Building Emma: The Architecture Behind Reliable Legal AI

Pieter started in AI more than 17 years ago, when it “was really just a fluke.” 

He did a PhD, added an MBA to translate research into value, then shipped real products. At Chatlayer in 2019, his team was among the first bot platforms to use transformers and to support true multilingual understanding. After an acquisition by Sinch, he led global machine learning and later ran an innovation lab. The startup itch brought him back to build Emma for legal teams.

"Alongside Emma, he runs mlconference.ai, a leading European ML conference, and trains DPOs on AI. Another reason why security and governance sit so highly on Emma’s agenda."

The inflection points shaped his view. In 2012, deep learning beat traditional computer vision. In 2022, ChatGPT showed what better training signals could unlock. 

Why legal documents suit LLMs

Legal language is specific, structured, and repetitive in the right ways. Clauses map cleanly to topics. Expectations can be defined in “golden standards.” That lets LLMs extract, compare, and summarize with clarity.

Due diligence is the sweet spot. It is largely “looking for specific information in documents and then comparing that with what you expect.” LLMs do not tire, can run in parallel, and reduce weeks of review to hours while keeping lawyers in control of judgement.

The fine-tuning debate is yesterday’s news

Early legal AI players trained bespoke models on firm data. Pieter’s team took a different path. Foundation models have absorbed enormous volumes of public legal text, and the gap has closed.

“Writing a good prompt is faster than fine tuning,” he notes, “and it lets you adopt new models sooner.” Off-the-shelf leaders improve monthly. Prompted well and grounded in the right workflow, they deliver better output without the latency and cost of constant re-training.

“You can stay ahead by adjusting prompts, not retraining models.”

Horizontal platforms vs Point solutions

Generic chat interfaces look flexible, but they push the work back on lawyers. Even legal-focused platforms often stop at a table of extracted facts. You still need to decide what is a liability, configure dozens of checks, and even ask yourself if you are sure that all the required documents are in the data room to start with..

For M&A lawyers, breadth is the problem. A platform built to handle litigation, IP, criminal, public law and M&A cannot mirror the diligence workflow tightly enough. The result is setup drag and partial time savings.

“If you want to save real time, tailor for one task and cut it into steps.”

Agentic workflows that you can trust

There are two “agentic” paths. Free-form agents plan their own steps, wander the data, and hope to land on an answer. They make great demos when the happy path holds. The problem is they also go off the rails.

Emma’s approach is different. The team pre-defines each step in the diligence process, then picks the right model for that step. Tasks run sequentially, with narrow prompts and checks that are easy to audit. Where generic agents guess, this system shows its work and stays within guardrails.

Recent experiments, including internal work at Apple, show “thinking” models only outperform standard models when the task is fuzzy or under-specified. When the task is clear and well-scoped, regular models win. “Thinking models only help when tasks are vague,” Pieter says. “Diligence is not vague. Define the steps and solve them.”

How the Emma workflow runs

Pieter’s team mapped how firms actually work, then went deeper.

1. Get the data
   Connect to the virtual data room or secure storage and ingest the set.
2. Check completeness
   Match the information request list against what is present and flag gaps.
3. Scan for liabilities
   Run focused checks across the entire data room, not just by folder label, so addenda or misplaced exhibits are still found.
4. Aggregate findings
   Roll up document-level issues into topic-level findings like IP transfer or restrictive covenants.
5. Report red flags
   Assemble a buyer-ready summary with links back to source passages.

Under the hood, each phase is split into smaller, verifiable steps, with model selection tuned per task.

Security you can actually show a DPO

Deal data never trains models. Processing happens in the EU on trusted clouds with written “no training” guarantees. The platform was built to ISO 27001 standards from day one and validated through pen testing with external experts.

Pieter’s bias is personal as well as professional. He trains DPOs on AI risk and cheerfully calls himself “a data privacy nerd.” He even nudges colleagues to use Signal because “I don’t want my data to be used by other companies.”

What changes next

Model performance will keep rising, especially in vision and multimodal tasks like image and video. Language models will improve too, but in “baby steps.” Prices will continue to fall as providers squeeze more performance out of smaller models. That shift favors prompt-and-workflow systems that can swap in better models quickly.

The constant, in Pieter’s view, his method. Strictly define tasks. Use the best model for each. Keep humans in the loop for judgement.

An M&A story from the COVID lockdowns

When Pieter sold his previous company in 2020, the deal closed as Europe entered lockdown. Banks still demanded in-person signatures. The team had to escalate through a major bank to find the right approver and invent a compliant workaround. 

It was a reminder that process can be the riskiest dependency; and that robust, adaptable workflows are not a luxury.

One last piece of advice

If you are exploring AI yourself, cut work into tightly scoped subtasks and run each in a separate chat. You will reduce hallucinations and improve precision tenfold.

If you are buying, look for tools that do the same thing under the hood. “Use agentic workflows, not free-form agents,” Pieter says. “If you are buying, look for tools that do the same thing under the hood. “Use agentic workflows, not free-form agents,” Pieter says. “Pick the solution that gets you from data room to red flag report with the least friction, supported by human-in-the-loop review so the speed never compromises judgment.”