Agentic AI and Process Optimization: Why 94% of Companies See No ROI

Table of Contents

• Introduction: The 2026 Paradox
• 88% Use AI, 56% of CEOs See Nothing Coming
• McKinsey’s Real Finding: It’s Not the Same Companies Winning
• Why 70% of Companies Skip the Step That Matters Most
• From Generative AI to Agentic AI: The Real Shift in Nature
• Why 70 to 95% of AI Projects Fail — and It’s Never the Tech
• The AI Act, the Digital Omnibus, and What Actually Changes in 2026
• The Product Owner’s Role in an AI-Driven Process Redesign
• A Process-First Framing Method in Five Steps
• Conclusion: The Bottleneck Isn’t Technical
• FAQ
• Expert Insight
• Let’s Move Forward
• References

Introduction: The 2026 Paradox

88% of companies now use artificial intelligence in at least one business function. And yet, 56% of CEOs surveyed by PwC at the 2026 Davos summit report that AI has generated, to date, neither measurable revenue growth nor cost reduction. These two figures aren’t contradictory — they describe the same reality from two different angles, and that reality is the real subject of this article.

Across fifteen years framing digital transformation, data, and AI projects — from Product Ownership to multi-project program governance — I’ve watched this paradox repeat with troubling regularity: tool adoption was never the problem. The problem has always been what happens, or doesn’t happen, around the tool.

2026 makes this paradox particularly visible because agentic AI — agents capable of executing complete sequences of business tasks, not just answering questions — has changed the scale of damage a poorly framed deployment can cause. A poorly framed chatbot produces a mediocre answer. A poorly framed agent executes a wrong action, potentially cascading, on a real business process.

This article isn’t trying to settle the « for or against AI in the enterprise » debate — that debate is largely overtaken by the facts, given the scale of investment and deployment already underway. The goal is narrower and more useful: pinpoint exactly what organizations capturing real value are doing differently, so a Product Owner, Business Analyst, or executive can replicate that pattern rather than discover it through trial and error, project after project.

What follows draws on real framing engagements across digital transformation, data, and AI programs — not vendor talking points. The intent is to give a decision-maker or a fellow practitioner a concrete, testable lens for the next AI initiative on the roadmap, grounded in what the 2026 data actually shows rather than what the market narrative promises.

88% Use AI, 56% of CEOs See Nothing Coming

The macro numbers for 2026 leave no doubt about the scale of adoption. Global AI investment reached $581.7 billion in 2025, up 130% year over year. Gartner forecasts worldwide AI spending of $2.52 trillion for 2026, up 44% year over year, with generative AI model spending specifically growing 80.8%. The AI agent market alone is projected to reach $11.5 billion in 2026, up from $7.8 billion in 2025.

Against these amounts, PwC’s finding — 56% of executives reporting no measurable impact — should be a collective alarm bell. It isn’t yet, partly because McKinsey’s answer to this paradox is reassuring on the surface: a minority of companies, roughly 6% of organizations according to their State of AI 2025 segmentation, capture an EBIT impact above 5% from AI. That figure is both encouraging — the value genuinely exists — and sobering: it remains rare.

Agentic startups and frontier models are advancing in parallel with this paradox. Anthropic claims more than 15,000 Claude for Work customers and a valuation that surpassed OpenAI’s in 2026. Specialized startups like Sierra, Decagon, and Cresta have combined $1.5 billion in ARR, mostly in customer service. The technology is clearly not the limiting factor — the market’s best tools exist, are accessible, and are already deployed at scale.

McKinsey’s Real Finding: It’s Not the Same Companies Winning

The question that matters, then, isn’t « why doesn’t AI work » — it works, for some. The real question is: what are these top 6% of high-performing organizations doing differently?

McKinsey’s answer, drawn from its State of AI 2025 study covering 1,900 organizations and cross-referencing 25 attributes predictive of AI’s financial impact, is unambiguous: these companies aren’t doing anything technologically different. They’re doing something organizationally different. They redesign their workflows around AI, instead of bolting AI onto existing processes without rethinking them.

