A practitioner's read on why enterprise AI spend is outrunning enterprise AI returns — and where the constraint actually sits.
The views below are my own, drawn from public research on B2B AI adoption and from what we have lived through deploying these tools inside Allari. Vendors and analysts would frame parts of this differently, and I welcome the argument.
A split-screen market
Sit with the 2026 corporate landscape for a moment and you see two pictures that refuse to reconcile. In one, capital markets are euphoric — but the euphoria is concentrated in a narrow band of frontier model labs, hyperscalers, and the chipmakers underneath them. In the other, the operating businesses that are supposed to be the customers for all of this are mostly still circling the pool: a license purchased here, a siloed pilot there, and a slow accumulation of sober reports that adoption, sustained usage, and economic return are all lagging the promise.
Both pictures are true at once. The technology has advanced at a genuinely startling pace in the three-plus years since ChatGPT went public. What has not kept pace is organizational maturity. Expectations for immediate, bottom-line results have run far ahead of where most enterprises actually are, and the transition from static copilots and prompting toward autonomous, agentic action is only just beginning.
The bottleneck is not the model
Driven by a wish to look past the noise, I used modern AI itself to do the kind of synthesis that would once have taken me the better part of a week — scanning sources, reading dense material, capturing findings by hand. That alone is a real productivity gain: without these tools I would not have attempted research of this depth, and the output would have been thinner if I had. What the research surfaced, read against our own experience, is the part worth sitting with.
The primary limitation in achieving meaningful economic returns from AI is not a failure of the technology. It is a failure of organizational design.
The constraint is not the AI. It is us. Value, implementation, and success rest on human execution, process redesign, and organizational maturity. Adoption is driven by the human capacity to define the problem, rank the problems, and find the right place for AI inside a solution. The fantasy of typing "go find how AI can deliver ROI for my company and implement it" is not a product you can buy. That humans still hold full agency over this is good news, and I do not see it changing soon.
Adoption is real, but it is shallow
Look at how the technology is actually diffusing and the gap between activity and integration comes into focus. After the U.S. Census Bureau's Business Trends and Outlook Survey widened its lens in early 2026 from AI in "core production" to AI in "any business function," the reported baseline of U.S. firms using AI roughly doubled to about 20 percent. Federal Reserve district data lands in a similar range. By scale, large enterprises (250-plus employees) report roughly 30 percent, while the smallest firms sit closer to 17 percent — held back by capital, scarce technical talent, integration cost, and the absence of clean, structured data.
The number that matters is not the headline percentage but what "adoption" means underneath it. As measured today, adoption mostly captures usage in its infancy: a software license bought, a department running a pilot to draft emails or summarize meetings. A firm counted as an "adopter" has rarely restructured anything. Mistaking that for maturity is the central error of the moment.
There is one genuine counter-signal: companies born into the AI era build it into their foundations from day one. By various measures, recent startup cohorts reach adoption milestones in months that took earlier cohorts years — a direct consequence of cloud-delivered, AI-as-a-service models collapsing the barriers to entry.
The productivity paradox, in plain numbers
Here is the tension that should hold a CFO's attention. Local, task-level efficiency gains are widely reported — a majority of organizations say they see them. Those gains rarely climb up to the P&L. Wharton research found that even as corporate AI deployments surged, only roughly 12 to 18 percent of companies captured meaningful bottom-line ROI. MIT's NANDA group reported that around 95 percent of enterprise AI pilots in 2025 produced no measurable P&L impact. Treat these as the researchers' figures, not gospel — but the direction is unambiguous.
That gap is what you get when surface-level license adoption is mistaken for workflow integration. The failures cluster into a few human-centric patterns:
- The "magic wand" problem. Per Deloitte's reporting, only about a third of enterprises are using AI to transform core processes or actively redesign workflows; the rest are overlaying generative AI on top of fragmented, legacy operating models. Automate a chaotic, ill-defined process and you have simply made the chaos run faster, at higher cost.
- Imprecise problem definition. Projects launch under vague mandates to "improve productivity" rather than being anchored to a specific, measurable pain point. Scalability and complexity reduction rarely make leaders' top-three outcomes.
- The data-hygiene gap. Poor data quality is cited as hampering value capture for the large majority of leaders, yet only about half establish a clean, structured foundation before scaling. AI is a mirror of organizational health: point advanced agents at fragmented data and undefined workflows, and they amplify the dysfunction.
