A practical playbook for legal teams in the AI decade

Last week I sat down with Gartner’s research group to talk about where law and technology are right now, and where they’re headed. Their project is trying to define the shifts legal departments need to make by 2030. My view is simple: most legal teams can move faster than they think, without waiting for a perfect “AI strategy deck,” and without turning the department into a science experiment. The path is: get the basics right, automate the low-value repeatables, institutionalize learning, and keep your data portable so you aren’t trapped when the tooling landscape changes again (because it will).

Here’s what that looks like in practice.

Start with the boring foundation, then build real leverage

If you don’t have legal intake, contract lifecycle management (CLM), and a functioning repository, you’re going to feel like AI is “promising” but not delivering. Not because the models aren’t good enough—because your department isn’t set up to take advantage of them. The quickest wins tend to appear when you already have clean pathways for work to enter the department, for documents to be found, and for outcomes to be tracked.

In my case at Exos, we’re fortunate to have the foundational software in place, so the question this year has been: where does incremental investment buy compounding returns?

Three categories of tools have mattered most.

First, contract review acceleration. We adopted a standalone redlining tool that can apply our playbook to take a first pass at inbound paper. That matters in real-world workflows like RFP-driven contracting, where you may receive an NDA or similar agreement as part of an RFP package. The tool doesn’t replace judgment, but it does a credible job at: (1) identifying red flags; (2) proposing playbook-aligned edits; and (3) reducing the time to a meaningful markup. The point isn’t “AI wrote my redlines.” The point is that it compresses time-to-first-draft, which is where humans often lose momentum.

Second, legal research. We moved from an existing platform to a different major research product and found it materially more advanced from a “does this get me closer to a real world answer” standpoint. That sounds abstract, but legal research tools succeed or fail on whether they reduce uncertainty. The best ones help you understand not just what the law says, but how the law tends to be applied.

Third, broad access to state-of-the-art models. This is the least expensive line item and the highest ROI. We maintain subscriptions across multiple model providers. Not because it’s trendy, but because different models are better at different tasks, and because the landscape changes quickly. From a budget perspective, this is not the kind of investment that needs a procurement war. From a learning and productivity perspective, it’s a force multiplier.

A rule of thumb for prioritization: revenue vs. non-revenue

One of the most useful mental models for prioritizing automation is to divide work into revenue-generating and non-revenue-generating categories.

If something is repeated, operational, and not tied to revenue generation, I’m comfortable automating it aggressively—even if it occasionally makes mistakes—because “good enough” is often better than “much better but slow.” It’s the classic trade: you can spend expensive legal attention validating low-value work, or you can build a process where the work is mostly handled, and exceptions bubble up.

A concrete example: vendor procurement agreements for routine operational needs. In our business, we operate many physical locations and procurement can involve regular contracting work that is predictable and templatable. In a perfect world, legal might barely be involved. In the real world, we are. Those are the first candidates for automation: structured intake, standard templates, automated first-pass review, and a human only when something breaks pattern.

On the revenue-generating side, the bar is higher, but the payoff is also higher. This is where tools that accelerate contract review and negotiation create direct value: faster turnaround, fewer bottlenecks, less friction in closing business. Here, the goal isn’t to “remove lawyers”; it’s to reserve lawyer time for the parts that actually require judgment, negotiation strategy, and risk tradeoffs.

The overlooked leverage point: “paralegal-grade” automation

A shift that is just starting to click for many teams is what I’d call paralegal-grade automation: high-volume document construction, reassembly, and packaging work that is rule-based and time-consuming.

A recent example: I needed to take 15–20 documents and attach them as schedules to five other documents. Traditionally, that’s tedious. With modern models, it became an instruction-following exercise. I described the task, and the model generated code on the fly to split and merge PDFs, assemble the right packets, and finish the job.

