For law firms, artificial intelligence has often arrived as a choice between speed and control. Stephen Costigan, founder of Atlas AI, argues that choice deserves a rethink. In this episode of The Geek in Review, we speak with Costigan about private legal AI infrastructure, knowledge graphs, and why a firm’s internal work product may become its most valuable long-term asset.
Atlas AI focuses on turning documents, matter history, precedents, clauses, parties, and obligations into a curated legal knowledge graph inside a firm’s own environment. Costigan contrasts this approach with standard vector search and retrieval systems, which find text with similar language but often lack context around clients, matters, entities, and relationships. A knowledge graph offers structure, linking people, documents, clauses, and legal concepts in ways closer to how lawyers understand their work.
The conversation also explores data quality, a subject with enough baggage to fill a records room. Costigan argues firms no longer need year-long cleanup projects before seeing results. Agent-led curation, entity extraction, duplicate resolution, and ontology mapping reduce much of the manual sorting traditionally associated with knowledge management. Human judgment still matters, especially around practice-area vocabularies and lower-confidence results, but the machines get assigned more of the janitorial work.
Security and governance sit at the center of Costigan’s model. Rather than asking firms to trust a vendor’s assurances around privileged data, Atlas AI runs within a firm’s Azure environment, under firm-controlled keys and policies. Costigan frames this as a shift from confidentiality as a contractual promise to confidentiality as an architectural decision. For legal organizations handling sensitive client information, the location of data, embeddings, audit trails, and model interactions matters as much as the interface lawyers see on screen.
Looking ahead, Costigan predicts a divide between firms renting generic AI tools and firms building durable knowledge infrastructure from their own experience. As routine drafting, diligence, and review work compress, firms with structured and reusable internal intelligence may productize expertise, offer new fixed-fee services, and rely less heavily on traditional leverage models. The future question, Costigan suggests, will not center on which AI tool sits on a lawyer’s desktop. The bigger question will ask who owns the knowledge behind the work.
Listen on mobile platforms: Apple Podcasts | Spotify | YouTube | Substack
[Special Thanks to Legal Technology Hub for their sponsoring this episode.]
Email: geekinreviewpodcast@gmail.com
Music: Jerry David DeCicca
Transcript:
Nikki Shaver (00:00)
Hello Marlene and Greg, coming to you live today from Toronto, Canada. This is Nikki Shaver, the CEO of Legal Tech Hub. I wanted to share something with your audience from the FT Innovative Lawyers Summit last week in London. It was a great event because it brought together people from across legal verticals, including lawyers from law firms, GCs, technologists, people in innovation, and more. One of the panels that really stuck with me was on positive psychology, which emerged as a field in the early 2000s.
A few takeaways from the panel: If you think you can, you can. If you think you can’t, you can’t. There really is something to believing in agency, in your own personal agency. Another couple of things are particularly important as we all look to drive adoption on one hand and increase or maintain engagement among lawyers and employees during this time of unprecedented change and uncertainty in the industry.
Do not let people sit in their little pockets of pessimism. It will spread. Instead, as a leader, one should focus on creating a sense of hope, agency, and a pathway forward. First, create a sense of hope, then provide a vision for the way forward. What is the path forward? Then provide people with a sense that they have agency to drive that path forward.
I love that. I think it is a good thing for leaders to remember at this time, and something for all of us to remember as we encourage people to change the way they work and adopt new technologies, tools, and ways of working. They are much less likely to do so if they do not feel that they have agency themselves. So, leaving you with that today, we will be writing about the FT Innovative Lawyers Summit. Look us up at legaltechnologyhub.com, and you will get a notification when that article comes out. Thank you so much.
Marlene Gebauer (02:26)
Welcome to The Geek in Review, the podcast focused on innovative and creative ideas in the legal industry. I’m Marlene Gebauer.
Greg Lambert (02:33)
And I’m Greg Lambert, and today we are exploring how law firms can harness their own internal data to power the next generation of legal artificial intelligence.
To do that, I am very happy to welcome Stephen Costigan, the founder of Atlas AI. Atlas AI is an enterprise-grade legal AI platform that helps professional services teams transform their internal knowledge into a powerful private legal knowledge graph. So Stephen, welcome to The Geek in Review. Good to have you.
Stephen Costigan (03:08)
Thanks, Greg. Great to be here.
Marlene Gebauer (03:10)
Stephen, can you start by giving our listeners an overview of Atlas AI and what led you to build a platform focused specifically on firm-hosted private legal AI?
