Agentic AI ROI: Are Companies Actually Seeing Results?

How Are Companies Evaluating Agentic AI Tools Right Now, And Are They Actually Working?

Ask ten executives whether agentic AI is delivering results, and you’ll get two completely different answers depending on who you ask. Ask the team that bought a narrow, well-scoped tool for a specific back-office workflow, and you’ll likely hear about real time savings. Ask the team eighteen months into a broad “AI transformation” pilot with no clear metric attached, and you’ll likely hear a much more uncomfortable answer — because the data increasingly shows those two outcomes aren’t random. They track a specific, identifiable pattern in how companies chose and evaluated the tool in the first place.

This piece works through what that pattern actually looks like: the real data on how many agentic AI projects deliver measurable value versus how many don’t, how the companies getting real results are evaluating and deploying these tools differently, and a practical framework for doing this yourself. At CodeStore, we get pulled into this conversation most often after a client’s first pilot didn’t produce the results they expected — see our agentic AI development services or contact us if that’s where you’re starting from.

The Number Everyone’s Citing, and What It Actually Means

The most-quoted data point in this conversation comes from MIT’s Project NANDA, whose July 2025 report — The GenAI Divide: State of AI in Business 2025 — found that despite $30–40 billion in enterprise generative AI investment, 95% of organizations saw zero measurable impact on profit and loss. Just 5% of pilots were extracting significant, measurable value. The research, based on 52 structured interviews, a survey of 153 senior leaders, and analysis of over 300 public AI deployments, is explicit that this divide isn’t caused by model quality or regulation — the report attributes it to what the authors call a “learning gap”: most deployed systems can’t retain feedback, adapt to context, or integrate into how work actually gets done.

It’s worth being precise about what that 95% figure covers before applying it directly to agentic AI specifically: MIT’s study measured broad generative AI adoption in 2025, not agentic systems exclusively. But the finding matters directly for this conversation for two reasons. First, the report explicitly frames agentic systems — tools with memory, persistent context, and the ability to act rather than just respond — as the direction most likely to close this gap, not a separate category exempt from it. Second, more recent data specific to agentic deployments shows a similar, though somewhat less extreme, pattern.

Is Agentic AI Actually Different From the Broader GenAI Story?

The honest answer is: better in some respects, but still following the same underlying pattern of a small group of disciplined adopters pulling away from everyone else. Capgemini Research Institute’s 2025 survey of 1,500 executives found that while 93% of business leaders believe scaling AI agents over the next 12 months will provide a durable competitive edge, only about 2% of organizations have actually deployed AI agents at scale, with 12% at partial scale and 23% still at the pilot stage. That’s a wide gap between stated confidence and actual production deployment — a close cousin of MIT’s “high adoption, low transformation” framing.

Gartner’s own forecast reinforces the same caution: more than 40% of agentic AI projects are projected to be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls — a specific, dated warning about exactly the failure mode MIT documented a broader version of.

At the same time, where agentic tools are properly scoped, the productivity numbers are real. Enterprise pilots in automation-heavy functions have reported productivity gains of up to 60%, compared to roughly 40% for search-only generative AI deployments — because an agentic system that can autonomously execute a task (resolving a ticket, generating a report, coordinating a workflow) produces completed work rather than only saving someone’s research time. The gap between the disappointing headline numbers and these stronger results comes down almost entirely to how narrowly and deliberately the use case was chosen and measured.

How Companies Are Actually Evaluating These Tools Right Now

Three shifts show up consistently across the organizations getting real value, based on current guidance from enterprise AI evaluation research:

Baseline metrics before deployment, not after. Capturing pre-deployment data — average time per task, cost per task including labor and tooling, error rate, throughput volume, and escalation rate — for two to four weeks before an agent goes live is what makes an ROI claim afterward credible. Organizations that skip this step consistently struggle to prove value even when the agent is genuinely delivering it, because they have nothing to compare the result against.

ROI measured per workflow, not per program. A practical formula gaining traction: (total benefit from the deployment, minus total cost of platform, implementation, and maintenance) divided by total cost — applied to a single workflow, not averaged across an entire AI program. This matters because averaging hides the truth: a company can have one genuinely excellent agentic deployment and four mediocre ones, and a program-wide average would understate the first and overstate the rest.

A broader definition of what counts as return. Salesforce research on this shift found that 61% of CFOs say AI agents are changing how they evaluate ROI entirely, expanding measurement beyond traditional cost metrics to a broader set of business outcomes. A four-pillar approach is becoming common in more rigorous evaluations: hard-dollar cost takeout, revenue acceleration, quality and risk reduction, and speed or throughput gains — with the expectation that a legitimate deployment should show measurable movement on at least two of the four, not just one easily-gamed metric like “number of employees given access.”

Vendor evaluation now includes lock-in and governance, not just capability. Enterprise buyers evaluating agentic AI platforms are increasingly checking whether agents can be exported, whether multiple LLM providers are supported, and whether the orchestration layer uses open standards — because proprietary, single-vendor lock-in is now understood as a real cost, not just a technical footnote. Evidence of investment in audit trails and compliance certifications (SOC 2, HIPAA, or EU AI Act readiness, depending on the industry) is treated as a signal of enterprise readiness in a way it wasn’t a year or two ago.

Where the Real Productivity Gains Are Actually Showing Up

Where the Real Productivity Gains Are Actually Showing Up?

