Search and context assembly
Retrieves policies, history, product data and prior interactions.
Intervenes when context is incomplete or contradictory.
←Back to AI
Applied AI / Operating model
2026 · Latin America
WORK → AUTONOMY
AUTONOMY → ECONOMICS
ECONOMICS → CONTROL
AI agents, unpacked
Where autonomy creates value, where humans remain essential, and what it takes to make the operating model pay.
The opportunity is not better conversation. It is better execution. The relevant question for a financial institution is not how autonomous an agent can become, but where agent execution improves capacity, unit economics and service without weakening accountability or control.
The objective is not maximum automation. It is maximum economic value inside an explicit operating and risk perimeter.
By José Ñáñez
Published April 7, 2026 · Updated August 22, 2026•9 min read
Most conversations about AI agents start in the wrong place. They start with models, prompts or autonomy. Financial institutions should start with work. A workflow that looks simple from the outside usually contains search, validation, recommendation, approval, execution and exception handling—each with a different control boundary and a different economic consequence.
That distinction matters because the business case for agents does not begin when a model produces an answer. It begins when a unit of work changes. If human effort falls, if completion cost changes, if more cases can be processed without expanding cost at the same pace, and if the institution can still explain who decided what, then the operating model has genuinely improved.
This page is therefore organized as an argument, not a feature list. It starts with the anatomy of work, moves to bounded autonomy, translates released capacity into economics, and only then asks what architecture and control model are required to let the system execute safely.
Work anatomy
Start by decomposing work—not by selecting a model
A serious agentic case begins with the unit of work. The same workflow can combine deterministic tasks, judgment, policy checks, sensitive decisions and exception handling. Each part should have a different execution boundary.
#WorkAgent roleHuman role
Classification and routing
Classifies intent, case type, priority and next operational path.
Owns ambiguous, high-consequence or policy-exception cases.
Document collection and completeness
Requests evidence, checks completeness and prepares missing-item follow-up.
Validates exceptions, authenticity concerns and non-standard evidence.
Recommendation
Synthesizes facts and proposes the next best authorized action.
Decides when regulation, materiality or conduct risk requires judgment.
Execution
Executes reversible, permissioned and policy-bounded actions through tools and APIs.
Authorizes or performs sensitive, irreversible or high-consequence actions.
Exception management
Packages context, evidence, prior actions and recommended next steps.
Resolves the exception and remains accountable for the decision.
READING NOTE
The unit of work changes the conversation
Once work is decomposed, the discussion becomes more precise. The question is no longer whether an institution should “use agents,” but which fragments of a workflow can be delegated, which require supervision and which should remain fully human because the cost of being wrong is too high.
That shift is important for executives because it connects technology with operating leverage. A bank does not scale simply because it added AI. It scales when the same number of people can process more customers, more products and more interactions without losing clarity, control or service quality.
The goal is not to automate more work. It is to change the economics of work without losing control.
Bounded autonomy
Autonomy is an operating decision, not a technology setting
The right level depends on consequence, reversibility, policy clarity, data sensitivity and required authority. Higher autonomy is not automatically better.
DESIGN RULE
The design target is economically optimal autonomy: enough execution authority to change the economics of work, but not more authority than the institution can safely govern.
Human-led
Sensitive or poorly structured work.
Agent
Searches, summarizes and prepares context.
Human
Decides and executes.
Copilot
Knowledge-intensive work with material judgment.
Agent
Analyzes and recommends.
Human
Approves and executes.
Agent-assisted
Structured workflows with selected sensitive steps.
Agent
Executes low-risk steps and prepares control points.
Human
Approves defined checkpoints and exceptions.
Bounded agentic
Repeatable, measurable and reversible workflows.
Agent
Executes end to end inside permissions, policies and thresholds.
Human
Supervises, handles exceptions and can intervene or stop.
Human-only
Actions where legal, conduct or prudential constraints dominate.
Agent
No execution authority.
Human
Maintains full control and accountability.
Autonomy is not the prize
Many teams still talk about autonomy as if it were a feature slider. In regulated industries it is an operating decision. The right level depends on consequence, reversibility, policy clarity, data sensitivity and the authority required to act.
That is why the most valuable agentic designs are rarely the most autonomous ones. The best design is the one that gives the system enough execution authority to alter throughput and unit economics, while keeping human intervention explicit where the institution still needs judgment, approval or accountability.
Agent economics
Released time is not the same as financial value
A defensible business case separates operational capacity, recognized economic value and the full cost of agentic execution. The model should make each bridge explicit.
