Industry-Specific AI Agents Move the Moat Above the Model
Industry-specific AI agents are becoming an enterprise operating-model decision, not simply a model-selection exercise. On 25 August 2026, Google Cloud introduced Gemini Enterprise for Legal and a parallel offering for financial services. The legal product combines reusable domain skills, permission-inheriting connectors, agents that execute work, an open partner ecosystem, and centralized governance. Google describes the offer as a preview designed for law firms and corporate legal departments. Reuters independently reported that Google is moving into vertical offerings as competition for legal AI intensifies. The report also noted collaboration with several major law firms. My thesis is that the next durable enterprise AI moat will not be model access. It will be the ability to encode proprietary judgment, permissions, evidence requirements, and escalation rules as governed, reusable execution.
That thesis matters beyond legal services. A general model can summarize a contract, alarm log, or quality report. A production-grade system must know which playbook applies, which repository the user may access, which exception requires escalation, what evidence must accompany the output, and who remains accountable. Domain capability therefore lives in the system around the model. The model supplies cognition; the enterprise supplies context, authority, and consequences.
What Changed: From General Copilots to Vertical Operating Stacks
Google’s architecture makes the verticalization pattern unusually explicit. Its purpose-built “skills” package instructions and context so an agent can apply a firm’s playbooks, citation rules, and house style. Secure connectors use Model Context Protocol and inherit existing user and document permissions. Specialized agents then carry work through tasks such as regulatory scanning, research, contract review, redaction, and drafting. A governed control layer sits underneath, with security policies, private-data isolation, grounding, and traceable citations. These are Google’s stated product characteristics, not independently benchmarked outcomes.
The shift is significant because earlier enterprise deployments often treated domain knowledge as retrieval content: upload documents, search them, and ask a model to answer. The newer pattern treats knowledge as an executable operating asset. A playbook is not merely a document to retrieve; it becomes a versioned package of instructions, constraints, preferred sources, output structure, and escalation behavior. That moves knowledge management closer to software engineering and process governance.
Why It Matters Now: Accountability Does Not Verticalize Automatically
The timing is important because adoption is advancing while professional obligations remain unchanged. On 17 August, the Solicitors Regulation Authority warned that inaccurate output, confidentiality failures, weak supervision, and inadequate governance can harm consumers. The regulator said it received 42 reports related to potential AI misuse between July 2025 and July 2026 and emphasized that individuals remain responsible for their work even when AI is used. The SRA notice is a primary regulatory source and does not claim that all reported cases resulted in violations.
Policy is also moving toward supervised experimentation. The UK government’s legal-services Advisory AI Growth Lab, published on 3 August, brings multiple regulators together to help innovators navigate existing frameworks. The program describes legal services as its first sector focus and aims to support responsible adoption. This creates a productive tension: organizations are encouraged to innovate, but they cannot delegate professional accountability to the agent or vendor. Vertical AI must therefore encode not only expertise, but also the boundaries of that expertise.
The Five-Layer Mechanism Behind Executable Domain Knowledge
A useful operating model has five layers. First, the enterprise converts expert practice into versioned skills: decision criteria, acceptable sources, fallback positions, exception thresholds, and required output structure. Second, identity-bound connectors expose only the records the user and task are permitted to access. Third, the agent orchestrates the workflow across approved tools. Fourth, evidence services attach source references, policy versions, transformations, and tool-call records. Fifth, an accountable professional approves, rejects, or escalates consequential action.
Each layer prevents a different failure. Skills reduce improvisation but can encode obsolete practice. Permission-bound connectors reduce unauthorized access but cannot judge whether access is necessary for the stated purpose. Agent execution reduces handoffs but can accelerate a flawed interpretation. Evidence improves contestability but does not guarantee correctness. Human approval preserves accountability but can become ceremonial if reviewers lack time, context, or authority. Production readiness comes from the combination, not from any single component.
In my view, the design object should be the decision path rather than the chatbot. Leaders should ask: what event starts the workflow, what evidence makes the case eligible, which policy version applies, which actions are permitted, what uncertainty forces escalation, and what record proves the final disposition? Those questions produce an executable operating system for a domain. A polished conversational interface without that backbone remains an assistant, not an accountable enterprise capability.
My Perspective: Manufacturing Should Productize Its Own Judgment
Manufacturers should not copy legal workflows literally, but the architecture transfers well. A maintenance skill could encode asset criticality, approved diagnostic evidence, lockout requirements, spare-parts policy, and escalation thresholds. A quality skill could package control-plan logic, sampling rules, genealogy requirements, concession authority, and customer-notification triggers. A production-planning skill could combine service priorities, constraint rules, changeover costs, and the conditions under which a schedule may be overridden.
My interpretation is that these skills should be treated as controlled operational assets. Each needs an owner, version, effective date, validation set, allowed data sources, tool permissions, and retirement procedure. A plant cannot safely scale agents if its work instructions conflict, its master data is unreliable, or its escalation rules exist only in veteran employees’ memory. Vertical AI makes those organizational debts visible because the system cannot execute a rule the organization has never made explicit.