Among the 25 attributes McKinsey studied, workflow redesign ranks highest in correlation with financial impact. And yet, a separate analysis building on the same body of work notes that 70% of companies skip this step entirely. They deploy AI on fragmented, poorly documented processes where humans have spent years compensating for dysfunctions that the tool itself cannot compensate for — it executes them faithfully, errors included.

High-performing companies are, per McKinsey, nearly three times more likely to fundamentally redesign their workflows in their AI approach than others. They’re also more likely to have defined clear processes determining when and how human validation should intervene on model outputs — and their senior leaders show real, not just declarative, commitment to AI initiatives.

That last point deserves emphasis, since it’s often the most overlooked factor in project failure diagnostics. Leadership commitment that stops at an approved budget and an internal press release isn’t the same commitment as leadership willing to rethink a business process established over years, even if it means shifting responsibilities or challenging deeply ingrained work habits. This second, structurally more uncomfortable, level of commitment is exactly what McKinsey consistently finds in high-performing organizations. On the ground, this commitment shows up very concretely as an executive sponsor’s willingness to sacrifice a short-term deadline for stronger framing, rather than caving to calendar pressure.

There’s a temptation to read the McKinsey finding as purely a data-quality story — clean up your data, redesign your pipelines, and the AI project follows. That reading misses the harder part. Workflow redesign isn’t a technical cleanup exercise; it’s an organizational negotiation about who owns a decision, what gets automated versus what stays judgment-based, and which shortcuts a team has quietly relied on for years that no longer hold once an agent takes over the step. That negotiation is precisely what a Product Owner is equipped to run, and what a purely technical rollout consistently skips.

Why 70% of Companies Skip the Step That Matters Most

48% of organizations cite the difficulty of locating their own data as the primary obstacle to their AI strategy, according to a 2026 Deloitte report. An analysis synthesizing this body of work states the diagnosis plainly: this isn’t a data problem. It’s a workflow problem. When a business process is structured around human judgment and informal adjustments, the resulting data has no usable context — no retrieval system can find it, because it isn’t stored with metadata a system recognizes.

The result of skipping this step is measurable and severe. 95% of enterprise generative AI pilots deliver zero measurable P&L impact, according to data compiled by MIT NANDA. 42% of companies abandoned most of their AI initiatives in 2025. These aren’t isolated failures — they’re consistent symptoms of the same failure mode, repeated across the market.

A concrete example cited in recent literature illustrates the mechanism: picture a loan origination process at a mid-sized lender. If that process has relied for years on the intuition and judgment of experienced analysts rather than documented rules, plugging an AI agent into it doesn’t make it faster or more reliable — the agent inherits the process’s ambiguity, without the intuition that previously compensated for it.

From Generative AI to Agentic AI: The Real Shift in Nature

There’s a structural difference between generative AI, which excels at creating content, making predictions, or answering questions, and agentic AI, which goes further: it doesn’t just produce insights, it has the ability to act on them. AI agents can collaborate with each other, make decisions, and trigger actions across different business functions — AI moves from suggestion to execution.

Gartner forecasts that by 2028, a third of enterprise software applications will incorporate agentic AI, translating into 15% of daily operational decisions being made autonomously. This isn’t an incremental improvement on existing capabilities — it’s a structural change in how enterprise operations function.

The market confirms this shift with cases already in production, not just announcements. C.H. Robinson’s fleet of more than 30 AI agents processed over 3 million freight shipment tasks in 2025, generating a documented 30% productivity increase. Walmart deploys autonomous agents in inventory management, supplier negotiation, and logistics — with verifiable financial results: AI-driven inventory precision enabled 5% sales growth for only a 2.6% increase in inventory.

But this shift toward autonomous execution also changes the nature of risk. One study finds that only one in five companies has a mature governance model for its autonomous agents. And Gartner projects that more than 40% of agentic AI projects will be abandoned by the end of 2027, due to runaway costs, unclear business value, and insufficient risk controls — a phenomenon some analysts already refer to as « agent washing, » the agentic equivalent of greenwashing.

