When work is no longer just human: the agentic operating model

This is Part 5 of the Rethinking the Target Operating Model series. Part 2 traced why the discipline broke and came back. Part 3 mapped what the major firms actually deliver. Part 4 named the three attributes of a model that works and the multi-sourced workforce as the forcing function behind them. This article goes deep on the newest source in that workforce — the one everyone is talking about and almost no one has actually operationalized: AI agents.

Start with the number that frames everything else. Of the enterprise "agent" deployments studied by Menlo Ventures at the end of 2025, only 16 percent qualified as true agents — systems that plan, act, observe the result, and adapt. The rest were fixed-sequence or routing workflows wrapped around a single model call, wearing the agent label. Gartner has a blunter name for the same gap: "agent washing." Before this is an operating-model problem, it is a definitions problem — because the word "agentic" is hiding the actual work.

What "agentic" really means — and what's being sold

The cleanest definition is Anthropic's. A workflow is a system where models and tools run through predefined code paths. An agent is a system where the model dynamically directs its own process and tool use. The difference is who decides the next step: the code, or the model. Most of what ships as agentic AI is the first kind — deterministic paths with a model call dropped inside — sold as the second.

This is not hair-splitting. The two fail differently, cost differently, and need entirely different governance. Gartner estimates only about 130 of the thousands of vendors marketing agentic AI are the real thing, and expects more than 40 percent of agentic AI projects to be cancelled by the end of 2027 — for escalating cost, unclear value, and inadequate risk controls. The label is running years ahead of the capability, and that gap is exactly where operating models break.

The three components of an agentic system — and where each stands today

Strip the marketing and an agentic system has three components. They are not equally built, and that imbalance is the whole story. Two of them — execution and orchestration — sit on a spectrum from deterministic to probabilistic. The third — governance — is not a determinism question at all; it is a maturity one. Here is each, and where it actually stands.

Autonomous execution: the part every enterprise already has

Execution is the work actually getting done, and it runs from deterministic to probabilistic. Deterministic execution follows fixed rules: extracting the fields off an invoice, matching a purchase order to a receipt and an invoice three ways, releasing a payment once the terms are met. Same inputs, same output, every time — the RPA and rules-engine world, in production for years. Probabilistic execution reasons over messy inputs and can vary from one run to the next: forecasting demand across thousands of SKUs, reading a contract for off-market clauses, scoring a new vendor. This is the LLM-agent world, and it is what "agentic" usually points at.

The difference shows up in practice: it decides which technology fits which task, and getting that match wrong is a common way agentic programs stall. Our sister platform BizBlocz frames the discipline as choosing the right mix of AI approaches for each process — document AI to read the invoice, a rules-based agent to release the payment, machine learning to forecast the demand: "same company, different process, completely different mix."

Either way, execution is the component the market has actually deployed. Gartner expects task-specific agents in 40 percent of enterprise applications by the end of 2026, up from under 5 percent in 2025; McKinsey's 2025 survey finds 23 percent of companies already scaling an agentic system in at least one function. When a company says it "did agentic," this is almost always what it means: execution, one task at a time. It is real — and it is the easy part.

Orchestration: the part almost no one has built

Execution does a task. Orchestration is what strings tasks into a process that delivers an outcome — setting the sequence, routing each task to the right executor (a deterministic bot, a probabilistic agent, or a person), governing the handoffs between them, and assembling the result the business actually wanted. A paid invoice is a task; procure-to-pay is the process; every supplier paid on time, at the right amount, with no duplicate payments is the outcome. Execution gets you the task. Orchestration is what gets you the outcome.

Orchestration also runs deterministic to probabilistic, and this is where the conflation lives. Deterministic orchestration is the workflow or BPM engine — predefined sequences, designed in advance, mature for two decades. Smart, or dynamic, orchestration is an agent deciding the flow at runtime — Anthropic's orchestrator-workers pattern, where a central model sets the subtasks "based on specific input rather than using pre-defined divisions." The first is everywhere. The second barely exists in production.

