data sovereignty

Agentic Features Won't Fix Shared-Cloud Legal AI's Custody Problem

Agentic AI features don't change where privileged documents live. Here's the architectural difference between shared-cloud legal AI and true data custody.

RAGbase Legal Research TeamNovember 16, 2026 9 min read

In November 2026, the Agentic AI Infrastructure report's "State of Play" survey asked a deceptively simple question: when a legal AI platform adds autonomous agent capabilities, does anything change about where the underlying documents live? Across the platforms reviewed — the major practice management and legal research incumbents among them — the answer was consistently no. The agents got smarter. The infrastructure underneath them did not move an inch.

That gap matters more than most firms realize, because the marketing has quietly conflated two unrelated things: what the AI can do and where your data sits while it does it. A platform can ship genuinely impressive agentic features — multi-step research, autonomous drafting, workflow orchestration — while leaving the exact same custody model in place: full client documents, indexes, and logs sitting on vendor-controlled, multi-tenant cloud infrastructure, governed by the vendor's terms of service rather than the firm's own data processing agreement.

The Conflation: Capability Upgrade, Custody Unchanged

Agentic AI vendors have every incentive to bundle "more autonomous" with "more secure" in the same release notes. It reads better. But the two are architecturally independent. Adding an agent orchestration layer to a SaaS platform is a software update to the application tier — it does not, by itself, touch:

  • Where the vector index and embeddings are hosted
  • Which organization's cloud account holds the raw document store
  • Whose logging and retention policies govern query history
  • What legal entity is named as data processor in the DPA

The report's central finding is blunt: of the mainstream legal AI platforms reviewed — spanning practice management, legal research, and dedicated AI-native tools — every one still processes privileged work product on vendor-operated cloud infrastructure, regardless of how agentic the front-end experience has become. Clio's agentic features, LexisNexis's Protege expansions, and Thomson Reuters's CoCounsel agent workflows all sit on the same custody foundation that existed before agents arrived. The agent is new. The server it runs next to is not.

This isn't a criticism of agentic capability itself — multi-step legal research and autonomous drafting are genuinely valuable. It's a correction to how that capability is being positioned relative to a separate and much older question: who controls the infrastructure processing privileged material.

Why This Distinction Gets Blurred

Three forces are driving the conflation, and understanding them helps explain why so few procurement conversations surface the custody question before signature.

1. Agentic demos are visually compelling; infrastructure diagrams are not. A live demo of an agent chaining five research steps and producing a memo draft is a better sales moment than a data flow diagram showing document storage location. Buying committees respond to what they can watch happen.

2. Vendors genuinely believe their security posture is strong — and it often is, relative to generic consumer tools. SOC 2 Type II, encryption at rest, access controls — these are real and meaningful. But they describe how well-protected the vendor's cloud is, not whether it's the vendor's cloud at all. A bank vault is still a vault even if it's not yours.

3. "Enterprise-grade" has become a synonym for "safe," which has become a synonym for "fine to send privileged work there." The report notes that 68% of firm respondents could not correctly identify whether their primary AI platform's retrieval index was hosted on vendor or firm infrastructure — not because the information was hidden, but because the question had never been asked internally.

What "Processing on Vendor Cloud" Actually Means for Privileged Work

When a firm uploads a matter file to a mainstream legal AI platform — even one with sophisticated agentic features — a predictable sequence occurs:

  1. The full document set is transmitted to the vendor's cloud environment.
  2. It's chunked, embedded, and indexed into a vector store the vendor operates.
  3. Every query against that matter re-reads from the vendor-hosted index.
  4. Logs of what was asked, retrieved, and generated are retained under the vendor's data retention schedule — not the firm's.
  5. If the vendor's agent calls an external LLM API to generate output, that call is governed by the vendor's contract with the model provider, not the firm's.

At no point in that sequence does the firm's own data governance policy — its retention rules, its conflicts protocols, its privilege log practices — actually control the data. The firm is a tenant. This is the pattern our your-data-is-their-moat analysis traced across the practice management category broadly; the agentic layer added in 2026 doesn't change the underlying tenancy model, it just adds a more capable application on top of it.

For non-privileged, low-sensitivity workflows, this may be an acceptable trade-off. For privileged work product, conflict-sensitive matters, or regulated-industry clients with their own data residency obligations, it's a custody question the firm often hasn't actually decided — it's been decided by default, by whichever platform procurement picked for other reasons.

The Architecture That Actually Changes Custody

There is a real alternative architecture, and it's not "avoid LLMs entirely." It's about where the heavy infrastructure sits relative to where the thin model call happens.

Shared-Cloud Agentic AI vs. Firm-Controlled Agentic Scaffolding

LayerShared-Cloud Platform (agentic features added)Firm-Infrastructure Deployment
Full document corpusStored on vendor cloud, multi-tenantStored on firm infrastructure (on-prem or firm-controlled cloud tenant)
Vector index / embeddingsVendor-hosted, vendor-controlledFirm-hosted, firm-controlled
Permissions & access controlsVendor's identity system, vendor RBACFirm's existing IAM, ethical walls enforced locally
Query & agent logsRetained under vendor's scheduleRetained under firm's own retention policy
Agent orchestration layer (connectors, workflows)Vendor-operatedDeployed on firm infrastructure
What reaches the LLM providerOften full context windows, sometimes full documentsOnly minimized, retrieved chunks relevant to the specific query
Governing terms for the model callVendor's contract with the model providerFirm's own API agreement with the chosen provider
Data residency controlVendor's regions, vendor's choiceFirm's choice of region and provider

The critical row is the second-to-last one. This is the honest version of the architectural argument, and it's not "we never touch the cloud" — RAGbase Legal also calls out to commercial LLM providers when a firm chooses to. The difference is what leaves the firm's walls and under whose terms.

