Published: October 29, 2026 | Category: AI | By Mahesh

AI native operations means redesigning how work flows through a company around AI agents that carry out multi-step tasks, instead of adding an assistant to each old step. The gap between talking and doing is wide. McKinsey's State of AI in 2025 survey found 62% of organizations experimenting with agents and only 23% scaling them, usually within a single function^[1][6]. Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027 because of cost, unclear value or weak risk controls^[2].

This cluster article sits under Depth Grid's guide to the AI native company. It explains why operations is where most AI native ambitions succeed or fail, and it gives three tools a manager can use today: a reliability table, a readiness score and a cost model. The central claim is simple. Agents do not fix a broken process. They make it faster, and a faster broken process fails more often.

What Does AI Native Operations Mean?

Direct answer: AI native operations rebuilds a business process so that agents own the routine steps and people own exceptions, judgment and sign-off. It starts from the outcome the process must deliver, then removes steps that existed only because humans had to pass work between each other.

A company is what an AI native company is when AI changes how work is organized and not only which tools staff open. In operations that shows up in order handling, procurement, logistics updates, claims, support escalations and compliance checks, where work moves through many small steps. Traditional automation encoded the steps. Agents can plan the steps and adapt when something unexpected happens, which is useful and also the source of new risk.

The practical difference is who holds the process. In a traditional setup a person follows a procedure and a system records it. In an AI native setup an agent holds the process, a person owns the outcome, and a written rule says when the agent must stop and ask. That last rule is the control that decides whether the arrangement scales.

How Far Along Are Companies?

Direct answer: McKinsey found 88% of organizations use AI in at least one function but only 39% report any profit impact from it, mostly under 5% of earnings. Gartner expects at least 15% of daily work decisions to be made autonomously by 2028, up from 0% in 2024.

The McKinsey State of AI in 2025 report, drawing on nearly 2,000 respondents, says 88% of organizations regularly use AI in at least one business function and that nearly two-thirds have not begun scaling it across the enterprise^[3]. Only 39% report any attributable impact on earnings before interest and tax, and for most of them it is below 5%^[3]. On agents, 62% are at least experimenting and 23% are scaling somewhere, most often in IT and knowledge management^[1][3].

Gartner paints a similar picture from the vendor side. Its forecast says at least 15% of day-to-day work decisions will be made autonomously through agentic AI by 2028, up from 0% in 2024, and that 33% of enterprise software applications will include agentic AI by 2028, up from under 1% in 2024^[2]. It also warns of "agent washing," the relabeling of chatbots and robotic process automation as agents, and estimates that only about 130 of thousands of agentic AI vendors are real^[2]. A January 2025 Gartner poll found 19% of organizations had made significant investments in agentic AI, while 42% were cautious^[4].

Large vendors keep pushing in. Salesforce and Oracle have both embraced agents, as the South China Morning Post noted when it reported the Gartner forecast^[5]. For an operations leader that creates a buying risk. The vendor pitch is loud. The real product count is small. The cancellation forecast says many projects will not survive contact with cost and control questions.

Why Does Redesign Beat Bolt-On?

The most useful statistic in the McKinsey data is about method and not about technology. AI high performers were 2.8 times more likely to report fundamental workflow redesign, at 55% against 20% of other organizations^[3]. They were also far more likely to have defined human-in-the-loop validation processes, at 65% against 23%^[3]. A commentary on the report from HiQ makes the same point from the leadership side: senior leaders in the successful group were more involved in owning AI projects^[6].

Bolting an agent onto an existing process keeps every handoff, approval and form that was designed for humans. The agent then spends its effort navigating a maze built for people. Redesign asks a different question for each step: does this step add value to the customer or the control environment, or does it exist only because one person needed to tell another person something? Steps in the second group can usually disappear.