That term isn’t incidental. It describes a precise market mechanism: as « deploying AI agents » becomes a competitive positioning argument for solution vendors, the pressure to announce agentic deployments far outpaces the pressure to frame them correctly. The visible market outcome is a proliferation of agentic pilots presented internally as successes, but which meet none of the value-measurement criteria an investor or executive committee would demand for any other type of project.

The distinction matters for how a Product Owner should scope a first agentic project. A generative AI pilot that misfires produces a bad draft someone edits or discards — the cost is measured in minutes. An agentic pilot that misfires can place an order, approve a payment, or reject a candidate before anyone reviews the output — the cost is measured in dollars, reputational exposure, or in regulated sectors, compliance findings. That asymmetry is exactly why the framing discipline described later in this article matters more for agentic projects than it ever did for earlier generations of AI tooling.

Why 70 to 95% of AI Projects Fail — and It’s Never the Tech

Depending on the study, between 70 and 95% of enterprise AI projects fail to meet their objectives. The 95% figure, from an MIT report relayed by Fortune, specifically measures generative AI pilots that never reach at-scale deployment with measurable impact, over a short time horizon. Consolidated analyses from Gartner, McKinsey, and BCG place the failure rate in a broader range, between 70 and 85%, depending on the definition of failure used.

In every case, the diagnosis converges: failure is almost always tied to framing, data, and method — almost never to the technology itself. The recurring documented causes are a vague objective from the outset, poor-quality or unusable data, no real integration into existing processes, lack of team skills and buy-in, neglected governance, and above all no measurement of return on investment once the pilot launches.

POC purgatory — projects that stay indefinitely stuck at the proof-of-concept stage — has become a documented norm rather than an exception. Gartner found that at least 50% of generative AI projects were abandoned after the POC phase by late 2025; IDC, on a different scope, puts the figure as high as 88% of projects that remain stuck without ever reaching production. 70% of organizations run into data-related difficulties — governance, integration speed, insufficient training data — according to the same McKinsey research already cited. That 70% figure is, notably, close to the same proportion of companies that, earlier in this article, skip the workflow redesign step — the two statistics likely describe, at least in part, the same population of organizations and the same methodological gap.

What distinguishes the 5 to 30% of projects that succeed is, again, never a more powerful model. It’s precise targeting of a measurable, bounded use case, serious data preparation ahead of any deployment, and continuous measurement of the value generated — not just at launch, but throughout the project’s lifecycle.

It’s worth dwelling on a detail often glossed over quickly in these analyses: measuring value isn’t a one-off exercise done three months after launch to justify the budget already spent. Successful organizations treat this measurement as a continuous flow, built into project governance from day one, with regular checkpoints that allow adjusting — or even stopping — a deployment before it has consumed disproportionate resources relative to its actual value. That’s a classic program portfolio management reflex, rarely applied with the same rigor to AI projects as to any other technology investment.

The AI Act, the Digital Omnibus, and What Actually Changes in 2026

On the regulatory front, 2026 is a pivotal year, though less abrupt than many companies anticipated. The European AI regulation (EU 2024/1689) has imposed an AI training obligation since February 2025 for any company deploying a covered system. Article 50’s transparency obligations have applied since August 2, 2026. But compliance for high-risk systems under Annex III, initially expected earlier, was pushed back to December 2, 2027 — a provisional agreement from May 7, 2026, known as the « Digital Omnibus package, » adjusted the timeline for several obligations to give companies more time.

This delay doesn’t exempt organizations from framing their deployments now. Article 14 of the AI Act remains the central principle for any high-risk system: it requires that systems allow a human to understand, monitor, and interrupt their actions. A fully autonomous agent making high-impact decisions — hiring, scoring, significant financial decisions — isn’t compliant, regardless of the deferred enforcement timeline.

A technical point worth noting for any Product Owner framing an agentic AI project: it’s the agent’s action, not its level of autonomy per se, that determines regulatory risk. An agent drafting emails falls under limited risk. An agent screening applications or making scoring decisions falls under high risk per Annex III, regardless of the underlying technical sophistication. Multi-agent systems further complicate compliance: when several agents interact, the chain of accountability must be clearly established between providers and deployers — a topic standard technical documentation for an AI project rarely covers satisfactorily.

