Review to answer that question before moving from a promising idea to an AI investment.