CEO Corner

An operating model is how the organization gets work done

I use operating model to mean the connection between the work, the people responsible for it, the systems they use, the decisions they can make and the measures that tell us whether the outcome is acceptable. Agentic AI touches all five. It can gather evidence, select tools and coordinate tasks, but introducing that capability into an unclear process can make the confusion move faster.

My starting point is a complete business outcome: a beneficiary receives a properly assessed claim decision; a banking exception reaches a reconciled resolution; a finance team closes a reporting period with evidence it can explain. Each outcome crosses departmental boundaries. The design must follow that whole journey.

Give the journey one owner and a shared definition of done

Consider a service team that measures how many requests it forwards. Its AI assistant may increase that number while creating a larger investigation queue for another team. Local productivity rises; the customer still waits. I want one journey owner accountable for completion, supported by operations, technology, data, risk and finance.

Together, those teams should map the normal path and the exceptions, agree what completion means and record the baseline. Which information is repeatedly requested? Where do cases wait for authority? Which checks are essential, and which exist because two systems do not share context? Genellipse helps make those dependencies visible before deciding where an agent belongs.

Design the boundary between assistance and authority

I favour a clear progression: an agent first prepares information, then recommends a next step, and only performs an action when that action falls within an explicitly authorized scope. The organization must define the permitted tools, records, transaction limits, time limits and escalation conditions. A persuasive response does not grant permission to act.

Use a predictable workflow when the sequence is known. Introduce agent-directed steps where the task genuinely requires adaptation, and test whether that flexibility improves the result. Anthropic’s engineering guidance makes a useful distinction between predefined workflows and agents that choose their process and tools. For leaders, the practical implication is to match autonomy to the work and its consequences.

The operating loop below is a design proposal. Its approval and verification stages should be implemented in the workflow and connected systems, with named human owners. They should never depend solely on asking the model to behave correctly.

Life insurance: prepare the case without surrendering the decision

Imagine a life-insurance claim with a policy record, beneficiary information and documents arriving through different channels. A document agent can structure the material; an evidence agent can retrieve permitted policy information; a coordination agent can identify missing items and prepare a case summary. Each finding should point back to its source, including unresolved conflicts.

The claims professional assesses the case and follows the insurer’s authority rules for the decision and any payment. Conflicting beneficiary records, uncertain coverage or missing evidence must enter an exception path. The system should support the reviewer’s investigation, preserve the decision record and confirm the downstream outcome.

This illustrates how AegisNow insurance workflows and Cortex AI orchestration could be configured together. It is an implementation pattern, not a reported customer result. The benefit to test is less repeated preparation and a clearer case for the professional responsible for the claimant’s outcome.

Banking: resolve the exception across systems and teams

A payment can appear submitted in one channel while the payment engine shows processing and the ledger has no matching entry. In an illustrative agent-assisted workflow, the system correlates identifiers, gathers status evidence and prepares an exception case. An operator sees the known facts, the uncertainty and the permitted next actions.

The bank retains authority over payment release, adjustments and resubmission. Execution should include duplicate-prevention controls, confirmation from the relevant banking system and reconciliation after the action. A failed API call must not quietly become an instruction to repeat a payment.

The operating-model improvement is a connected resolution process: one accountable owner, a shared case history and an explicit escalation path. Genellipse can help connect these elements around existing banking systems rather than requiring a wholesale replacement before useful progress begins.

Equip people for the work that changes

Human oversight has to be a real job with enough time, information and authority. A reviewer needs to see source evidence, understand what the agent attempted and be able to correct, stop or escalate the work. Measuring that person only on speed creates the wrong incentive.

I recommend a small, durable team around each priority journey. The business owner defines the outcome; a domain specialist judges quality; a data owner maintains meaning and permissions; engineering owns integration and reliability; and risk and security agree controls and escalation. Frontline employees should help design exception handling because they understand where the process breaks.

Genellipse can support role design, practical training and operating runbooks alongside implementation. Success includes people becoming more capable of supervising the work and improving it. Workforce decisions require a considered plan for skills and redeployment, rather than an assumption that every minute released translates into a role removed.

Connect agents through controlled APIs and MCP tools

Existing policy administration, core banking and finance systems contain business rules that still matter. Our approach is to expose selected capabilities through documented APIs and, where appropriate, Model Context Protocol (MCP) tools. Keep the authoritative record clear, validate tool inputs and preserve the transaction controls of the underlying system.

MCP provides an integration mechanism; using it does not by itself establish authorization or trust. Its official security guidance addresses risks such as token passthrough and confused-deputy attacks. Organizations must apply appropriate authentication, scoped access and independent enforcement at the integration boundary.

I also want explicit tests for instructions hidden in retrieved documents, attempts to access another customer’s data, unavailable dependencies, duplicate actions and an agent exceeding its allowed budget. Record tool activity without unnecessarily exposing sensitive information. Maintain a way to pause the workflow, revoke access and return the case to an operational team.

Measure the economics of a completed, acceptable outcome

Counting agents, prompts or generated summaries tells me little about business value. I want elapsed time to resolution, total handling effort, rework, exception age and the quality of the accepted outcome. Include the costs of integration, models, infrastructure, human review, training and support.

The chart below is deliberately hypothetical. Suppose preparation and validation improve while professional judgment and execution checks retain their full time allocation. Handling effort would fall from 50 to 30 minutes per completed case. At 1,000 comparable cases a month, that arithmetic represents roughly 333 hours of gross capacity released before additional operating overhead.

Those hours are not automatically cash savings, and the chart is not a Genellipse performance claim or forecast. The team must validate case quality and decide how released capacity will improve service, absorb volume or reduce an actual expense. Finance should sign off the assumptions and compare the result with the original baseline.

Make governance part of everyday operation

NIST’s voluntary AI Risk Management Framework organizes the work around Govern, Map, Measure and Manage. I see that as a useful reminder that oversight continues after an initial approval. In practice, the journey owner needs a regular view of errors, escalations, costs, changes and unresolved risks.

A model update, a changed prompt, a new data source or a revised tool permission can change system behaviour. Record the change, evaluate representative and difficult cases, and agree rollback criteria. Expand autonomy only when the evidence supports it. A failed evaluation should result in a correction or a narrower scope, with accountability for the next decision.

Start with a focused MVP within six weeks

For a suitably scoped engagement, Genellipse targets an MVP within six weeks once the required access, representative data, environments and decision-makers are ready. In the first two weeks, map the journey, agree the baseline and set the authority boundary. In weeks two to four, connect the selected sources and build the bounded workflow. Use weeks four to six for evaluation, exception testing, user preparation and the agreed release decision.

The deliverable should include a working capability, the evidence supporting its acceptance, a named owner, monitoring and a recovery procedure. Wider deployment, core conversion and external approvals need their own plans. The first release should teach the organization how to operate the next one responsibly.

The question I would put to every leadership team

Choose one important journey and ask: if an agent joined this process tomorrow, could we explain its job, its authority, its evidence and its escalation path? If the answer is unclear, that is valuable information about the operating model itself.

I believe the lasting opportunity is an organization that makes better use of its people, information and technology together. Agentic AI can help us reach that outcome when we design the work and accept responsibility for how it runs.