Intelligence Foundry

The Intelligence Foundry: How Elite Professional Services Firms Are Building Firm-Owned AI

By David H. Deans

An Intelligence Foundry is a pipeline that turns the by-products of daily work — emails, meetings, negotiation threads, collaboration patterns — into a governed knowledge graph, uses that structure to synthesize a training curriculum, and fine-tunes specialist small language models that reason over a firm's own proprietary judgment rather than merely retrieving fragments of it.

It's the architecture already replacing generic retrieval-augmented generation (RAG) in the law, consulting, and corporate-finance firms moving fastest on AI — and it's what the industry's "Sovereign Intelligence" pitch actually describes underneath the marketing: not just AI that's secure and isolated, but AI that knows what makes one firm different from its competitors.

I previously wrote about why frontier AI training had moved inside the enterprise: the open web no longer supplies the cognitive density that frontier models need, and the next phase of competitive advantage runs through an organization's unwritten expertise, not its public data.

This analysis follows that thread into the sector where it matters most acutely, and names the firms and vendors actually building it.

Why This Isn't Just Better RAG

The dominant pattern of 2023–2025 — a frontier model bolted onto a vector database — has a structural ceiling. It retrieves chunks of text ranked by similarity, stitches them together at query time, and offers no persistent model of who decided what, when, or why. 

A governed knowledge layer works differently: it models entities, relationships, and decisions before retrieval happens, so an AI agent can answer questions a flat document index never could — what was decided, what's top of mind for a team, which obligations are still open.

The performance gap between the two approaches is now measured, not just argued. In first-party engineering studies, ontology-grounded retrieval has shown average answer-quality gains in the 36 percent range over conventional RAG, with the advantage widening sharply — roughly doubling — on questions that require synthesizing three or more sources, exactly where flat vector search degrades worst.

Ontology-backed agents also produced "excellent" responses at more than twice the rate of non-ontology agents in controlled comparisons. These figures come from vendor-run studies (chiefly Microsoft's Azure AI Foundry and Fabric teams) rather than independent audits, so I'd treat them as directionally credible rather than settled fact — but the direction is consistent across every source I reviewed.

The Flagship Case: Kirkland & Ellis and Palantir

The clearest public signal that this architecture has moved from thesis to capital allocation is Kirkland & Ellis's roughly $500 million, multi-year AI investment program, a portion of which funds a proprietary "Fund Formation Engine" built with Palantir.

The engine is designed to manage private-equity fund documentation, side-letter drafting, obligation tracking, and closing commitments across the firm's fund-formation practice. Kirkland selected Palantir specifically for its ontology-modeling capability, and has described the goal in explicit terms: centralizing and compounding the expertise of its most senior lawyers so it's usable across more than 1,000 attorneys firm-wide.

What Kirkland's own materials do not confirm is whether the arrangement extends to firm-owned model weights, as opposed to a firm-owned ontology sitting on top of Palantir's platform. That distinction matters more than it sounds: a knowledge graph is portable and durable; a fine-tuned model generally is not.

Firms that build the graph first and negotiate weight ownership second are making the more defensible long-term bet.

Kirkland has reportedly paired this with a second, separate build — a litigation-focused engagement with Syllo — suggesting a repeatable pattern: acquire a commercial substrate, secure exclusive build rights, layer proprietary knowledge on top, and repeat per practice area.

The Wider Applied-AI Vendor Landscape

A handful of other players are building pieces of the same stack, each with a different answer to the question of what the firm actually gets to own:

Harvey has scaled to a large footprint across AmLaw 100 firms and in-house teams, and in August 2026 shipped Tenet, its first in-house legal reasoning model, post-trained on a third-party open-weight base. Harvey's stated ambition is to let each firm eventually train its own version on its own lawyers' working patterns — but the offering is not yet live with named firm deployments, and outside coverage has cautioned that the benchmark claims deserve scrutiny since they're vendor-graded.

Intapp frames its Celeste product around a sharp version of the sovereignty pitch: every firm has access to the same underlying models, so the only real differentiator is the firm's own deals, clients, and prior judgment. Intapp's approach encodes that judgment as firm-authored "playbooks" rather than fine-tuned weights — firm-owned logic on a shared model layer.

Microsoft has assembled the most complete commercial version of a unified context layer, spanning collaboration-signal mining (Work IQ), an enterprise ontology layer (Fabric IQ, still in preview), and a managed retrieval service (Foundry IQ) built on Azure AI Search. The architecture is real and shipping, but the most differentiating pieces are not yet generally available, and Microsoft's own economics for the stack are measured mainly through Copilot's broader ROI studies rather than isolated to the knowledge layer itself.

iManage, Glean, and Hebbia occupy the substrate layer beneath many of the above — governed document and collaboration graphs that other vendors integrate with — without themselves offering firm-owned weights.

Twin1, a recent entrant, is pursuing passive capture at the individual level: building a "digital twin" of each professional's working patterns directly from email, meetings, and documents, with early legal-industry customers including Linklaters, Orrick, and Dechert.