This also changes the role of subject-matter experts. Their highest-value contribution is not repeatedly answering the same question. It is designing the boundary conditions that let routine cases flow while unusual cases reach the right person with the right evidence. That requires operations, engineering, IT, security, quality, and legal teams to co-author the skill. The result is both technical configuration and institutional design.
Four Strategic Implications
1. Domain skills become a new form of enterprise IP. Models can be replaced; a validated library of playbooks, exception logic, evidence requirements, and workflow outcomes is harder to reproduce. The moat deepens when each completed case improves the skill and validation set without exposing confidential content.
2. Connectors become policy infrastructure. A connector is not merely plumbing to a repository. It carries identity, permissions, matter or asset boundaries, and audit obligations into the agent’s context. Leaders should evaluate connector behavior with the same rigor applied to privileged application access.
3. Knowledge governance converges with software lifecycle management. Skills need testing, change control, deployment environments, rollback, observability, and ownership. A policy update that changes an agent’s behavior should be reviewed like a production release, not uploaded like a new brochure.
4. Vertical AI will reshape operating leverage unevenly. High-volume, rules-rich work is a plausible early target; ambiguous, adversarial, or irreversible decisions require stronger review. I infer that enterprises will gain more by redesigning portfolios of work than by applying one automation percentage to every role.
Counterargument and Limits
The strongest counterargument is that vertical layers can become expensive wrappers around rapidly improving general models. If foundation models absorb more domain capability, specialized skills may have a shorter life than expected. Proprietary connectors and partner ecosystems can also increase switching costs, while “open” does not automatically mean portable. Vendors may describe workflows as agentic even when the result still needs substantial professional checking.
Those limits are real. Google’s announcement does not provide independent accuracy, productivity, error-rate, or return-on-investment benchmarks, and preview availability is not proof of production maturity. The SRA’s warning demonstrates why source verification and human responsibility remain necessary. The appropriate response is not to reject vertical AI, but to measure it at the workflow level: accuracy by task class, evidence completeness, escalation quality, cycle time, exception rate, rework, and business outcome.
Five Leader Actions for the Next 90 Days
1. Select one decision-rich workflow with high volume, bounded authority, reliable source systems, and a clear human owner. Avoid beginning with an enterprise-wide assistant.
2. Convert the current playbook into a structured skill: entry conditions, approved sources, steps, prohibited actions, uncertainty thresholds, evidence, and escalation.
3. Test permissions end to end. Confirm that agent access never exceeds the initiating user, role, matter, plant, product, or purpose boundary.
4. Build a representative validation set that includes normal cases, rare exceptions, conflicting evidence, obsolete documents, and attempted policy bypasses.
5. Establish release governance with a named skill owner, technical owner, risk approver, performance dashboard, review cadence, rollback plan, and retirement criteria.
Conclusion
Industry-specific AI agents will matter because enterprises do not compete on generic intelligence alone. They compete on how consistently they apply knowledge under real constraints. The strategic opportunity is to transform tacit expertise and static documentation into versioned, permission-aware, evidence-producing execution—while keeping humans accountable for consequential judgment. In my view, the winners will not be the organizations with the most agents. They will be the ones whose domain knowledge can be executed, tested, improved, and stopped with discipline.
Frequently Asked Questions
What are industry-specific AI agents?
They are AI systems configured for a particular profession or operating domain through specialized skills, trusted data connections, approved tools, evidence requirements, and governance. The domain system around the model is as important as the model itself.
How are domain skills different from retrieval-augmented generation?
Retrieval supplies relevant content. A domain skill also packages instructions, decision rules, output structure, constraints, and escalation behavior so the agent can execute a repeatable workflow rather than only answer a question.
Can industry-specific agents replace professional accountability?
No. Regulators and enterprise policies can still hold qualified people and organizations responsible. High-impact decisions need explicit review, escalation, evidence, and override mechanisms.
What is the best first manufacturing use case?
Choose a high-volume workflow with bounded authority, strong source data, a stable playbook, measurable outcomes, and a named human owner—for example a constrained maintenance triage or quality-document review process.
References
- Thomas Kurian. “Now introducing Gemini Enterprise for Legal.” Google Cloud, 25 August 2026. Read the original source
- Mike Scarcella. “Google expands Gemini Enterprise AI platform for law firms, lawyers.” Reuters, 25 August 2026. Read the original source
- Solicitors Regulation Authority. “SRA cautions profession about safe and responsible use of AI in legal sector.” SRA, 17 August 2026. Read the original source
- Department for Business, Innovation, Science and Trade. “Advisory AI Growth Lab: legal services.” GOV.UK, 3 August 2026. Read the original source
Related reading
- Enterprise AI Control Plane for Scalable AI Agents
- Human Review as a Safety Layer
- Why AI Pilots Scale Activity but Not P&L Value
Have a question or an idea for a research or industry collaboration? Get in touch with Industry AI Decision.
Leave a Reply