Appendix G — Adopting GenAI in an Organization

G.1 From Assistance to Delegation

Individual assistance is a useful place to begin because it exposes where GenAI actually helps. People discover which tasks are easy to describe, which require extensive context, which outputs are easy to check, and which mistakes recur. Little organizational machinery is required because the user remains inside the loop. If an assistant misunderstands a requirement, the engineer corrects it; if it proposes an implausible design, the engineer rejects it; if it needs a fact about the system, the engineer supplies the fact. Much of the engineering environment remains implicit in the person operating the tool.

Delegation changes that arrangement. The relevant unit is no longer a prompt but a work unit with an intended outcome, available context, permitted actions, acceptance conditions, and a boundary beyond which the agent may not proceed on its own. A useful first organizational step is therefore to identify work that already has a reasonably clear boundary. A maintenance issue with an executable regression test is easier to delegate than an open-ended product decision. A document transformation with explicit preservation requirements is easier to delegate than deciding whether the resulting document communicates effectively to its intended audience. The distinction is not whether the task is difficult. It is whether the organization can state enough about the work and its acceptance to govern the delegation.

This suggests expanding delegation along engineering surfaces rather than assigning an autonomy level to an entire organization or product. One service may have strong interface contracts, architecture rules, deployment checks, and operational evidence while another remains poorly understood. Within the same task, dependency policy may be mechanically governable while an architectural tradeoff remains judgmental. The appropriate boundary therefore depends on the particular obligations at issue: their consequence, recurrence, representability, available evidence, and the reliability with which the environment can evaluate them. Part VI developed this as an economic and governability question rather than a maturity ladder. Figure G.1-1 shows the transition as a moving boundary, not a ladder of levels.

From assistance to bounded delegation: responsibilities move into the environment and the delegation boundary expands Two panels. On the left, under assistance, a worker supplies context, judgment, and acceptance directly to an agent, and a narrow dashed boundary marks the little the agent may do on its own. A wide arrow points right: selected responsibilities move into the environment. On the right, under bounded delegation, a governed environment surrounds the agent — models carry context, evidence establishes properties, and mechanisms constrain and admit results — while the worker governs that environment rather than each act, and a wider dashed boundary marks the expanded authority the agent may exercise on its own. The change is a shifting boundary, not a sequence of maturity levels. FROM ASSISTANCE TO BOUNDED DELEGATION ASSISTANCE WORKER context judgment acceptance AGENT the worker supplies each judgment directly selected responsibilities move into the environment BOUNDED DELEGATION WORKER governs GOVERNED ENVIRONMENT models context evidence properties mechanisms constrain · admit AGENT selected responsibilities move into the environment The dashed boundary of delegated authority expands — a shifting boundary, not a sequence of maturity levels.
Figure G.1-1. From assistance to bounded delegation. Under assistance, the worker supplies context, judgment, and acceptance directly around the agent. As work is delegated, selected responsibilities move into the environment — models carry context, evidence establishes properties, and mechanisms constrain and admit the result — and the boundary of what the agent may do on its own expands. The change is a shifting boundary, not a sequence of maturity levels.

Early delegation should consequently favor work with short feedback loops and independent evidence. Tests, static analysis, schema validation, differential comparison, build results, simulations, and operational measurements can establish useful properties without asking the producing agent whether its own work is correct. The point is not to eliminate human review. It is to learn which parts of review repeatedly establish the same facts and which parts require judgment that the environment does not yet adequately support. That distinction supplies the raw material for the next stage of adoption.

© James C. Davis, 2026–present