What Is Agentic AI — and Why the Future of Enterprise Transformation Is Agentic

Everyone has AI. Almost nobody has been transformed by it.

McKinsey's global survey found that while 88% of organizations now use AI in at least one function, only 39% can attribute any EBIT impact to it — and just ~6% qualify as high performers seeing 5% or more.[7][10] MIT's Project NANDA put a sharper number on the same gap: despite an estimated $30–40 billion in enterprise GenAI investment, roughly 95% of organizations were getting zero return, with only about 5% of integrated pilots extracting millions in value.[2]

The researchers were explicit about the cause. The divide was not driven by model quality or by regulation. It was determined by approach.[2]

That sentence is the whole enterprise AI conversation right now. And agentic AI is where the approach question finally gets decided — for better or worse.

A copilot recommends. An agent executes.

A copilot retrieves, drafts, and suggests; a human still does the work. An agent is handed a goal, a set of tools, and a permission boundary — then it plans, calls real systems, executes multi-step work, verifies the result, and escalates only the exceptions.

Deloitte's maturity ladder captures the shift: assist, advise, coordinate, orchestrate, self-evolve. The human role slides from operator to supervisor to orchestrator of a team of agents.[10]

What changes the economics is memory and action. MIT traced stalled pilots to a "learning gap" — static tools that don't retain feedback, adapt to context, or integrate with how work really happens.[2] Agents close it by acting inside the workflow rather than beside it.

The honest counterweight

Gartner projects that over 40% of agentic AI projects will be canceled by the end of 2027 — due to escalating costs, unclear business value, or inadequate risk controls — and warns that much of what is sold as agentic today is "agent washing," rebranded assistants and chatbots.[1][8] Its analyst was blunt: "Many use cases positioned as agentic today don't require agentic implementations."[1]

This is not a story where everyone wins. It's bimodal — which is precisely why the redesign matters more than the tooling.

What agentic AI looks like in production

Four real examples:

• Customer service — Klarna. An OpenAI-powered assistant handled 2.3 million conversations in its first month, two-thirds of the company's service chats, doing the equivalent work of 700 full-time agents. Resolution dropped from 11 minutes to under 2, satisfaction stayed on par with human agents, repeat inquiries fell 25%, and Klarna estimated $40 million in profit improvement.[3][9]
• Field operations — Walmart. The retailer has stood up four agent platforms, mixing custom-built and external models, and certifies associates on agentic AI through its Squiggly platform — now open to 1.7 million US and Canadian associates, on the way to all 2.1 million employees.[4][11] Walmart also reports the honest complication: associates spend real time correcting AI errors and supplying the context agents lack.[12] That is what a mature rollout actually looks like.
• IT operations. Forrester found organizations deploying ServiceNow ITSM agents achieved an average 210% ROI over three years, with payback in under six months.[6]
• Back office — the least glamorous, highest-ROI territory. A financial services firm cut manual effort in credit approval and fraud detection by 30%, with compliance designed into the architecture from day one. A 600-location hospitality group compressed a 160-hour data migration and reporting cycle into four hours.[5] A January 2026 Forrester Total Economic Impact study commissioned by Microsoft found its agentic AI deployments delivered 120% ROI and $24.2 million in net present value over three years, payback in 15 months.[5]

The uncomfortable part: the value is in the redesign

Four research houses, measuring differently, reach the same conclusion: most of an initiative's value comes from reworking the work, not buying the tool. PwC puts it at 80%; Bain at roughly two-thirds.[10] High performers are about 2.8x more likely to have fundamentally reworked workflows — 55% versus 20%.[10]

Meanwhile Deloitte found only 23% of companies use agentic AI even moderately, and just 21% have a mature governance model for autonomous agents. The other 79% are running agents without adequate oversight.[6]

The failure mode is predictable: agents layered on top of an unchanged process, standing on data nobody trusts. Agents amplify whatever they're given — ungoverned data produces outputs nobody believes, and a workflow needing four systems and three approvals can't be saved by a smarter agent.[5]

The shift that works is from human-by-default — every action initiated and approved by a person — to human-by-exception, where only the decisions that matter reach one.[5]

Where this lands in your ERP

For mid-market enterprises, the agentic question is an architecture question with two honest answers. Unify operations under a single roof — a modern ERP like Odoo, with HR, accounting, procure-to-pay, CRM, and helpdesk sharing one data model — so agents inherit clean, trusted context for free. Or keep the legacy core and build a custom intelligence layer over it via APIs, modernizing without a teardown.

Either way the sequencing is the same: fix the data and the process first, embed agents second.

And keep the unit economics honest — agentic workflows consume far more tokens than chat. Choosing capable, cost-efficient models such as DeepSeek and Kimi over the most expensive tier is often what makes a use case viable at all.

The bottom line

Agentic AI is not a tool rollout. It is an operating-model shift, and the organizations treating it as the former are the 40% Gartner expects to cancel.[1]

Take one small, defensible next step: pick a single bottleneck with an existing P&L owner, baseline its cost and cycle time before you build anything, redesign the process around human-by-exception, then embed an agent and measure monthly.

That is how the 5% got there.[2] It is not a technology strategy first. It is a work-redesign strategy that happens to be powered by agents.

Sources

1. Gartner: Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 (June 2025)
2. MIT NANDA: The GenAI Divide — State of AI in Business 2025 (July 2025)
3. Klarna: AI assistant handles two-thirds of customer service chats in its first month (Feb 2024)
4. CIO Dive: Why Walmart is rolling out AI to 2M employees (April 2026)
5. HSO: The State of Enterprise Agentic AI (2026 briefing)
6. AI to ROI: Where Agentic AI is and is Not Working Well (April 2026)
7. IntuitionLabs: Enterprise AI Deployment Failures and Outcomes in 2026
8. Reuters: Over 40% of agentic AI projects will be scrapped by 2027, Gartner says (June 2025)
9. OpenAI: Klarna's AI assistant does the work of 700 full-time agents
10. A4BEE: From Adoption to Redesign — What 12 Agentic-AI Reports Reveal
11. Fortune: Walmart has a message for its 2.1 million workers — AI to improve job, not take it (June 2026)
12. Business Insider: Walmart workers' new responsibility — correcting AI errors (July 2026)