The Minimized-Chunk Principle

In a properly scoped retrieval-augmented architecture, a query doesn't trigger a bulk document transfer. It triggers:

  • A retrieval pass against the firm-hosted vector index (stays internal)
  • Selection of the specific passages — typically a few hundred to a few thousand tokens — relevant to the question
  • Transmission of only those passages, plus the query, to the model provider via an API call the firm itself has contracted
  • A response that's logged back into firm-controlled audit infrastructure, not vendor infrastructure

The full matter file — the other 400 pages that weren't relevant to that specific question — never leaves the building. The agent's reasoning, its tool calls, its memory of prior steps in a multi-step workflow — all of that orchestration logic, the part that's actually "agentic" — runs on infrastructure the firm controls. Only the minimized evidentiary slice needed to ground a specific answer crosses to the model provider, and it does so under an API agreement the firm negotiated, not a platform's bundled terms.

This is a meaningfully different exposure profile. A privilege log entry for "document transmitted to third-party AI processor" looks very different when the transmission is a few retrieved paragraphs under the firm's own data processing terms versus an entire file set under a SaaS vendor's standard contract.

Where the Pattern Shows Up Across the Market

The report's findings track a pattern visible across categories, not a single vendor's failure:

  • Practice management platforms (Clio and peers) have added agentic task automation — drafting, intake triage, matter summarization — while document storage architecture remains the same multi-tenant cloud model that predates the agent features by years.
  • Legal research platforms (LexisNexis's Protege, Thomson Reuters's CoCounsel) have extended agentic research chains that pull from proprietary case law databases and, increasingly, firm-uploaded documents — with the uploaded material processed on the same vendor cloud that hosts the research product.
  • AI-native platforms (Harvey, Legora) and general-purpose agentic assistants (Claude Cowork, ChatGPT with agent features) each bring real capability gains, but the same question applies: is the firm's full document set sitting on infrastructure the firm doesn't control, governed by terms the firm didn't write? Our claude-cowork-privilege piece walks through this specifically for consumer-grade agent tools entering firm workflows through the back door.

None of this means these platforms are insecure in an absolute sense. It means agentic capability and architectural control are two different procurement questions, and the market has been answering only one of them in its marketing.

A Due-Diligence Checklist for Managing Partners and CIOs

Before the next agentic AI renewal or pilot, five questions separate marketing claims from actual architecture:

  1. Where does the full document corpus physically reside — vendor cloud account, or infrastructure the firm controls?
  2. Who operates the retrieval index — and can the firm export or delete it independently of the vendor relationship?
  3. What specifically is sent to the underlying LLM — the full document, the full context window, or a minimized, retrieved chunk?
  4. Whose contract governs the model call — the vendor's agreement with the model provider, or the firm's own negotiated API terms?
  5. Who controls log retention and audit trails — the vendor's default schedule, or the firm's records policy?

Firms that can't answer question 3 with specificity are, by definition, unable to represent to a client or a court exactly what left the building during a privileged workflow.

Where This Leaves the Buying Decision

Agentic features are not the problem — they're the most useful development in legal AI since retrieval-augmented generation itself, and our agentic-ai-law-firms analysis covers why multi-step autonomous workflows are worth adopting. The problem is treating "the platform added agents" as equivalent to "the platform solved data custody." Those are separate upgrades, and only one of them has shipped broadly across the market reviewed in the November 2026 report.

The firms getting this right aren't rejecting commercial LLMs. They're separating the two layers deliberately: keeping the document corpus, retrieval index, permissions, and workflow orchestration on infrastructure they control, and sending only the minimal chunks required to answer a given question to the model provider of their choice, under their own terms. That's the architecture behind private AI deployment — agentic scaffolding on firm infrastructure, with case search and retrieval running locally, and only the thin, necessary slice of context reaching the model.


If your firm's next AI renewal conversation is framed entirely around agentic capability, it's missing half the procurement question. The more useful starting point is our AI for law firms guide, which walks through how to evaluate agentic features and data architecture as two separate line items — because right now, most vendor pitches are only answering one of them.

Frequently Asked Questions

Does adding agentic AI features to a legal platform change where privileged documents are stored?
No. Agentic capabilities — autonomous research, drafting, or workflow orchestration — run on top of existing infrastructure. If a platform stored full client documents on vendor-controlled multi-tenant cloud before adding agents, it still does after, unless the vendor has published a specific architectural change to custody, indexing, or retention.
What's the difference between 'agentic AI' and 'data sovereignty' in legal tech?
Agentic AI describes what the software can do autonomously — chain tasks, call tools, execute multi-step workflows. Data sovereignty describes where your documents, indexes, logs, and permissions physically and legally reside. A platform can be highly agentic and still process everything on vendor-owned cloud infrastructure outside firm control.
Is it possible to use commercial LLMs like Claude or GPT-4 while keeping privileged data on firm infrastructure?
Yes. A properly architected retrieval layer keeps the full document corpus, vector index, permissions, and audit logs on firm-controlled infrastructure, and sends only the minimal retrieved text chunks needed to answer a specific query to the model provider under the firm's own API terms — not the vendor platform's terms.

Related Articles

R
RAGbase Legal Research Team
Research

RAGbase builds private AI systems for law firms: deployed on the firm's own infrastructure, zero data retention, full ownership.

See How RAGbase Works on Your Data

30-minute call. We scope your use case and show the system live.

We use audience and marketing cookies (Google Analytics, LinkedIn). No tracker loads without your consent. Learn more