> ## Content Index
> Fetch the complete content index at: https://unhyd.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# AI ROI Looks Strong. Explainability Doesn't. Here's What Three Studies Reveal
- URL: https://unhyd.com/article/ai-roi-accountability-explainability-gap/
- Published: 2026-10-14T02:52:37.000Z
- Updated: 2026-10-14T02:52:37.000Z
- Description: New research from FICO, Maryland and Workday points to a gap between using AI, getting value from it and governing the decisions it makes.
- Author: Jonas Muthoni
- Tags: Business, #home-spotlight

Businesses are increasingly comfortable claiming strong AI ROI, or return on investment. They are much less comfortable explaining exactly how artificial intelligence makes consequential decisions. That contrast is the most revealing finding in a new round of enterprise AI research, and it exposes a weakness in how the technology's success is often measured.

In a survey released October 7, [FICO reported](https://fico.gcs-web.com/news-releases/news-release-details/only-52-tech-leaders-have-full-confidence-explaining-their-ai?ref=unhyd.com) that 85.1% of respondents said AI had met or exceeded their return-on-investment expectations. Yet only 5.2% were very confident they could explain AI-driven decisions to regulators and customers. The figures come from different questions, not from a matched assessment of which organizations both made money and lacked explainability. They nevertheless reveal a crucial distinction: a company can be satisfied with a project's return while still lacking confidence in its accountability.

Unhyd compared that global survey with Maryland's October business AI benchmark and Workday's new workforce research. These are three different datasets with different samples and methods, so their percentages cannot be pooled into one national AI score. Read side by side, however, they suggest a more useful way to evaluate enterprise AI: separate *adoption*, *benefit*, *control* and *workforce readiness* instead of treating a successful pilot as proof that the organization is ready to scale.

## Why AI ROI and AI accountability are not the same measurement

FICO's 2026 State of Responsible AI research, conducted with Corinium Global Intelligence, surveyed 1,004 senior technology, data and risk leaders across seven industries. About two-thirds of respondents were based in North America. The company says 71.5% of customer-impacting decisions in respondents' organizations are still made without AI, while just 8.2% of organizations have a single shared AI deployment platform fully in place. Another 71.2% say such a platform is only partially in place.

These results do not mean that 71.5% of decisions ought to be automated. Some decisions are better left to people or transparent rules. Nor does a shared platform guarantee a sound decision. What the results show is that favorable ROI perceptions can coexist with fragmented infrastructure and uncertainty about why a model produced a particular result.

The definition of ROI matters. FICO asked leaders whether AI met or exceeded their expectations; the public release does not establish that an independent auditor verified every respondent's cost savings or revenue gains. An organization might have set modest targets, achieved genuine savings on a narrow task and still be unable to defend the system's output in a disputed customer case. That is a successful pilot and an incomplete governance story at the same time.

## Maryland's benchmark shows how broad adoption can hide shallow use

Maryland's [October 1 benchmark](https://governor.maryland.gov/news/press-releases/governor-moore-convenes-maryland-innovation-summit-releases-new-maryland-business-ai-benchmark?ref=unhyd.com) offers a different lens. Among nearly 300 senior decision-makers surveyed in the state between June and July 2026, 91% reported some AI use. But 58% described basic use, such as standalone tools or features embedded in existing software, rather than deeper integration. Among regular users, 92% reported a positive productivity effect; 70% characterized the gain as slight and 21% as significant.

Those are self-reported experiences, not a measured experiment showing that 91% of Maryland businesses have transformed their operations. The sample is also different from the business populations measured by national statistical surveys. Its value is narrower and more practical: it separates having access to AI from reorganizing work around it.

Imagine two businesses that both report using AI. One asks employees to summarize meetings and draft emails. The other has redesigned an invoice process, checks outputs against source records, logs exceptions and measures the time staff spend correcting errors. A simple adoption question counts both. A serious operating review does not.