This is the gap that matters. That 16 percent — the share of enterprise agent deployments Menlo found to be genuine agents — is the orchestration gap quantified. The other 84 percent are workflows — deterministic orchestration with a probabilistic model call inside. And smart orchestration is rare not only because it is new, but because it is fragile. A Berkeley study cataloguing why multi-agent systems fail found their gains over single-agent setups "often remain minimal," while introducing fourteen distinct failure modes. The builders say the same thing: Cognition, maker of a leading coding agent, tells engineers to "just use a single-threaded linear agent"; Anthropic reports its own multi-agent system burns roughly fifteen times the tokens and is "not a good fit" for work that needs shared context. There is no clean public figure for how much "agentic" deployment is scripted versus genuinely orchestrated — but Menlo's number, Gartner's agent-washing, and the builders' own caution all point the same way. For most organizations, the orchestration that turns task execution into a process outcome is not there — and where it exists, it is the old deterministic kind.

Governance and oversight: the part that decides if you can run agents at all

Governance is what determines whether you can run any of it safely: the human boundary (when a person must approve before an action versus monitor after it), the guardrails on what an agent may touch, and a named owner accountable for what it does. This is not a deterministic-versus-probabilistic question — it is simply whether the thing has been built, and mostly it has not. Deloitte found in 2026 that only 21 percent of organizations have a mature governance model for agentic AI, even as 74 percent expect to be using agents at least moderately within a year. IBM's study of CIOs and CTOs is sharper: two-thirds say they are already accountable for AI systems they do not fully control, and only 11 percent feel ready for the scale of agent deployment coming. Gartner's cancellation forecast names "inadequate risk controls" as a primary cause. Deployment is outrunning the ability to govern it.

The three components, read together — with maturity tied to each form, not averaged:

What an AI operating model has to do differently

It is tempting to read the three components as a checklist — you have execution, now add orchestration and governance — and conclude the operating model is the same one with two pieces bolted on. It is not.

A traditional operating model coordinates work by arranging people in layers. Spans of control, reporting lines, escalation paths, the org chart — that is the machinery for making human work add up to an outcome, and it assumes the executor is a person who can be hired, managed, supervised, and held responsible. An agentic operating model coordinates work a different way: by orchestrating task execution across a mix of executors — deterministic bots, probabilistic agents, and people — around the process outcome itself. The unit you organize around is no longer the role on a chart; it is the task, routed to whichever executor should do it, with the handoffs governed. Operating a few thousand people through org layers and orchestrating a shifting mix of agents and people toward a concrete outcome are not the same job with two add-ons. They are different operating models. An AI operating model is not the old one with agents bolted on; it is organized around orchestrated outcomes rather than managed headcount. Deloitte's line captures the mismatch: bolting autonomous agents onto an operating model built for human workers is like fitting "a jet engine to a bicycle."

Commonwealth Bank of Australia learned this in public. In mid-2025 it made 45 contact-center roles redundant and credited its generative-AI voice assistant — the execution component, deployed and working. But it had not re-orchestrated the work around the new mix or governed the boundary: call volumes rose rather than fell, the bank ran overtime and pulled team leaders back onto the phones, and within weeks it reversed the redundancies and apologised, admitting the assessment "did not adequately consider all relevant business considerations." It swapped an executor and left the operating model untouched — and the model was built for the executor it removed.

So the agent is not a project you finish; it is a new executor the operating model has to coordinate alongside all the others — employees, BPO partners, contractors, gig labor, deterministic bots, and probabilistic agents — as one workforce. Mercer calls this the human-agent hybrid workforce; Staffing Industry Analysts calls it the blended workforce. Designing for it is not the old model plus two features. It is a different model, organized around orchestrated outcomes rather than managed headcount.