This is more important than it looks. A lot of “legal work” isn’t legal reasoning. It’s document manipulation: assembling exhibits, reconciling versions, extracting key terms, formatting, packaging, and routing. When models reliably do these tasks, the leverage is enormous. Not because it changes the law, but because it changes the throughput of the department.

If you want a practical takeaway: identify one workflow in your team that is mostly document assembly, and automate it end-to-end. You’ll learn more from that than from another round of “AI use case brainstorming.”

How to drive adoption: normalize AI as the first and second proofreader

One of the early adoption steps is unglamorous: set the expectation that LLMs are the first and second proofreader for written work. If someone delivers work product that could clearly be improved by running it through a model (e.g., clarity, structure, tone, completeness) that becomes a coaching moment. The aim isn’t policing. It’s habit formation.

Within a few cycles, this stops being a conversation because it becomes default behavior. People don’t want to hand in work that hasn’t been improved with an available tool, in the same way they don’t want to hand in a document full of typos.

But proofing is table stakes. The bigger lever is creating a culture where experimentation is expected.

Culture matters more than tooling: blameless isn’t enough—mistakes should be encouraged

Lawyers are often described as risk-averse. That’s not wrong. But it’s also not destiny. You can build a culture where reasonable mistakes are not just tolerated, but encouraged because they are a signal that the team is moving fast enough to discover better ways of working.

Most mistakes are never noticed. The ones that might be noticed can often be solved through relationships and responsiveness. In practice, I can’t think of a mistake we’ve made that cost us real money in a way that wasn’t recoverable. That matters because it changes the calculus: if the downside is limited and the upside is compounding learning, you should push the organization toward experimentation.

There’s also a talent-retention angle that legal leaders sometimes underweight. Part of the job of a general counsel is to protect the careers of the people on the team. People leave when they stop learning. AI is an unusually open field where motivated professionals can learn quickly and become more valuable in the market. Encouraging that learning is both the right thing to do and a practical retention strategy.

Hiring: prioritize brains and comfort with technology over “pure legal ops pedigree”

I’ve always taken a technical view of legal operations. Early on, I hired someone with a product management background into a role that had previously been more administrative. That was not a conventional legal ops hire, and it was one of the best decisions I’ve made. A product mindset of problem framing, user empathy, systems thinking and iteration translates extremely well to legal department transformation.

The principle is to hire for intelligence and fearlessness, then aim that capability at the highest-value problems for your department. Domain knowledge can be learned. Curiosity and technical comfort are harder to retrofit.

Transformation case study: adoption is a human problem, not a software problem

We replaced a legacy CLM that was poorly configured and disconnected from intake. The prior state involved manual copying and pasting between a web form and a contract repository, and the system didn’t support smooth matter intake. We moved to a tool that connected intake to the contract repository, supported better metadata extraction, and enabled more self-service for templated agreements.

None of that is novel. Many departments have made similar moves. The difference-maker was adoption.

We tried quick “how-to” videos. They didn’t work. People experience videos as homework. What worked was a human touch: when someone asked for help, we did a 10-minute screen share. Multiply that by the right 40 people—those who submit the most requests—and you get adoption faster than any training library.

The broader lesson is that legal tech adoption is a behavior change problem. The shortest path is personal enablement at the moment of need.

DIY automation: start with small wins that have contained risk and clear truth sources

Low-code/no-code experimentation is often framed as “many lawyers aren’t technical.” That’s true, but it’s not the point. You don’t need everyone to be technical. You need a repeatable pattern for small wins that build confidence.

A good starter example is insurance compliance checks against client requirements. The truth sources are the insurance policy PDFs. All answers are in those documents. You can build a constrained prompt (or a “custom GPT” / “gem”) that answers only with citations to those PDFs and flags uncertainty. Done well, this becomes reliable enough that the human role shifts from “hunt for answers” to “confirm and respond.”

The win isn’t just time saved. It’s the psychological shift: the team experiences that models can be trusted in bounded domains when you control inputs and require traceability.

