Assist
Summarize, search, draft, classify and prepare information.
Human remains the decision maker and executor.
Grounding · data access · quality review
ARTIFICIAL INTELLIGENCE
Artificial Intelligence
From assistance to execution: operating models, controls and economics for AI in regulated environments.
The relevant question is no longer whether a model can generate a useful answer. The question is what work it can change, what decisions it can influence, what it can execute safely and whether the resulting economics justify the operating risk.
AI creates value when it changes how work is executed, decisions are made and economics move.
Assist→Recommend→Decide→Execute→Control
FROM ASSIST TO EXECUTE
Not every AI system should have the same autonomy
The operating model changes as AI moves from producing information to taking action. Control should increase with the consequence of the decision — not with the novelty of the technology.
Recommend
Propose next-best actions, prioritization or decision options.
Human accepts, rejects or changes the recommendation.
Evidence · explainability · threshold monitoring
Decide
Select an action inside defined policy and materiality boundaries.
Human designs policy, reviews exceptions and remains accountable.
Decision rights · policy constraints · exceptions
Execute
Call systems, update records, initiate workflows or complete authorized transactions.
Human supervises material exceptions and can interrupt or reverse execution.
Identity · permissions · limits · rollback
Control
Observe outcomes, detect anomalies and trigger escalation.
Human owns risk appetite, thresholds and incident response.
Monitoring · audit trail · kill switch · incidents
The more AI moves from language to action, the less useful “human in the loop” becomes as a generic control. What matters is where human judgment is actually required and where policy can be encoded safely.
APPLIED AI
Four assets show how AI changes the operating model
The work is organized around execution, governance and measurable operational change — not around model capability alone.
01Operating model
AI agents, unpacked
The real opportunity is not better conversation. It is better execution.
Problem
Fragmented channels, interrupted origination, low commercial productivity and repeated service work.
Intervention
A network of specialized agents coordinating commercial, service, origination and early-collections activity.
Control
Defined mandates, authorized rules, escalation boundaries and technical execution controls.
→
Outcome
Higher continuity, lower interaction cost and stronger operating capacity.
02Assessment
AI Governance Maturity Assessment
Pilots do not constitute governance.
Problem
Organizations often scale AI with fragmented ownership, partial controls and weak traceability.
Intervention
Assess accountability, risk, data, model operations and value/scale across a structured maturity model.
Control
Scoring, domain interpretation, failure patterns and explicit next actions.
→
Outcome
A clearer view of institutional readiness and what must be built before scale.
03Framework
Governance for AI in regulated environments
Governance should constrain consequence, not experimentation.
Problem
Generic AI policies do not define who can decide, what can execute or how material incidents are handled.
Intervention
Translate accountability, model oversight and human intervention into an operating control system.
Control
Roles, decision rights, monitoring, escalation and model lifecycle discipline.
→
Outcome
AI can move from experimentation to controlled institutional capability.
04Applied research
Measuring Generative AI Productivity
AI adoption is not productivity. Productivity begins when the economics of work change.
Problem
Licenses, prompts and time saved can show activity or local efficiency without proving higher process or enterprise productivity.
Intervention
Measure the chain from AI-exposed work and task efficiency to released capacity, process throughput and enterprise productivity.
Control
Baseline metrics, observed workflow change, value-capture rate and net productivity after rework and control costs.
→
Outcome
A defensible view of how much AI-created capacity becomes operating leverage — and where value is lost.
GOVERNANCE
Control intensity should follow consequence
A low-materiality assistant and an agent that moves money should not pass through the same governance process. Control should be proportional to what the system can affect and how difficult the action is to reverse.
OPERATING RULE
More autonomy + more materiality + less reversibility = stronger control before execution.
Autonomy
Can the system only suggest, or can it decide and execute?
More autonomy requires stronger permissions, monitoring and intervention.
Materiality
What is the financial, customer, legal or operational consequence of an error?
Higher materiality requires stronger evidence, approval and accountability.
Reversibility
Can the action be corrected quickly and completely?
Low reversibility increases the need for pre-execution controls.
Human judgment
Does the decision require context that cannot safely be reduced to policy?
Place human intervention at material ambiguity, not mechanically at every step.
Evidence
Can the organization reconstruct what the system knew, decided and executed?
Material AI processes require traceability for review and incident response.
AI ECONOMICS
A working AI system is not necessarily a valuable AI system
Technical performance is only one input. AI creates value when the economic benefit of changed work or decisions exceeds the full cost of operating, supervising and correcting the system.
DECISION EQUATION
Net AI value = economic benefit from changed work and decisions − full cost of execution and control
Economic benefit
Capacity released, cycle time reduced, conversion improved, loss avoided, service cost reduced or revenue enabled.
Model and inference
Model consumption, context, retrieval, evaluation and supporting AI infrastructure.
Integration and operations
APIs, orchestration, observability, data pipelines, maintenance and reliability engineering.
Human oversight
Review, exception handling, quality assurance, investigation and operational supervision.
Error and risk
Expected economic cost of incorrect decisions, incidents, fraud, remediation and control failure.
The business case should be built around the unit of work or decision being changed — not around tokens, prompts or model accuracy in isolation.
THESIS
AI does not create advantage because it can generate an answer. It creates advantage when an organization can trust it to execute the right decision at scale.
José Ñáñez