This regulatory framework isn’t unique to Europe, even though it’s currently the most developed. For an organization operating internationally, the question of regulatory framing now arises at agent design time, not at the moment of deployment in a given jurisdiction — the costs of retroactive compliance, once an agent is already in production across multiple markets, far exceed those of framing anticipated as early as the product specification phase.

For a Product Owner working across US, Canadian, or other non-EU markets, the practical takeaway isn’t to treat the AI Act as someone else’s regulation. Multinational organizations increasingly default to the strictest applicable standard across their footprint rather than maintaining separate compliance postures per market, simply because the engineering cost of maintaining two versions of the same agent — one EU-compliant, one not — rarely justifies the savings. Framing an agent to Article 14 standards from day one, regardless of where it first ships, is turning into the pragmatic default rather than the cautious exception.

The Product Owner’s Role in an AI-Driven Process Redesign

Faced with this landscape, a Product Owner’s or Business Analyst’s role on an agentic AI project is structurally not that of a technical project manager overseeing a tool rollout. It’s that of a process architect, who must answer a question neither data teams nor IT teams can settle alone: which process deserves to be rethought, and how far should that redesign go before an agent gets plugged into it?

This is exactly what the McKinsey data makes visible: workflow redesign is the factor most correlated with an AI project’s financial success, but it’s also the most demanding in terms of skill — it requires understanding the business well enough to distinguish a process step worth automating from a step that informally compensates for a deeper design flaw.

A Process-First Framing Method in Five Steps

Across my AI and data framing engagements, five steps systematically structure the move from a pilot to a deployment that delivers on its promises:

1. Map the process before choosing the tool. Document how the process actually works — not how it’s supposed to work on paper — before evaluating a single technical solution. This is the reverse of the approach most commonly observed in the field, which starts from the tool.
2. Pick a bounded, repetitive, measurable use case. A high-volume use case with a quantifiable outcome is the best guarantee of proving — or disproving — the project’s value quickly, rather than discovering it eighteen months later.
3. Prepare and qualify data before deployment. An agent is only as good as the information it handles; skipping this step is the most documented cause of failure across every study cited above.
4. Define the agent’s scope and guardrails. What it can do alone, what it must always escalate to a human — accounting, for any high-risk system under the AI Act, for the Article 14 obligation of effective human oversight.
5. Measure continuously, not just at launch. Tracking generated value must start on day one of deployment and continue over time — organizations that track their synergies from day one show significantly higher success rates than those that only set up this tracking later.

This method isn’t a methodological luxury. It answers, point by point, the failure causes documented by McKinsey, Gartner, and MIT NANDA: the workflow redesign before the tool answers the number-one identified failure cause; the bounded use case answers POC purgatory; data preparation answers the governance problem 70% of organizations run into; the guardrails answer both operational risk and regulatory obligation; continuous measurement answers the fact that 56% of executives still see no impact, precisely because they never measured it correctly. None of these five steps depends on a particular technology choice — they apply identically whether the agent is built on Claude, GPT, or a proprietary model.

Conclusion: The Bottleneck Isn’t Technical

The 2026 paradox — mass adoption, rare ROI — isn’t going to resolve itself as long as most companies keep treating AI as a layer to bolt onto existing processes rather than an occasion to rethink them. The 6% of organizations capturing significant EBIT impact don’t hold a secret, better-performing model. They simply did the organizational work the remaining 94% skipped.

For a Product Owner or Business Analyst, that’s good news disguised as a constraint: the bottleneck isn’t technology, which is available, mature, and increasingly accessible. It’s the ability to frame a process, arbitrate what needs to change before plugging in an agent, and honestly measure what happens next — exactly the job no model, however powerful, replaces.