Stephen Costigan (03:20)
Absolutely. I’ll start by saying our platform turns a firm’s documents and matter history into a private, curated knowledge graph that the firm owns and runs inside its own environment. It then puts research, drafting, review, and enterprise search on top of that graph.
What led me to do this is that I have spent years building software inside demanding enterprises, including elite law firms. Most of my experience is in the AmLaw 50 space, and I kept seeing the same trap. Every AI tool asked a firm to choose between productivity and control. You either get speed by shipping your most sensitive work to someone else’s cloud, or you get control by building something you can barely operate. For a profession whose entire value rests on confidentiality and privilege, that is not a real choice.
Greg Lambert (04:19)
Yeah, and I have a side question on that. Do you think the foundational companies, Gemini, Claude, and ChatGPT, really understand the legal market and how sensitive the data is that we have?
Stephen Costigan (04:39)
Well, I think a firm should be able to use frontier AI without its knowledge ever leaving its walls. The firm’s knowledge should become an asset that it owns, not fuel for an outside vendor’s model.
Greg Lambert (04:54)
So, Stephen, when you and I were prepping for this and had a couple of conversations, you really got me thinking. I have been thinking about this for a while, but it was good to talk with somebody who is actually doing it about the underlying architecture of the knowledge graph systems you have developed at Atlas AI.
For the non-engineering types listening to this, can you explain what a legal knowledge graph is, why it is a better foundation than dumping documents into a vector database and creating a basic RAG system, and what benefit the knowledge graph provides?
Stephen Costigan (05:42)
Sure. I’ll start with an analogy. A standard vector database is like a clerk who hands you the pages that sound most similar to the question you asked. It is fast, but it has no understanding and very little context. It does not know that the company in one agreement, ACME Holdings, and the company in another agreement are the same client.
So it only knows which pages use similar words. A knowledge graph is the opposite. It is a map of your knowledge where the things that matter, clients, matters, parties, clauses, and obligations, are represented as connected entities. The graph knows that Client A has these matters, each matter has these documents, and each document contains these clauses and parties.
It captures the relationships, not only the text to which it refers. And really, AI is only…
Greg Lambert (06:45)
Yeah, I wrote an entire section of a book on this late last year. It was interesting because, in the legal industry, we all started with vector databases and thought, “This is great because it knows the kind of similar words that we can look for.” It was like the natural-language search we were promised in the early 2000s and that never really worked out.
Stephen Costigan (07:18)
Right.
Greg Lambert (07:24)
It would be great if law were flat, if our business were flat and every document were equal to every other document, or every client were equal to every other client. But it is a complicated batch of information that we deal with. I think even the legal research vendors have finally figured out that it is not flat information. There is nuance, and there are levels to it.
I geek out a little bit when we start talking about adding knowledge graphs to data. But it is interesting that you are doing that with internal information. That is something most people do not think about.
Stephen Costigan (08:01)
Right. AI is only as reliable as what you ground it in. Even with traditional search, adding vector search, adding a knowledge graph, and bringing all of those together into a hybrid search pipeline, you are still not going to achieve the level of accuracy that law requires.
That is why we have added an additional layer, which focuses on bringing the ontology into the index. We bring the firm’s ontology into the index, its information architecture, if you will, in a traditional way of describing it, and then use agents to map that data. At a high level, that is what we are doing, and that is why we are seeing big increases in accuracy when drafting and performing enterprise search.
Marlene Gebauer (08:54)
So, speaking of accuracy, these tools tend to work better when they have clean data. How much work does a firm need to put into cleaning and curating its precedents, templates, and contracts before Atlas AI can generate reliable insights?
Stephen Costigan (09:13)
I would expand this beyond Atlas AI and say far less than people fear. This is the part that many firms get wrong about AI projects. They assume a year-long data-cleaning program is necessary before any value is derived. That is not really the case anymore. Agent-led curation, as we like to frame it, is speeding up the process.
Greg Lambert (09:14)
Is the internet out?
Stephen Costigan (09:40)
Curation can now be automated and continuous. In our platform, a component we call the Librarian runs over every document as it comes in and extracts the entities and relationships. It resolves duplicates, reconciles the same party appearing in different forms, and maps everything to the firm’s ontology. This is a controlled vocabulary for legal concepts, and the graph cleans itself as it grows.