Gartner's advice points the same way. It recommends pursuing agentic AI only where it gives clear value or return, and notes that many use cases labeled agentic do not need an agent at all^[2]. A rules engine or a simple model is often the better answer for a stable, well-defined step. Agents earn their place where the path varies and judgment is needed.

Why Do Multi-Step Agents Fail More Than You Expect?

Direct answer: Success rates multiply across steps. An agent that completes each step correctly 95% of the time finishes a 10-step workflow correctly only about 60% of the time, and a 20-step workflow only about 36% of the time.

This is arithmetic and not a prediction about any one product. If each step succeeds independently with probability p, a workflow of n steps succeeds with probability p raised to the power n. Real steps are not fully independent, and good systems add checks, but the multiplication explains why a demo that looks impressive can fail in production.

Source: Depth Grid calculation. End-to-end success = per-step success raised to the number of steps, assuming independent steps.

Two design rules follow. First, shorten the chain. Splitting a 20-step process into four 5-step sub-workflows with a checkpoint after each does not change the arithmetic by itself, but it lets a person or a validator catch an error after step 5 instead of step 20. Second, measure per-step success and not just final success, because a step at 90% quietly drags everything downstream.

Which Processes Should Agents Take First?

Direct answer: The Ops Readiness Score rates a process from 1 to 5 on four factors: volume, rule clarity, tolerance for error and data readiness. A total of 16 to 20 suits an agent now, 11 to 15 suits a pilot with human review and 10 or less should stay with people.

This is a Depth Grid working framework, not a published standard. It exists because the cancellation forecast suggests many firms pick use cases for excitement and not for fit. Each factor is scored from 1 (poor) to 5 (strong). Tolerance for error is scored high when a mistake is cheap and easy to reverse.

Source: Depth Grid framework with illustrative scores. Bands: 16 to 20 start now; 11 to 15 pilot with human review; 10 or less keep with people.

Invoice matching wins because it is repetitive, rule-based and reversible. Contract negotiation scores low because volume is small, rules are fuzzy and an error is expensive. The middle case, refund exceptions, shows the typical pattern: an agent can prepare most of the work while a person approves anything above a set value.

What Does One Completed Workflow Cost?

Leaders often compare the price of an agent call with the price of an hour of staff time. That comparison misses the cost of failures. A better measure is the cost of one completed workflow, including the human work spent on the cases the agent cannot finish. The model below uses invented but plausible numbers to show the mechanism.

Source: Depth Grid illustration. Assumptions: $0.40 agent cost per attempt, $6 human cost per exception, $5 human cost per workflow as baseline, $40 loss per silent error.

Three lessons come out of the table. The savings stay large even at 70% success, so a mediocre agent can still pay for itself on cheap tasks. The savings fall quickly as exceptions rise, because every failed case costs more human time than the baseline did. And silent errors, the ones nobody catches, can erase a large share of the gain. That is why McKinsey's finding on human-in-the-loop validation matters: it is a cost control as much as a safety control^[3].

What Operating Model Do Agent Workflows Need?

Three roles keep an agent workflow healthy. A workflow owner is accountable for the outcome, the metrics and the decision to expand or stop. An exception desk handles the cases the agent hands back, and its queue size is the best early warning that something has drifted. An evaluator samples completed work each week, scores it against a checklist and reports per-step accuracy. Small companies can combine the roles in one or two people. The point is that each job has a named owner.

Controls matter as much as roles. Each agent needs its own identity, a limited toolset and logs of what it did, a theme covered in CrowdStrike Built a Product for Your Unapproved AI Agents, which describes a product built to find and govern agents nobody approved. Open tooling is emerging too, as in our summary of the Nvidia Open Agent Safety Platform. Agents with clear job scopes are also easier to govern, which is the reasoning behind Salesforce Gave Its AI Agents Job Titles, Not Features.

Operations also depends on good data about the work itself. SAP Buys TechWolf: What the AI Work Deal Means shows a vendor buying a graph of tasks and skills so agents have context. A mid-size company can start smaller by writing down, for each process, the inputs, the rules, the owner and the exceptions. That document is the context an agent needs, and it is often the first honest map of the process the company has ever had.