From copilots to agents — meeting the real architecture
Through 2024 and 2025 most deployments were static, human-directed copilots. By early 2026 the market began moving toward agentic systems that reason, use tools, and execute multi-step work, with roughly a quarter of enterprise buyers reporting at least moderate use. Vendors project a large orchestration market and, further out, "agentic commerce" in which AI agents negotiate and execute B2B transactions at enormous scale. Those forward numbers are vendor and analyst projections — useful as a direction, not a forecast to budget against.
The collision is with the architecture these agents land in. The average enterprise runs hundreds of distinct applications with only a minority integrated. Reported consequences: a large share of deployed agents operating in disconnected silos, a meaningful fraction of paid AI licenses sitting inactive for 90-plus days, and employees still spending hours each week moving data by hand between the very systems AI was meant to bridge. Procurement has outrun integration.
The labor story is more complicated than the headline
The 2026 reality contradicts the simple narrative of immediate, broad job replacement. Of the large wave of U.S. corporate layoffs in 2025, only a small minority were attributed to AI; the rest trace to macro pressures and post-pandemic over-hiring. AI is sometimes a convenient label for cost cuts leaders would rather not explain otherwise.
Targeted, technology-led reductions are real in specific functions — several large tech firms have trimmed support and administrative roles as automation absorbed routine queries and code generation. But Forrester's research suggests a majority of employers will come to regret AI-attributed layoffs, with a large share of those workers quietly rehired once it becomes clear that automation cannot carry complex problem-solving and relationship work. Meanwhile, cutting entry-level roles starves the talent pipeline, and a wide AI skills gap persists. Again, treat the specific percentages as the source's, not as settled fact.
The capital-to-value disconnect
Across the global technology market, the gap between capital invested in AI and revenue actually generated from it has, by some estimates, ballooned into the hundreds of billions of dollars. The pressure on leaders is rising — a majority report heavier scrutiny to prove ROI, and many investors expect returns on a horizon shorter than the work realistically allows. Spending keeps accelerating anyway; BCG's AI Radar reports corporations roughly doubling AI's share of revenue in 2026.
The uncomfortable counterpoint: only about a third of initiatives are meeting internal ROI targets, more than half of CEOs report no revenue gain yet, and Gartner projects a sharp rise in project cancellations as integration cost, poor data, and unclear value catch up with the spend. Today, most of AI's financial contribution shows up as cost avoidance and efficiency, not new revenue.
Where it is actually working
The wins are real and they are local. In supply chains, agents evaluate combinations of partial shipments, transfers, and re-sequencing in real time, breaking trade-offs that planners used to manage by hand; reported outcomes include strong logistics-automation ROI and large administrative-overhead reductions at a global consumer-goods enterprise that moved from manual to agent-led replenishment. In go-to-market, where sellers historically spend only about a quarter of their time with buyers, automation of CRM entry, research, and reporting is reported to give hours back each week. The pattern holds: AI is diffusing through horizontal tasks first, not transforming proprietary core operations.
What I would tell a CIO or CFO to do
The long-run trajectory is not in doubt — affordable, cloud-delivered AI is leveling the field, and smaller, agile competitors can now match incumbents' operational capability without the legacy debt. That is precisely why the discipline matters now. A few mandates I would hold to:
- Reject the magic wand. No tool fixes a broken process. Do the friction-heavy work of mapping workflows end to end and redesigning them around a hybrid human-AI model before you deploy.
- Reclaim precision. Anchor every pilot to an explicit, measurable pain point and map the solution's logic by hand before a single agent goes live.
- Fund maturity, not just licenses. Inactive seats are proof that acquisition has outpaced capability. Shift budget toward change management, training, and incentive realignment, led by business-unit owners who know how the work actually flows.
- Assign human ownership of data. Skip the wait for a perfect enterprise data lake. Use a thin-slice approach: pick a high-value use case and task a specific team with cleaning exactly the data that workflow needs.
- Govern the agents. As copilots give way to autonomous networks, hardcode human validation thresholds and keep the workflows auditable and explainable.
The success of AI does not belong to the machine. It belongs to the organization that prepares the ground for it. The technology is far enough along; the path to value is visible. The open question is whether an organization has the courage to redesign itself to walk it.
Analysis by Ravi Madhavan, offered as one practitioner's reading of the state of B2B AI. Originally published on LinkedIn. Substantive disagreement welcome. Allari runs the ERP support layer for JDE, SAP, Oracle Fusion, and NetSuite estates — finding the repeatable manual work these reports describe and finding where it can be automated.