The three attributes, made concrete for agents

Part 4 named the three attributes of an operating model that works — Operational Depth, Continuous Adaptation, Execution Accountability. They are not generic virtues here. In an orchestrated, multi-executor model each one means something specific, and something harder.

Operational Depth — decompose to the task, because the task is the unit you assign. You cannot route work to the right executor, or place the human/agent boundary, until the process is broken down far enough that each task can be labelled: deterministic bot, probabilistic agent, or human. A model that stops at the function or org-chart layer has nothing to orchestrate — there are no units to route. This is why value has to be assessed at the subprocess level, not the function level: depth is what makes the workforce mix assignable in the first place.

Continuous Adaptation — re-cut the executor mix as the agents change, not on the budget calendar. The boundary between human and agent does not hold still. A task that needed a person last quarter is an agent task this quarter as the models improve; a fragile multi-agent flow that failed last quarter may be production-ready by the next. The allocation of tasks across executors has to be re-cut on a cadence triggered by capability, not signed off annually. An agentic operating model that freezes its human/agent split is wrong within months.

Execution Accountability — the agent executes, but it cannot be accountable. When work moves to an agent, the accountability does not move with it. An agent has no manager, no performance review, no one to answer a regulator. Someone has to own the orchestrated outcome and the agents inside it — the fleet, the guardrails, the escalation when an agent acts wrongly at machine speed. The org chart used to assign this automatically through reporting lines; with agents in the workforce it has to be assigned deliberately, or it falls straight through the gap the agent opened.

The agents handle execution. Depth, adaptation, and accountability — the orchestration and governance that turn execution into a coordinated, owned outcome — are the operating model's job. That is the line between a company with agents and a company with an agentic operating model.

Where this lands first: the offshore spokes

There is one place the shift arrives first and hardest: the operations built on labor. Global capability centers, BPO providers, shared-services organizations — the spokes of the global operating model — were designed around arbitraging human execution. When routine human execution is exactly what agents do first, the spoke's value proposition is the question on the table.

The early signs are already in the open. In mid-2026 Opendoor shut its India operations — a team it had built, in its CEO's words, to "handle manual workflows across fragmented systems" — and said it was shifting to "smaller AI-native teams." Analysts read it as a pattern rather than a one-off: the work that once justified an offshore team is the work agents target first, so the response is to redesign the operating model, not relocate the labor. The repricing shows up in the markets too — Concentrix, one of the largest customer-experience outsourcers, took a goodwill writedown of roughly $1.5 billion in its fiscal-2025 results as investors reassessed how durable a per-seat, headcount-based model is in the agent era.

That is where this series goes next. We are going to leave the boardroom view — how the operating model is designed from the top — and go to the floor: how the shift looks from inside the centers, the providers, and the teams actually absorbing it. The operating model reads very differently when you are the spoke rather than the hub.

The test

We have spent more than three decades designing operating models across every technology wave — ERP, workflow, cloud, automation, and now agents. Each arrived as a tool and was absorbed as a change to how work gets sourced, structured, and owned. Agents are the same pattern, moving faster, and the companies that get this right will not be the ones with the most agents. They will be the ones whose operating model already supplies what the agents don't: the orchestration that coordinates the work and the governance that owns it.

The test is concrete. Pick one process where you have deployed agents, or plan to. Can you name, today, which tasks are the agent's and which are the human's, who orchestrates the handoffs between them, who is accountable when the agent is wrong, and what it is allowed to touch? If you can, you have an agentic operating model. If you can only point to the agent doing the task, you have execution — and a boundary you will discover in production, the way Commonwealth Bank did, instead of in the design.

If you are working on this now rather than reading about it, our AI operations practice is where this becomes an engagement. For the design discipline underneath all of this, start with the three attributes of a target operating model that actually works and the series hub. For the full landscape of operating-model labels — agentic, composable, product, and the rest — see the operating-model-types glossary.