Economic bridge
01
Current work → eligible work → straight-through execution + escalation + review + rework
02
Released human capacity → captured efficiency + productive growth + non-monetized operating benefit
03
Gross recognized value − agent runtime − platform − observability − AgentOps − governance = net annual benefit
Capacity first
Measure what human effort actually disappears after eligibility, escalation, review and rework—not a theoretical automation percentage.
Economic use second
Released capacity creates financial value only when it becomes captured efficiency, avoided future hiring or productive incremental volume.
Charge the full operating model
Runtime is only one cost. Include tools, orchestration, platform, observability, escalation, AgentOps and governance.
Separate resource cost from P&L capture
A lower modeled cost per case can coexist with little short-term P&L benefit if the organization does not economically redeploy or capture released capacity.
Time saved is not value created
This is where many AI business cases become inflated. Released time is not automatically savings, because an institution can free capacity and still fail to capture any P&L benefit. The bridge from capacity to value only becomes real when management explicitly converts that capacity into efficiency, avoided future hiring or productive additional volume.
The model below separates those layers on purpose. It charges runtime, platform, observability, AgentOps and governance, and it forces a choice about how released capacity is used. That makes the economics less dramatic—but much more believable.
Released time only becomes economic value when leadership decides how it will be captured.
MODEL MECHANICS
What the simulator calculates — and in what order
The model follows a causal chain. It first reconstructs the current human workload, then splits agent-eligible work between straight-through execution and human intervention, calculates the human effort that remains, converts released capacity into explicitly chosen economic uses, charges the full agent operating layer, and only then calculates annual benefit and payback.
Current human workload
Current human hours = annual cases × current human minutes per case ÷ 60Establishes the labor baseline before any agent intervention. This is workload capacity, not payroll savings.
Agent execution split
Eligible cases = annual cases × eligibility %; straight-through cases = eligible cases × STP %; escalated cases = eligible − straight-through; rework cases = straight-through × rework %Separates what the agent can attempt from what it completes without human execution, what returns to a person, and what requires rework.
Human effort after agents
Transformed human hours = non-eligible work + escalation work + straight-through review + reworkThe model does not assume that eligible work becomes fully automated. Human review, escalation and rework remain visible.
Released capacity
Released hours = max(0, current human hours − transformed human hours); equivalent capacity = released hours ÷ 1,760Released hours measure operational capacity. The 1,760-hour denominator is an illustrative annual productive-capacity assumption (220 days × 8 hours), not a workforce-reduction claim.
Full agent operating cost
Agent OPEX = eligible cases × variable agent cost per case + platform/observability + AgentOps/governanceThe model charges more than inference. It includes the recurring operating layer required to execute, observe and govern agents.
Recognized economic value
Efficiency value = released hours × loaded hourly cost × efficiency capture %; growth cases = released hours × growth allocation % × current cases/hour; growth value = growth cases × contribution per additional caseCapacity is monetized only when management explicitly assigns it to captured efficiency or productive additional volume. Unallocated capacity remains an operating benefit, not P&L value.
Net benefit and investment recovery
Net annual benefit = efficiency value + growth value − agent OPEX; payback months = transformation investment ÷ net annual benefit × 12; 3-year net value = annual benefit × 3 − initial investmentThis is the final economic bridge. The model does not claim ROI until capacity use and the recurring execution cost have both been recognized.
Two views are intentionally kept separate
Modeled cost per case compares the resource cost of the current and agent-enabled operating models. Net annual benefit is a different view: it recognizes only the portion of released capacity management chooses to monetize, then subtracts agent OPEX. This avoids treating every released hour as an automatic cash saving.
Interactive model
Agent economics, with your numbers
Model one recurring workflow. The tool separates operational capacity from economic value: released time only reaches the business case when management explicitly captures efficiency or uses capacity for additional productive volume.
The model recognizes value only after separating capacity, economic use and the full cost of execution.
Operating-mode presets
Illustrative starting points—not market benchmarks. Every parameter remains editable.
No FX conversion is performed. Use the same currency for every monetary input.