Sharing learnings: make it social, not bureaucratic

We use a company-wide channel for AI wins. It matters that it’s not confined to legal. Cross-functional sharing produces better ideas because different departments see the same data and workflows from different angles.

We also maintain cross-departmental governance through an AI committee, which is useful when use cases cut across functions and data sources. The key is that the committee enables experimentation rather than slowing it down.

A useful trick: automate the AI tool review process itself

One practical way we reduced friction was to build an AI-assisted intake for AI tool approvals. We use a prompt-driven process where someone can paste a tool’s terms of service into a structured prompt. The output categorizes the risk (“likely bad,” “maybe fine,” “looks good”) and, critically, points to the exact contractual language that triggered the conclusion.

This does two things. It filters out obviously bad tools before legal ever sees them, and it makes legal review faster by focusing attention on a short list of issues that matter when you have no negotiating leverage (e.g., ownership of inputs/outputs, whether the vendor trains on your data, acceptable use, confidentiality, retention).

The “meta” point is important: AI adoption accelerates when legal is seen as an enabler with fast, pragmatic pathways rather than a slow gatekeeper.

Data strategy: insist on openness, because the model “winner” isn’t decided

If there’s one area legal departments are underestimating, it’s data portability.

Many systems, especially CLMs, become proprietary data stores. When your workflow data is trapped behind a vendor’s gateway, you lose the ability to run your own analysis, build your own tools, and switch models or platforms quickly.

My bias is that data should be stored in ways that keep it ready for whatever tool is best this month, next quarter, or two years from now. It’s far too early to declare a winner among model providers. The practical implication is that legal leaders should push vendors to sync data into systems you control (for many companies, that’s a cloud file system or data warehouse). If a vendor refuses, that becomes a strategic risk and, eventually, a reason to move.

Data retention is also going to get messy. AI is upending established instincts about what you keep, what you delete, and what “searchable memory” means inside and outside your organization. Many teams don’t have a mature answer here yet, and pretending otherwise doesn’t help. But the direction should be: reduce silos, increase controlled accessibility, and design retention with the expectation that AI will amplify both value and risk.

2030 and the external market: outside counsel, staffing, and the coming renegotiation of value

When asked what’s underappreciated, I pointed to the inside-versus-outside counsel boundary. In my view, outside counsel is in trouble. Not because clients will successfully “demand” new billing models, but because work will simply move in-house as tools increase leverage.

Top-tier firms will still have a place for bet-the-company matters, board-level risk, and reputational validation. The middle tier is likely to feel pressure sooner because its traditional value proposition is less differentiated and more exposed to automation.

The legal staffing industry is also exposed. A lot of overflow work is routine commercial contracting and document-heavy tasks. That sits in the same landscape as paralegal-grade automation. When those workflows become tool-driven, demand for that labor changes materially.

Looking further ahead, even automated negotiation is plausible: your company’s negotiation agent interacting with a counterparty’s agent, with lawyers operating at the level of risk tolerances and business priorities. This requires the parties to establish a shared model of relative bargaining power, which is no easy task given historical negotiation norms. But if established and coupled with standardized playbooks, it can remove substantial friction from everyday contracting.

Where I’m going next

My intent over the next few years is to keep pushing in three directions.

First, treat AI as a daily utility, not a special initiative. That means proofing, drafting, summarizing, extracting, assembling, and routing work as default behavior.

Second, keep building “small wins” into repeatable systems. Especially where the truth sources are bounded and the risk is contained. Each win builds confidence, adoption, and the muscle memory of improvement.

Third, fight for open data architectures. The department that can flexibly connect its documents, contracts, communications, and institutional knowledge to whatever model is best at the moment will move faster than the department locked into a vendor’s silo.

Thought leadership in this space doesn’t require grand predictions. It requires operational clarity, a bias for learning, and the willingness to redesign legal work around what is now possible. The legal departments that do that will look very different by 2030. Not because they bought more tools, but because they changed how they work.