It connects to iManage and SharePoint and builds the structure automatically. What the firm contributes is judgment, not janitorial work. That is what we are removing from the equation. Sorry, go ahead.
Marlene Gebauer (10:21)
Yeah, because you would still have to determine, even if you have the infrastructure in place, which documents are important versus others.
Stephen Costigan (10:32)
Exactly. That is now possible through automation. I would say the firm needs to define and approve the ontology for practice areas and maintain a light human-in-the-loop review queue, where the system can flag lower-confidence data extractions before they are committed to the graph. That allows the firm to govern quality.
The firm does not need to hand-clean precedents anymore. I am not getting into a ton of detail about our product, but at a high level, that is what you are able to achieve now, and it is pretty incredible.
Marlene Gebauer (11:12)
And this will deploy directly into the firm’s infrastructure, such as its Azure environment.
Stephen Costigan (11:21)
Yeah.
Marlene Gebauer (11:21)
I think that is key for large law firms because of the security question. How are you seeing this change the conversation around client data security and firm governance?
Stephen Costigan (11:30)
It is changing confidentiality from a promise into an architecture. That is the best way I can describe it. In a standard SaaS platform, you are trusting a vendor’s contract. You are signing something that basically says, “Trust us with this privileged client data. We adhere to all these controls. We have enough funding. Trust us.”
Greg Lambert (11:48)
All right. Dive deeper into that.
Marlene Gebauer (11:52)
Sounds good. Tell me more.
Stephen Costigan (12:13)
“We can handle this. There will not be a data breach. And we have indemnity clauses to back it up if something happens.” That does not change anything. Client data is exposed. Trust us that we will handle it correctly. Trust that it does not train a model. For a firm with duties to its clients, trust is a weak control.
That has been our thesis since the beginning. We started as a plain private version of ChatGPT, with a few legal twists in our prompt library. Now we have expanded to cover many different focus areas and features. But with Atlas AI, the platform runs inside the firm’s own Azure environment under the firm’s keys.
The data, the graph, the embeddings, and the audit trail, none of it leaves the firm’s environment.
The models are accessed under zero-data-retention terms, so nothing is retained and nothing trains a third party’s system. That flips the governance conversation entirely. Instead of asking, “Can we get comfortable with the vendor’s data-handling procedures?” it becomes, “Can we show our general counsel and conflicts partner exactly where the data flows?” The answer is yes, because it never leaves. That is why firms will run their most sensitive matters on our platform.
They will run them on private AI infrastructure because sovereignty is total in that configuration.
Greg Lambert (13:50)
Interesting.
I have a saying that I use probably a little too much now: Lawyers tend to do better with a red pen than a blue pen. They like to have something to edit rather than create whole cloth. They do their best work when they have a solid first draft to edit and refine.
How do the agentic workflows in Atlas AI provide attorneys with that critical first draft for complex tasks such as due diligence or bulk contract review? What benefits let them dive in much faster?
Stephen Costigan (14:35)
Right. I love that framing. At a high level, we give the lawyer a first draft that has already been argued against by a second, adversarial system. They get to do their best work, the red-pen work, instead of assembly. I can go into more detail if you would like.
Marlene Gebauer (14:55)
Yeah, please do. Go ahead.
Greg Lambert (14:57)
Yep, please do.
Stephen Costigan (15:00)
What I mean is that the system is not there to replace a lawyer’s judgment. It is there to deliver a strong, cited first draft. The lawyer spends time editing and deciding, not assembling.
For due diligence, an attorney points the system at a deal document set, and it produces a diligence checklist or an issues list. Every line item carries an inline citation to the exact subsection from which it came, not simply Section 3, but Section 3.2(a).
For bulk contract review, you define the question once, such as change of control, governing law, or termination, and the system extracts structured answers across hundreds of documents into a reviewable grid. Each cell is traceable to the source. The part that makes the draft trustworthy is an adversarial verification step. One model drafts, and a separate model is tasked with arguing against it, checking that the deliverable does what was asked and that every citation is precise.
That happens before the lawyer or anyone else sees it. So the red pen the attorney picks up is editing a verified draft, not catching the machine’s mistakes or, even worse, having someone else catch them.
Marlene Gebauer (16:20)
There is always a question of build versus buy, and that continues in the market. What do you think is the primary differentiator for firms that say, “We are going to choose to build our own capabilities. We are going to build our own knowledge model, relying on our own data rather than a centralized vendor model?”
