Model-Agnostic Legal AI: Workflow Control Layer

Technical architecture commentary — not legal advice. The “workflow control layer” is a Nesthing positioning hypothesis and proposed architectural pattern, not an official standard or a statement of deployed Nesthing capabilities. Every integration, control, and provider configuration requires evidence, testing, and firm approval.

AI models are improving quickly, but a law firm’s operating rules should not need to be rebuilt whenever a model changes. The durable architecture separates model capability from workflow authority.

In that design, models can extract, classify, compare, or draft within defined boundaries. A control layer decides what context is available, which tools may be called, where approval is required, and what evidence must be retained. This pattern is an inference from governance needs identified by Singapore’s agentic-AI framework and legal-sector guidance, not a product claim (IMDA Agentic AI Framework, 2026; MinLaw, 2026).

Separate five concerns

A model-agnostic workflow can keep these concerns distinct:

Workflow state: where the matter or task is in its approved process.
Policy: which data, destination, action, and role combinations are allowed.
Model route: which approved model configuration may perform a defined task.
Tool authority: which systems may be read or changed, and within what scope.
Evidence: what must be recorded for review, testing, and reconstruction.

This separation reduces the chance that a prompt becomes the hidden source of business authority.

Define a stable task contract

Each model-assisted step can receive a structured input and return a structured output. The contract can specify required fields, source references, allowed output types, exception conditions, and validation rules.

Changing the model should not change the contract silently. If a new route produces different fields or behaviour, the workflow should fail validation or enter a controlled evaluation path.

Use capability adapters

Different models and tools expose different interfaces. An adapter can translate the stable task contract into a provider-specific request and normalise the response. Provider settings, destination, and evidence should remain explicit.

An adapter is not proof that two models are equivalent. Each approved route needs its own acceptance tests for the tasks and data it may receive.

Keep policy outside the model

The model may assist with classification, but the final routing and permission decision should use approved policy inputs. Unknown data classification, missing authority, or an unsupported destination should produce an exception rather than a best guess.

The same principle applies to tool use. The model can propose an action; the control layer determines whether the action is permitted and whether human approval is required.

Plan for change and rollback

A model update can affect output structure, quality, latency, or tool behaviour. Versioned routes allow a pilot group to evaluate a change before broader use. A rollback path should restore the previous approved configuration without losing workflow state or audit history.

Evaluate the whole workflow

Model-level benchmarks do not establish operational fitness. A controlled evaluation can test source traceability, structured-output validity, exception routing, approval enforcement, tool permissions, retry behaviour, and reconstructability. NIST’s AI RMF likewise treats governance, mapping, measurement, management, documentation, and monitoring as lifecycle activities (NIST AI RMF 1.0).

Results should be tied to the tested version, configuration, task set, and evidence date. They should not be generalised into promised outcomes.

The strategic advantage is not owning a single model. It is preserving the firm’s control as models, tools, and integrations change. That is the role of a legal AI workflow control layer: capability can evolve while authority, evidence, and professional oversight remain explicit.

Evidence boundary

The final paragraph is a strategic proposition for Owner evaluation. It is not an independently verified market outcome, customer result, or statement that Nesthing currently supplies every described control.