Across this landscape, only Palantir currently makes an explicit, first-party commitment to customer-owned model weights. Everyone else is selling governed and isolated — which is a meaningful security posture, but not the same claim.

The Underlying Science: Ontologies, Curricula, and Two Kinds of Collapse

The technical case for this approach isn't limited to legal-tech marketing. A 2026 Princeton preprint on domain-specific superintelligence argues that general-purpose training is structurally insufficient for deep domain expertise, because expertise requires composing simple domain concepts into complex ones — exactly the structure a knowledge graph provides.

The researchers demonstrated the mechanism in medicine: they synthesized roughly 24,000 reasoning tasks from a medical knowledge graph, verified each one against the graph itself, and used the resulting curriculum to fine-tune a mid-sized open model that outperformed both open-source baselines and larger frontier models on in-domain reasoning.

The paper is explicit that the method generalizes to other domains — which means the limiting factor for law, consulting, and finance isn't the training technique, it's whether a firm has a reliable, firm-specific ontology to train from.

This raises the two failure modes any firm considering this path needs to understand, because they are not the same risk:

Model collapse is the degradation that occurs when synthetic training data replaces real data across generations, gradually losing the tails of the original distribution. Current research suggests this is avoidable, not inevitable — accumulating synthetic data alongside real data, rather than substituting for it, keeps the error bounded.

Knowledge collapse is the more dangerous risk for an elite firm specifically. It's the tendency of any centralized AI assistant to average differentiated expertise toward a generic middle, because large models naturally regress toward the statistical center of their training distribution.

For a firm whose entire value proposition rests on judgment competitors don't have, routing everything through a shared, vendor-hosted model risks quietly erasing the differentiation clients pay a premium for. This is close to Kirkland's own stated rationale for building rather than buying: off-the-shelf tools trained on broad market knowledge tend to converge on a lowest common denominator.

The Real Applied-AI Constraint Isn't Technical

Every credible account of this market points to the same conclusion: the bottleneck is economic and contractual, not technical.

Roughly 90 percent of legal billing still runs through the hourly model, and knowledge-sharing is rarely factored into partner compensation — which means any foundry design requiring senior partners to spend unbilled hours labeling data or writing playbooks is fighting the firm's own incentive structure.

Outside counsel guidelines increasingly go further, explicitly barring firms from billing for AI-driven time savings, which means the ROI case for this technology has to be made through win rate, new productized revenue, or capacity redeployment — never through hours saved.

The winning designs I found all solve for this the same way: capture expertise passively, from signals firms already generate (email, meetings, document collaboration), rather than by scheduling expert time. Microsoft's Work IQ, Twin1's digital twins, and Harvey's in-flow document retrieval are all variations on the same answer — mine the exhaust of normal work rather than asking anyone to stop and teach the AI tool.

There's a second, quieter constraint worth naming for anyone in a general counsel or CFO seat: model-weight ownership is genuinely unresolved contract territory. Legal commentary on AI vendor negotiations already flags the specific terms to fight for — whether a client gets equivalent rights in resulting model weights, and whether a vendor can sidestep that by training on a "roughly equivalent" synthetic dataset instead of the client's actual data.

No public litigation has tested this yet; it's still being fought in term sheets, which is exactly when it's cheapest to get right.

What This Means for Intelligence Foundry Buyers

For a general counsel, CFO, or managing partner evaluating this category, three practical conclusions follow from the evidence:

The ontology is the asset, not the model. A firm-specific knowledge graph is portable, inspectable, and durable across model generations. Weights, by contrast, are usually locked to a single AI vendor relationship. Build or acquire the graph first, and negotiate weight rights as a separate, explicit contract term.

Any design that depends on un-billed partner time will stall. The billable-hour economics of professional services make voluntary knowledge-labeling programs a poor bet. Favor architectures that mine existing collaboration signals over ones that ask experts to stop and document.

Consulting is the open competitive space. Every named case study in this market — Kirkland, Harvey, Intapp's flagship clients — sits in legal or corporate finance. No AI vendor has yet published a comparable sovereign-intelligence case study in management consulting, which makes it the highest-opportunity white space for a firm willing to move first.

A Note on Market Research Sources

This analysis draws on primary vendor materials (Palantir investor relations, Microsoft Learn and Azure AI Foundry documentation, Kirkland and Harvey press materials), a 2026 Princeton pre-print on domain-specific super-intelligence, peer-reviewed research on model collapse, and reporting from Reuters, Bloomberg Law, and the Thomson Reuters Institute, current as of August 2026.

Several figures — particularly vendor-reported ROI and engagement metrics — are self-reported or vendor-controlled studies rather than independently audited, and I've flagged those inline rather than presenting them as settled numbers. Given how quickly this market is moving, I'd treat anything here tagged as "not yet generally available" as subject to change within a quarter or two.

Interested in what a Proprietary Intelligence Foundry strategy would mean for your organization?

Learn more > Contact us now