Stephen Costigan (16:50)
My thoughts are, number one, you are not going to build a differentiated practice area in your firm or maintain differentiation by buying the same product and using the same data set as everyone else.
Number two, data is the firm’s greatest asset. The signals that come off that data are being extracted. Why are you giving your data away or training another organization’s environment on those signals? But the primary differentiator is not a feature. It is ownership.
And with the… Sorry, go ahead.
Marlene Gebauer (17:33)
No, go ahead.
Stephen Costigan (17:51)
I am really backing up the points that I made. In the centralized vendor model, you are renting access to their product, interface, roadmap, and data-handling promises.
Your knowledge ultimately improves their system. With our platform, or with private AI infrastructure in general, you build a curated graph that you own outright. That compounds with every matter you work on and bring into the platform.
You can extend and build on it. You can create differentiated products for your firm, and you can do that more easily now, especially with recent advancements in agentic AI. Three things follow from that.
First, the ontology is yours and editable. You govern how your knowledge is structured, not some vendor. Second, because it is structured and resolved, the asset becomes more valuable over time rather than being consumed and forgotten after each query, as in a RAG model. Third, because it runs in your environment, you are never exposing client data to build someone else’s moat.
The crowded part of the market is selling chatbots and tooling that sit on top of a model using the same closed pattern. The durable position is helping a firm own its intelligence and knowledge infrastructure. The model layer is becoming a commodity. The curated graph is part of the defensible asset, and it should belong to the firm.
Marlene Gebauer (19:22)
What would you say?
Greg Lambert (19:23)
Is there a certain type of expertise that firms need to maintain this? If we are going to use a third party and rely on its infrastructure, is there a different type of expertise that we need internally to maintain our own version of that infrastructure? I hope that question made sense.
Stephen Costigan (19:56)
Yeah, it does. In terms of requiring an entire team to manage an infrastructure like that, I think that in the near future it will not be as much of an ask to build a part of your organization that can manage private AI infrastructure.
I think existing knowledge management roles can be adapted to the curation and management of that aspect of the environment. I say that because of the advancements in agent-based infrastructure management. For instance, in our environment, we have what we call the Enclave. It is an agent environment where pretty much everything in our infrastructure is managed automatically, in a highly governed way.
It is not that difficult to roll your own infrastructure-management agents. When we look at the amount of code written agentically in our organization, we have gone from around 10 percent to 60 or 65 percent of our platform being written agentically. Everyone thinks they are going to need a huge team to manage their own private infrastructure, but that is simply not the case anymore.
Greg Lambert (21:20)
Yeah. I know a lot of us are looking at the Kirkland advertisements for GPU professionals and seeing that kind of build-it-almost-from-scratch, nearly-on-your-own approach. With Atlas AI, I am not going to have to hire people to stand up GPUs and monitor them, right?
Stephen Costigan (21:42)
Exactly.
Marlene Gebauer (21:43)
I had a question because you mentioned before that the legal AI market is incredibly saturated, which it is. But I am not sure the same is true from a knowledge management perspective. There seems to be a more limited set of tools that deal specifically with that. What do you think differentiates those types of tools?
Stephen Costigan (22:10)
I do not think there are many highly effective knowledge management platforms for law firms. Again, the team size is typically one to five people in that particular area of the firm, if they are lucky. I do not think there are really any solutions out there to compare to, honestly.
Marlene Gebauer (22:13)
Correct. Mm-hmm.
Stephen Costigan (22:36)
But I think what is critical is automating the curation of the DMS. If you can do that, you have achieved a great deal for a firm at the outset.
Marlene Gebauer (22:48)
Yeah, that is what I am saying. There seems to be a limited number of tools that say they do that. The ones that are out there still say you need to offer the system something to start with or clean up the data. I was curious whether you see other differentiators between tools in that space.
Stephen Costigan (22:55)
Yes. The thing is, it is not just about curating the data. It is about what happens after that, too. There are many paths one can take as a founder building a product. The path we selected was to get away from building features to compete with everyone else on the end-user side of the product and to enable the democratization of feature development.
You can take Claude Code, for instance, use our MCP, access your curated data set, and build whatever applications you want, either separate from or integrated into our platform. That is something I think is truly unique.
It takes firms away from having to follow the vendor roadmap and allows them to start building differentiated products for their firm immediately.