How Does AI Native Operations Differ From Traditional Operations?

Source: Depth Grid analysis.

What Will AI Native Operations Look Like by 2031?

The table below is Depth Grid analysis built on the dated forecasts above. It is a projection and not a measured result.

The riskiest assumption is the 2031 row. It requires per-step reliability to rise well above today's levels, and the compounding table shows how demanding that is. Improvement in models helps, but so does process design, which is entirely in a company's own hands.

How Does This Play Out in India, the US and Europe?

India runs a large share of the world's outsourced operations, including finance processing, customer support and claims handling. Those centers are among the first places where agents will redesign flows, and the commercial model will shift from billing per person to billing per outcome. Data handling falls under the Digital Personal Data Protection Act, 2023, with the Ministry of Electronics and Information Technology responsible for policy, and teams should confirm client-specific obligations. In the United States, adoption pressure is driven by margin targets and vendor competition. In Europe, documentation, worker consultation and the status of AI Act obligations, which should be checked against the European Commission text, shape the pace.

What Should an Operations Leader Do in the Next 90 Days?

The plan below is Depth Grid's suggestion and fits a team of any size.

1. Days 1 to 30: map and score. List your ten highest-volume processes, count the steps in each and score them on the Ops Readiness Score. Pick one process scoring 16 or more.
2. Days 31 to 60: redesign before you automate. Remove steps that exist only for handoffs, write the exception rules and name the workflow owner, the exception desk and the evaluator.
3. Days 61 to 90: run and measure. Launch the agent on a share of live volume, track per-step success, exception rate and cost per completed workflow, and sample work weekly for silent errors.
4. Then decide: expand, fix or stop. Use the cost model with your own numbers and set a rule in advance for when the pilot ends.

A single well-chosen process that reaches production teaches more than five pilots that never leave the demo stage.

What Is Still Uncertain?

Several cautions apply. The Gartner cancellation figure is a forecast made in 2025 and has not been tested against outcomes. The McKinsey figures are survey responses from 2025, and some details reach this article through secondary commentary. The reliability table assumes independent steps, which real systems only approximate. The readiness score and the cost model use illustrative numbers and are teaching tools and not benchmarks. Vendor claims about agent capability change quickly, so any product comparison needs fresh testing.

The direction is stable whatever the exact figures. Companies that redesign a process, score it honestly, measure failures and keep a person accountable will get more from agents than companies that buy the biggest platform.

Common Questions

Sources

1. The Vault, "AI at work: how McKinsey's State of AI report changes the game for job seekers," December 5, 2025, summarizing McKinsey's survey on agents. vault.com
2. Gartner prediction on agentic AI cancellations, as reported by Outlook Business, "Over 40% of Agentic AI Projects Will Be Scrapped by 2027, Says Gartner." outlookbusiness.com
3. CX Today, "McKinsey's State of AI: The Scaling Gap Is Now CX's Problem," February 23, 2026. cxtoday.com
4. CDO Magazine, "Over 40% of Agentic AI Projects Likely to Be Abandoned by 2027," citing a January 2025 Gartner poll. cdomagazine.tech
5. South China Morning Post, "Over 40% of agentic AI projects forecast to be scrapped by 2027 due to lack of value." scmp.com
6. HiQ, "The major AI shift has begun, but leadership is falling behind," analysis of McKinsey's State of AI in 2025. hiq.se

Related Analysis

• The AI Native Company: The Complete Guide to Building AI-First Businesses in 2026
• What Is an AI Native Company? Complete Definition, Examples and Guide
• Salesforce Gave Its AI Agents Job Titles, Not Features
• CrowdStrike Built a Product for Your Unapproved AI Agents
• SAP Buys TechWolf: What the AI Work Deal Means

Article by Mahesh | Depth Grid