MIT’s research is specific about where the 5% of successful deployments are concentrated: back-office automation delivers the highest returns, largely by reducing outsourcing costs and streamlining processes that were already well-defined before the AI was introduced. This tracks closely with the general lesson from earlier in this content series — the strongest agentic AI use cases are high-volume, meaningfully variable, and produce output that’s easy to verify, which describes back-office and administrative workflows far better than it describes open-ended strategic work.

Function-level payback timelines back this up. Finance functions currently show the fastest average payback period among enterprise agentic deployments — around eight months in current data — largely because financial workflows tend to have clean, structured data and clearly measurable outcomes (an invoice processed correctly or not, a reconciliation completed or flagged). Software engineering and IT show a similar pattern: McKinsey’s research on agentic AI scaling found organizations in these functions reporting double-digit cost reductions, and more granular benchmarking from a 2026 study of 150 enterprises found meaningful reductions in routine coding time and shorter code review cycles among organizations using agentic coding tools properly.

So: Productive Automation, or Just a Waste of Money?

The honest, unsatisfying answer is: it’s genuinely both, and which one you get depends almost entirely on execution rather than on the technology category itself. The organizations landing in MIT’s failing 95% and Gartner’s projected 40%-plus cancellation rate share a consistent set of mistakes: they measure inputs instead of outcomes (“we rolled agents out to 5,000 employees” is adoption, not ROI), they point agents at workflows with no dollar value clearly attached to them, and they never captured a baseline, so there’s no way to prove what actually changed.

The organizations getting real value share the opposite pattern: they picked a single, well-defined, high-volume workflow; they measured its cost and performance before deployment; they redesigned the workflow around what the agent could actually do rather than layering the agent on top of an unchanged process; and they held the deployment to a specific, pre-agreed metric rather than a vague sense of “productivity.” McKinsey’s broader research on agentic AI scaling found this same distinction — organizations that redesigned workflows around AI, rather than adding it to existing processes, were far more likely to be among the group actually reporting cost reductions.

There’s also a quieter, less-discussed finding in MIT’s research worth taking seriously: employees at many organizations are already getting real value from generative and agentic tools on their own, informally, ahead of — and sometimes despite — their company’s official pilot program. This “shadow AI” pattern suggests that a meaningful part of the “95% failure” story is really a story about formal program design lagging behind what individual employees have already figured out works, rather than a story about the underlying technology not working at all.

Governance Belongs Inside the Evaluation, Not After It

NIST’s AI Risk Management Framework — built around four functions (Govern, Map, Measure, Manage) — is increasingly treated as part of the evaluation process itself, not a separate compliance step that happens after a tool is chosen. This matters for the ROI conversation directly: retrofitting governance and audit trails onto an agent already in production is consistently more expensive and more disruptive than building them in from the pilot stage, and a tool that can’t support the level of oversight your use case requires isn’t actually the cheaper option, even if its sticker price is lower.

A Practical Framework for Evaluating an Agentic AI Tool

A Practical Framework for Evaluating an Agentic AI Tool

1. Capture a real baseline before you deploy anything. Time per task, cost per task, error rate, and volume, for two to four weeks minimum. Without this, no ROI claim afterward is defensible.
2. Pick one workflow with a dollar value clearly attached to it. “Improve productivity” is not a workflow. “Reduce average handling time on tier-1 support tickets” is.
3. Measure ROI per workflow, not per program. Averaging hides both your best and worst deployments.
4. Redesign the workflow around the agent, not the other way around. The organizations getting real returns rebuilt the process; the ones that didn’t just added a tool to an unchanged one.
5. Build governance in from day one, including how you’ll monitor for behavior drift when the underlying model updates, and what level of human oversight the specific use case actually requires.

At CodeStore, this is the sequence we walk clients through before recommending a build — starting with whether a workflow is genuinely ready for this kind of evaluation in the first place. Contact us if you’re trying to figure out whether your first pilot is set up to succeed, or explore our agentic AI development services.

Common Misconceptions

“A 95% failure rate means the technology doesn’t work.” MIT’s own researchers attribute the divide to organizational integration failures, not model quality — the 5% that succeed are using largely the same underlying technology as the 95% that don’t.

“Buying a mature vendor platform is always safer than building custom.” Not universally — MIT’s research found buying from a specialized vendor tends to outperform internal builds on average, but the deciding factor was still whether the tool was scoped to a real, measurable workflow, not the build-vs-buy decision alone.

“If the pilot has high adoption, it’s working.” Adoption and ROI are different measurements. A tool that 5,000 employees have access to but that isn’t tied to a measurable outcome is an adoption statistic, not a return.

“Agentic AI’s higher risk means it’s inherently a worse bet than generative AI.” Risk profile and ROI potential are separate questions. Properly scoped agentic deployments in automation-heavy functions are showing stronger productivity gains than search-only generative AI use, precisely because they complete work rather than only assisting with it.

Frequently Asked Questions

The Bottom Line

Companies evaluating agentic AI tools right now are increasingly doing so with real rigor — baseline capture, per-workflow ROI, multi-pillar outcome measurement, and governance built in from the start — but the data shows most organizations still aren’t there yet. The honest answer to “productive automation or wasted money” is that it’s currently both, split almost entirely along execution lines rather than technology lines: a small group of disciplined adopters is getting real, measurable results from largely the same tools that a much larger group is getting nothing from. The deciding factor isn’t which vendor or which model — it’s whether the workflow was narrow enough, measured carefully enough, and redesigned around what the agent can actually do.

Trying to figure out whether your next pilot is set up to be in the 5% or the 95%? Contact us or explore our agentic AI development services.