This reading also changes how an executive should evaluate a candidate or a mission proposal for an AI project. The most revealing question isn’t « which tools do you know » — the answer is almost always the same list of product names from one candidate to the next. The question that truly distinguishes a profile capable of joining the top 6% of high-performing organizations is: « tell me about a process you had to rethink before you could plug AI into it, and what you discovered doing it. » That experience, rarely highlighted on a résumé, best predicts whether a project will deliver documented ROI or join the abandonment statistics — far more than a technical certification or familiarity with a particular tool.

It’s also worth naming the counterintuitive part directly: none of this argues for moving slower. The 94% aren’t failing because they moved too fast — they’re failing because they moved fast on the wrong layer, activating tools before touching the process underneath. Moving fast on the process layer first, then deploying the agent onto a workflow that’s actually ready for it, produces faster time-to-value than the reverse sequence, not slower. The organizations in McKinsey’s top 6% aren’t the cautious ones; they’re the ones that sequenced the work correctly.

FAQ

Why do most agentic AI projects fail despite increasingly powerful tools?
Because failure almost always comes from framing, data, and process — not the model. Companies that achieve real ROI redesign their workflows around AI rather than adding it to existing, unrevised processes.

Do we need to wait until 2027 to comply with the AI Act on agentic AI projects?
No. Only compliance for Annex III high-risk systems has been deferred to December 2027; transparency and effective human oversight obligations apply now for any covered system.

What’s the concrete difference between generative AI and agentic AI in terms of project risk?
Generative AI produces suggestions a human validates before action; agentic AI directly executes sequences of business tasks, which changes the scale of consequences from poor framing — an error can propagate before any human intervention.

How long does it take for an agentic AI project to demonstrate real ROI?
This varies significantly by use case and upfront preparation, but organizations that track generated value from day one of deployment show significantly higher success rates than those that wait several months to assess impact.

Expert Insight

What strikes me most about the 2026 data isn’t the failure rate itself — it’s remained remarkably stable across technology generations, from big data to cloud to generative AI. It’s the consistency of the diagnosis: every time, the obstacle limiting value isn’t the sophistication of the tool, but the clarity of the process it’s handed. An AI agent plugged into a fuzzy process doesn’t make that process clearer; it executes the confusion faster. The Product Owner’s job, in this context, is precisely to do that sorting upfront — and it’s a job that can’t be outsourced to the tool itself, however powerful it’s marketed to be. Fifteen years across different technology waves has taught me one durable lesson: the tool changes every cycle, but the discipline of asking « what process are we actually automating, and does it deserve to survive in its current form » never goes out of date.

Let’s Move Forward

Considering an agentic AI project and want to avoid joining the abandonment statistics rather than the documented-ROI ones? Process-first framing — mapping, bounded use case, guardrails, continuous measurement — makes all the difference before you even pick a tool. Let’s talk about your data and AI context.

References

• PwC — Davos 2026 CEO Survey
• McKinsey — The State of AI in 2025: Agents, Innovation, and Transformation (1,900 organizations, 25 attributes)
• McKinsey — AI-driven enterprise transformation for workflow automation
• Deloitte — 2026 report on data findability as an AI strategy obstacle
• MIT NANDA — State of AI in Business, data relayed by Fortune and derived analyses
• Gartner — AI spending forecast 2026; agentic AI application adoption by 2028
• Gartner — GenAI pilot abandonment rate post-POC
• IDC — POC purgatory data on blocked AI projects
• Deloitte Canada — The Agentic Supply Chain (C.H. Robinson, Walmart cases)
• SAP — Agentic AI in Global Supply Chain
• Stema Partners — AI Agent in Enterprise: Use Cases & ROI 2026, Process First methodology
• Ayi Nedjimi Consultants — AI Act and GDPR for AI Agents 2026
• aiacto — AI Agents in Enterprise: What Are the AI Act Obligations (European Commission, February 2026 guidelines)
• iaCockpit — AI Agents in SMBs: What the August 2, 2026 AI Act Changes (Digital Omnibus agreement, May 7, 2026)
• KlapDoc — AI Act Enterprise Obligations: Regulating Agentic AI in 2026
• alexitauzin.com — AI Agents in Enterprise 2026: Mass Adoption, Rare ROI, the 2026 Paradox