1. Current workload
012. Agent execution
023. Economic use of released capacity
034. Program cost
04Economic bridge
From human work to recognized economic value
Current model
Human hours / year
72,000
Current modeled labor cost
$2,016,000
Current modeled cost / case
$8.40
Agent-enabled model
Human hours / year
38,390
Modeled human + agent resource cost
$1,399,514
Agent-enabled modeled cost / case
$5.83
Released-capacity allocation
Efficiency 35%Growth 40%Retained 25%
Straight-through agent cases
85,800
Escalated eligible cases
70,200
Rework cases
2,574
Released human hours / year
33,610
Equivalent capacity released
19.1
Annual agent operating cost
$324,600
Explicit efficiency captured
$329,380
Additional cases supported by growth allocation
44,814
Incremental contribution
$358,509
Net annual economic benefit
$363,289
Simple payback
14.9 months
3-year net value at steady-state benefit
$639,866
CALCULATION TRACE
See the formulas with the values currently in the model
Every row below updates with the inputs above. The objective is to make the model auditable: you can see the formula, the substituted values and the resulting output rather than relying on a black-box ROI number.
CalculationFormula with current valuesResult
01Current human workload
240,000 × 18 ÷ 60 = 72,000 h72,000 h02Eligible cases
240,000 × 65% = 156,000156,00003Execution split
156,000 × 55% = 85,800 STP; 156,000 − 85,800 = 70,200 escalated; 85,800 × 3% = 2,574 rework85,800 / 70,200 / 2,57404Human effort after agents
(84,000 × 18 + 70,200 × 9 + 85,800 × 1.5 + 2,574 × 12) ÷ 60 = 38,389.8 h38,389.8 h05Released capacity
72,000 − 38,389.8 = 33,610.2 h; 33,610.2 ÷ 1,760 = 19.1 FTE33,610.2 h · 19.1 FTE06Agent operating cost
156,000 × $0.35 + $180,000 + $90,000 = $324,600$324,60007Agent-enabled cost per case
(38,389.8 × $28.00 + $324,600) ÷ 240,000 = $5.83$5.8308Captured efficiency value
33,610.2 × $28.00 × 35% = $329,380$329,38009Growth-supported additional cases
33,610.2 × 40% × (60 ÷ 18) = 44,81444,81410Incremental growth contribution
44,814 × $8.00 = $358,509$358,50911Net annual economic benefit
$329,380 + $358,509 − $324,600 = $363,289$363,28912Simple payback
$450,000 ÷ $363,289 × 12 = 14.9 months14.9 months133-year net value
$363,289 × 3 − $450,000 = $639,866$639,866Equivalent FTE capacity uses 1,760 productive hours per year as an illustrative assumption. Change the interpretation for institutions with a different productive-hours standard.
What the model recognizes
01
Non-eligible work remains human at the current minutes per case.
02
Eligible work is split between straight-through execution and human escalation.
03
Review and rework consume human capacity before any benefit is recognized.
04
Released hours are not automatically treated as savings.
05
Only explicit efficiency capture and productive growth allocation create gross economic value.
06
Agent runtime, platform, observability, AgentOps and governance are charged against that value.
INTERPRETATION
How to read the instrument
Read the simulator in three passes. First, examine how much human work actually disappears after eligibility, escalation, review and rework. Second, decide how that capacity is used: captured efficiency, productive growth or non-monetized operating benefit. Third, compare that recognized value with the full cost of the new execution layer.
This is also why the most important output is not ROI alone. The more revealing signals are modeled cost per completed case, released capacity, additional throughput supported by growth allocation and the payback period after charging the real operating cost of agents.
Latin America
Scale the platform; localize the operating perimeter
A regional agent strategy should not clone a country implementation. The reusable layer is the execution platform; the local layer is the policy, product, integration and conduct perimeter that determines what the agent may actually do.
Regional economics improve when common AgentOps, observability and orchestration are reused while local decision rights remain explicit.
Shared regional platform
Orchestration and identity
Agent routing, session context, permissions, authentication and delegation patterns.
Observability and AgentOps
Execution traces, quality, incidents, cost telemetry, deployment controls and rollback.
Knowledge architecture
Reusable retrieval, content lifecycle, grounding controls and evidence management.
Security and control primitives
Tool allowlists, secrets, authorization boundaries, policy enforcement and audit patterns.
Country rule packs
Products and eligibility
Country-specific products, suitability, credit, servicing and commercial rules.
Privacy and consent
Local data requirements, disclosures, consent, retention and sensitive-data treatment.
Conduct and language
Required wording, prohibited behavior, local terminology and customer-protection boundaries.
Systems and exceptions
Local cores, payment rails, identity services, document standards and operational exception paths.
Execution requires a control system
Once the economics are clear, the architecture has to answer a harder question: how can a bank let a system act without losing control? In agentic environments, architecture is no longer just a way to generate answers. It becomes the mechanism that defines identity, permissions, policy boundaries, escalation paths and continuous evidence.
