Key Takeaways
- Before applying AI to a contracting workflow, question the problem, who should own the work, and whether all of that work needs to exist in the first place.
- Standardization, self-service, and rules-based automation can simplify the work that remains before AI ever needs to enter the equation.
- The goal is not to eliminate human involvement, but to apply AI where it adds value and intentionally preserve judgment where it matters.
Remember PEMDAS?
Parentheses. Exponents. Multiplication. Division. Addition. Subtraction.
We all learned this in grade school math. I am pretty sure I have not used PEMDAS since then, except when one of those Facebook memes pops up with the math problem to solve and everyone starts arguing in the comments about the answer. I may not always solve the math problem, but I certainly look at the comments. I digress…
The basic lesson stuck with me though. You cannot just work through a problem in whatever order you want and expect to get the right result. There is an order to how you approach the problem.
I increasingly think we need the same thing for legal AI.
There is so much excitement about what AI can do in contracting. It can review agreements, compare language against a playbook, suggest redlines, summarize changes, extract data, draft communications, and connect those actions into broader workflows. All of that is impressive.
But I think we are sometimes starting too far downstream.
The first question should not be, “How can we use AI to do this?” A better question is, “What are we actually trying to solve?” And we may not know the full answer yet. That is where the order of operations comes in.
Validate Ownership: Should Legal Even Be Doing This?
This has become one of my favorite questions to ask in Legal Operations.
Someone recently mentioned to me that they were waiting to develop a playbook for NDAs. My immediate reaction was: Should your attorneys even be negotiating NDAs?
That changed the conversation.
The visible problem was the lack of a contract playbook. But underneath it was an ownership question. Which NDAs require an attorney review? Which can follow approved standard terms? Which can be handled by the business? Which exceptions genuinely require legal judgment?
Suddenly, the problem looks different.
“We need an NDA playbook” can actually mean too many NDAs are reaching attorneys.
“Contract review is too slow” can mean Legal is reviewing work that should never have reached Legal.
That is why “What are we trying to solve?” is not a question you answer once and move past. You keep asking it.
By the way, if you are in need of an NDA playbook, here’s one you can download and adapt for free!
Eliminate Unnecessary Work: Does This Step Need to Exist?
Ownership asks who should do the work. The next question is whether all of that work needs to happen at all.
Take contract intake. Requests arrive through email, Teams messages, hallway conversations, and an intake form that half the organization does not use. Attorneys spend time chasing down information before they can begin reviewing the contract.
AI could summarize the emails, extract the details, and help route the requests. But why are requests arriving through four channels in the first place?
Which approvals actually affect the outcome? Which handoffs are necessary? Where are we requesting the same information twice? Which steps exist because they serve a purpose, and which exist because that is simply how the process developed?
Sometimes the problem is not that a step is slow. The problem is that the step should not exist.
Standardize What Remains
Once the work is in the right place and unnecessary steps are removed, make the remaining work as consistent as it reasonably can be.
Do we have approved templates? Clear clause positions? Fallback language? Defined escalation points? Consistent intake information? Known paths for common scenarios?
And can some of that work become self-service?
Go back to the NDA example. Instead of asking AI to help attorneys review every NDA faster, create an approved path for agreements that meet defined criteria. Route predictable exceptions appropriately. Send Legal the agreements that actually require legal judgment.
Now we have changed the problem before asking AI to solve it.
Sometimes the better outcome is not helping lawyers process more work. It is making sure unnecessary work never reaches them.
Automate the Predictable
Next, ask: Can a rule handle this? Not every technology problem requires AI. If a decision is predictable and can be expressed reliably through rules, traditional automation may be the better tool.
A workflow can route an agreement based on contract type, geography, business unit, or another defined criterion. A system can trigger an approval when a known condition is met or send a reminder at a particular milestone.
Those decisions require consistency, not interpretation. AI can be tempting because it can do so much. But capability is not the same as fit. If a rule can solve the problem reliably, use the rule.
Apply AI to the Variable Work
Now I want to talk about AI.
What work remains that benefits from interpretation, comparison, synthesis, or pattern recognition?
AI can compare third-party language against a preferred position, identify deviations, summarize an agreement, extract obligations, or explain why language falls outside an established playbook.
This is where AI becomes powerful because we are applying it to the problem that remains after working through the earlier questions. That distinction also changes how we evaluate technology.
Instead of asking a vendor, “What can your AI do?” we can say, “Here is what is breaking down in our workflow. Here is what we have already removed, standardized, or automated. Here is the work that remains. How well can your AI solve that?”
That is a much more useful conversation.
Design the Human Role
There is one more part of the equation: the human.
I keep hearing some version of this promise in legal tech demos: You’ll never have to redline a contract again.
I understand the appeal.
But efficiency is not the only objective.
A playbook can tell AI our preferred positions, fallbacks, and escalation points. It does not eliminate the need to verify that those instructions were applied correctly or to read the language in context.
Let AI do the first pass. Let it identify deviations, suggest language, compare versions, and surface what deserves attention.
Then intentionally decide what happens next.
Where is verification required? What should escalate? Who has authority to accept risk? Which decisions require business context or legal judgment?
That time is not automatically inefficiency. Sometimes it is simply the cost of doing the work well.
As AI gets better, we should be careful not to treat any human involvement as evidence that the workflow has failed.
The promise does not have to be, “You’ll never have to redline again.”
A better promise is, “You’ll never have to start a redline from scratch again.”
Legal AI Needs Its Own Order of Operations
So my order of operations looks like this:
Question the problem. Validate ownership. Eliminate unnecessary work. Standardize what remains. Enable self-service. Automate the predictable. Apply AI to the variable. Design the human role.
Unlike PEMDAS, this is not a rigid formula.
As you work through the steps, you may discover that the problem you started with is not the problem you actually need to solve.
That is the point.
Each step helps you solve the problem, but it also helps you diagnose it. Sequence affects the answer.
If we introduce AI before questioning the work, the owner, the process, and where judgment belongs, we may build a remarkably sophisticated system for solving the wrong problem.
Sometimes the smartest use of AI starts with deciding what AI should never have been asked to do in the first place.
You’ve got this.
Stay tuned for more practical legal ops tips and strategies from the trenches in my column Beyond the Fine Print, right here with Contract Nerds.