Greg Lambert (24:05)
Stephen, before we get to our crystal ball question, we have been asking guests to share some of the ways they keep up with the market. There is so much to try to keep up with. Do you have any resources you do not mind sharing that help you stay current on the transitions in technology?
Stephen Costigan (24:17)
We really rely on Legal Tech Hub for market signals, and Artificial Lawyer.
From a knowledge graph and data architecture standpoint, the foundational knowledge graph work by Hogan and colleagues is great. Those are the high-level sources. Ethan Mollick’s Applied AI is helpful for understanding how professionals actually adopt these tools. Those are my recommendations.
Greg Lambert (24:39)
Yeah. I have an Artificial Lawyer story to tell on this. I was on Richard’s podcast a few weeks ago, and I was in my Austin office when one of the attorneys came up and said, “I heard your podcast interview.” I said, “Which one?” They said, “The one you were on with Richard.” I said, “Well, have you ever listened to my podcast?” They said, “No, I have not listened to that yet.”
Stephen Costigan (25:20)
Exactly.
Greg Lambert (25:27)
So Richard does a pretty good job.
Marlene Gebauer (25:29)
Yes.
Stephen Costigan (25:30)
Absolutely. I try to read across two lanes that I do not know if they always talk to each other, but I think that is changing, and we want to drive that change. Those lanes are the legal innovation world and the knowledge representation world. The interesting work is at the seam between them. That is something no one is really attacking right now, so it is something that really excites me.
Marlene Gebauer (25:55)
Yeah, I think that is spot on.
It is time for a crystal ball question. Looking ahead, do not be scared. We will only come back next year to see whether you were right.
Stephen Costigan (26:03)
I have no idea what that is, so I am scared.
Greg Lambert (26:10)
I am just guessing.
Stephen Costigan (26:12)
Okay.
Marlene Gebauer (26:24)
What is the single biggest shift that you see coming for the traditional law firm business model?
Stephen Costigan (26:30)
Ownership. The biggest shift is that a firm’s accumulated knowledge stops being a byproduct and becomes its most leveraged asset. That really breaks the math on which the billable hour rests. The traditional model monetizes leverage.
Partners sell associate hours against precedents that live in people’s heads and scattered files today. That is changing, obviously. When that precedent becomes a curated graph, routine production work compresses dramatically.
The firms that own that graph will deliver senior judgment with far less junior leverage. They will be able to productize their expertise through fixed-fee, on-demand, client-facing offerings in ways the hourly model never allowed. The divergence I see is ownership.
Marlene Gebauer (27:34)
It is ownership and usability.
Stephen Costigan (27:36)
Usability. Firms that treat AI as a tool they rent will compete on price against everyone renting the same tools. Firms that build and own their own knowledge graphs, curate their data into those graphs, and use them effectively will turn their expertise into a durable, compounding asset. That becomes the real moat for the firm, not head count. The losers will be those who gave their knowledge away and do not have much to show for it. So there is my crystal ball response.
Marlene Gebauer (28:06)
Okay, kids, you heard it here. Do not give your knowledge away.
Greg Lambert (28:10)
It is surprising that you have to tell people that, is it not?
Stephen Costigan (28:10)
Ha ha.
Just to land this point, in five years the question will not be which AI tool your firm uses. It will be whether your firm owns its knowledge or rents access to it from a bunch of knowledge providers. That is the line on which the next generation of firms will be drawn.
Greg Lambert (28:37)
Yeah, I think that is a pretty solid prediction. Stephen Costigan from Atlas AI, I want to thank you for taking the time to join us, break this down, and geek out with us on knowledge graphs and private AI. I really appreciate it.
Marlene Gebauer (28:40)
Yeah, me too. Mm-hmm.
Stephen Costigan (28:52)
Thank you so much. Thanks for having me.
Marlene Gebauer (28:55)
Yeah, thanks, Stephen.
Thanks to all of you for listening to The Geek in Review. If you enjoyed the show, please share it with a colleague. We would love to hear from you on LinkedIn and Substack.
Greg Lambert (29:06)
And Stephen, where is the best place for listeners to learn more about you and what you are doing at Atlas AI?
Stephen Costigan (29:14)
Sure. You can find me on LinkedIn, Stephen Costigan, Stephen with a PH. I am happy to talk with any firm thinking about owning its AI rather than renting it. The site is atlas-ai.io.
Marlene Gebauer (29:29)
And as always, the music you hear is from Jerry David DeCicca. Thank you, Jerry, and goodbye, everybody.