That is why AgentOps matters. At scale, the institution is not managing a chatbot. It is managing a fleet of operators with memory, tools and execution rights. Observability, rollback, intervention and traceability are not support functions around the model. They are part of the operating model itself.
Reference architecture
The architecture must control execution, not only generate answers
An LLM connected to a channel is not an operating model. Agentic execution requires identity, policy, tools, state, observability and human intervention as first-class architecture elements.
Channels and work entry
Web, mobile, WhatsApp, contact center, branch, advisor desktop, batch queues and event-driven work.
Identity, consent and delegation
Who is acting, on whose behalf, with which consent, authority, session and permission boundary.
Agent orchestration
Intent, planning, agent selection, state, memory scope, task decomposition and multi-agent coordination.
Policy and decision controls
Country, product, customer, channel, risk and action rules applied before tools are invoked.
Knowledge and enterprise tools
Grounded content plus CRM, origination, payments, servicing, collections, workflow and core APIs.
Human intervention
Approval points, exception queues, intervention, override and full context transfer without forcing a restart.
Observability and AgentOps
Prompts, traces, tool calls, latency, cost, policy outcomes, incidents, drift, quality, rollback and audit evidence.
Governance
Treat agents as operators with permissions and accountability
The control model must follow the agent across design, deployment and operation. Governance that only reviews a model before launch is insufficient when the system can call tools, preserve state and execute actions.
Named owner
One accountable business owner for the workflow, its economics and its permitted outcomes.
Explicit permissions
Tool access, data scopes, transaction limits and action rights are defined and revocable.
Human intervention policy
Clear triggers for approval, escalation, supervision, override and forced stop.
Continuous evidence
Traceability must show what the agent saw, decided, called, changed and escalated.
Quality and failure thresholds
Operational metrics include completion, escalation, rework, error, policy breach and adverse-outcome indicators.
Release and rollback discipline
Versioning, simulation, controlled rollout, canaries, restrictions and rollback are part of normal AgentOps.
Adoption path
Increase autonomy only when the evidence supports it
The maturity path is not a race toward autonomy. Each phase should expand execution rights only after economic, quality and control thresholds are demonstrated.
01
Phase 1
Measure the work
Baseline volumes, handling time, rework, escalation, service levels, cost and decision points.
Gate: credible unit economics and a decomposed workflow.
02
Phase 2
Assist humans
Introduce retrieval, summarization, recommendations and case preparation without sensitive execution.
Gate: measurable quality and productivity improvement.
03
Phase 3
Execute bounded steps
Allow low-risk tool use, registrations and reversible actions with explicit checkpoints.
Gate: stable permissions, observability and exception handling.
04
Phase 4
Operate bounded workflows
Move selected workflows end to end with human-on-the-loop supervision and exception ownership.
Gate: sustainable economics, control effectiveness and acceptable failure rates.
05
Phase 5
Scale regionally
Reuse the platform and AgentOps while introducing country-specific rule packs and integrations.
Gate: local validation without fragmenting the regional control model.
METHOD / LIMITS
Methodology, assumptions and limitations
The interactive model is deterministic and designed for executive scenario analysis. It is not a forecast, accounting opinion, workforce plan or regulatory assessment.
Illustrative operating-mode presets are starting points only. They are not presented as industry benchmarks. Formal analysis should use observed workflow data and institution-specific costs.
The model deliberately separates modeled resource cost from recognized economic value. Released hours are only monetized when the user explicitly allocates them to captured efficiency or productive growth.
The model excludes credit losses, fraud losses, conduct remediation, regulatory capital, taxes and adverse-customer-outcome costs. Those effects should be added in the formal business case when relevant.
Regional implementation should distinguish shared platform economics from country-specific enablement, integration and control costs.
Research context
McKinsey — The paradigm shift: How agentic AI is redefining banking operationsFebruary 2026. Supports the shift from isolated pilots toward end-to-end redesign of work, decisions and delivery.→Accenture — Top banking trends for 2026Frames agentic AI as a capacity model, emphasizes intent-driven workflows and proposes AgentOps as a control tower for scaled adoption.→Deloitte Insights — Agentic AI risks in banking2026 analysis of agent permissions, accountability, observability, human oversight and controls for autonomous execution.→
Closing thesis
AI agents create value when institutions stop treating them as better interfaces and start treating them as a new unit of controlled execution.
The winning operating model will not maximize autonomy. It will allocate work between humans and agents so that capacity, economics, accountability and risk improve together.
What matters most is not how advanced the agent appears. It is whether the institution can explain where value is created, where control is retained and how both can scale together.
José Ñáñez · Strategy, technology and operating models in regulated industries.