Epona

Menu

Prompts are cute, but agents change the game.

By Bart van Wanroij

AI in 2026 is
all about building better pipelines

Let’s be honest about something. Most legal professionals tried Copilot. Some were impressed and many were underwhelmed. A lot quietly went back to Word, Outlook and “the way we’ve always done it.” Not because AI doesn’t matter. But typing a clever prompt into a chat window never really felt like legal work. It felt like experimenting and was rarely transformative. That phase is over. As Bart van Wanroij puts it: “Prompting is 2025 and an agent is now.”

Let’s start with saying that the shift is bigger than it sounds. Because prompting is an individual skill and agents are a team capability. And that’s where the real change begins. In this article I will dive deeper into this.

“Suddenly, Copilot stops feeling like an intern you must brief every time.
It starts behaving like a junior associate who understands the playbook.”

The limitation of prompts

A prompt is essentially a one-off instruction. You tell the model what to do, maybe reference a document, and hope the result is good enough to refine. That works for summaries, comparisons and maybe for quick drafts. But legal work isn’t a single-step activity. It’s layered and contextual and structured. Often repetitive in shape, even if the facts differ.

Take a typical corporate legal scenario: a formal claim arrives, or a demand letter or a structured complaint that requires a response. You don’t just summarize it. You:

  • review the factual context
  • check the applicable contract or policy
  • look at prior similar cases
  • confirm internal position guidance
  • apply a structured response format
  • ensure nothing violates internal risk boundaries


That is not a prompt but a workflow. And workflows are exactly what agents are built to handle.

What an agent really is

Forget the jargon for a moment. An agent is not “a smarter prompt.” It is a predefined pipeline.
It guides the AI through a sequence of steps and tells it where to look and what sources to trust. But also, what structure to apply and what standards to check against. And last but not least, it tells what the final output must resemble. Instead of asking the system to “please write a response,” you give it a repeatable process:

  1. First, read the incoming document.
  2. Then summarize the relevant claims.
  3. Then compare against internal precedents.
  4. Then review the applicable contract or policy language.
  5. Then generate a structured draft response.
  6. Then validate that response against internal quality criteria.


Now imagine that process being reusable and shareable, improved over time.

Why Copilot felt weak and why that’s changing

Earlier complaints about Copilot weren’t entirely wrong. It often felt too generic or too unaware of context. But that wasn’t a model problem but more an instruction problem. Copilot is powerful, but it needs direction. Without context, it guesses. With context, it becomes precise. Agents solve the blank-page problem by embedding context into the workflow itself. The user no longer needs to remember what to ask or how to phrase it. The agent already knows:

  • which SharePoint location to consult
  • which KMS documents to prioritize
  • which templates to apply
  • which boundaries to respect

Suddenly, Copilot stops feeling like an intern you must brief every time. It starts behaving like a junior associate who understands the playbook.

The hidden power: standardization without rigidity

One of the most underestimated benefits of agents is quality consistency. Legal teams don’t struggle because their best lawyer produces weak work. They struggle because quality varies across people, time pressure, and volume. Agents create a baseline. Bart described testing AI-generated responses against the strongest expert in a team. The verdict wasn’t “this is better than me.” It was: “This is not as good as I would write it, but it’s better than many drafts we see internally.” That’s what we can call a structural uplift. And when you refine the agent over time, incorporating senior feedback into the workflow, the baseline keeps rising. The result is controlled consistency and in legal, consistency reduces risk.

The generational advantage

There’s another insight that shouldn’t be ignored. The strongest agent implementations don’t come from tech alone. They come from collaboration. The 25-year-old who understands how to build and iterate quickly and the 55-year-old who understands what “good legal work” actually means. Agents become powerful when senior judgment shapes the pipeline and younger fluency accelerates iteration. Without that mix, you either get technically clever but legally naïve workflows, or brilliant legal minds that never scale their thinking. Together, they build institutional intelligence.

The bigger shift

Agents don’t replace legal professionals, but they encode best practices. They guide junior staff; they reduce repetition and enforce standards invisibly. And: they turn AI from a novelty into infrastructure. Which brings us full circle to something Marcel Lang emphasized earlier: AI becomes strategic only when it connects to your internal knowledge and governance environment. And agents are the bridge.