The Maryland figures also resist a popular shortcut in the jobs debate. The state reported that 64% of respondents expected existing employees to do more with AI rather than add staff, while 4% expected to reduce headcount. Those are plans and expectations, not observed layoffs or guarantees that employment will be preserved. They do suggest that the first challenge for many organizations is changing work, not eliminating it.

## Workday's data points to the missing workforce layer

The workforce evidence adds another distinction. In its [October 5 report](https://newsroom.workday.com/2026-10-05-Workday-Global-Workforce-Report-AI-Is-Rewriting-Jobs-More-Than-Its-Cutting-Them?ref=unhyd.com), Workday said demand for basic AI skills in job requisitions from roughly 550 enterprise employers using Workday Recruiting fell 25% after peaking in January 2026\. Demand for hands-on skills associated with building AI tools, automation and AI engineering rose 51% between September 2025 and July 2026.

The percentages have different comparison windows, so they should not be treated as a single net change. Nor do requisitions from Workday customers represent every employer. But the direction is consistent with a shift from using a chatbot to integrating AI into actual processes. Unhyd's earlier [analysis of Workday's skills data](https://unhyd.com/article/workday-ai-skills-demand-beyond-prompting/) examined that distinction in detail.

Workday's separate survey found that 79% of workers said they knew what skills they needed to succeed, while 66% said their employer helped them develop those skills. That is a 13-percentage-point difference within the reported survey results, not proof that training alone will close it. It does raise a straightforward management question: who is responsible for teaching staff to supervise and improve the workflows companies now want to automate?

Governance cannot be outsourced to a technical team after deployment. Unhyd previously reported [Gartner's finding on HR and IT alignment](https://unhyd.com/article/gartner-hr-it-ai-alignment-gap/), which highlights the difficulty of agreeing on how AI will change jobs. A model can pass a technical test while the organization has no clear owner for training, appeals, process redesign or error correction.

## A more useful AI scorecard: four questions instead of one

These studies do not prove that weak explainability causes low productivity or that better training guarantees higher ROI. They do provide a basis for an editorial framework that leaders can use to interrogate their own claims. Rather than asking only whether an AI tool saved time, Unhyd proposes four separate questions.

### 1\. Adoption: where is the tool actually used?

Record the tasks, teams and frequency of use, not just the number of licenses or employees who tried a chatbot. Distinguish optional assistance from a system that changes a customer's outcome. This is a coverage measure, not a value measure.

### 2\. Benefit: what changed after all costs?

Compare a defined process before and after deployment, ideally with a suitable comparison group. Count software costs, integration, review, rework and training alongside saved time or additional revenue. A faster first draft is not necessarily a faster finished decision if employees must spend the savings checking it.

### 3\. Control: can someone reconstruct a decision?

For consequential use cases, identify the inputs, system version, rules, human approvals and escalation route. Ask whether the organization can explain and challenge an output without relying on a vague claim that the model is intelligent. The appropriate level of documentation will depend on the risk of the use case.

### 4\. Readiness: can people operate the new process?

Measure whether staff know when to trust an output, when to escalate, how to spot unsupported claims and who owns a failure. Track whether the process survives ordinary staff turnover. A workflow that depends on one expert's undocumented prompt is not a durable operating capability.

This is an analytical checklist, not a validated industry index or a new statistical finding. Its purpose is to prevent four different kinds of evidence from being collapsed into a single success story.

![Infographic showing four dimensions of enterprise AI performance: adoption, business benefits, explainability and governance controls, and organizational readiness.](https://storage.ghost.io/c/df/e4/dfe4d080-643c-453a-8a60-0452bdbc0efd/content/images/2026/10/AI-performance--adoption--business-benefits--explainability-and-governance-controls.jpg)

Unhyd's four-part AI scorecard examines adoption, measurable benefits, governance controls, and organizational readiness. The framework offers a broader approach to evaluating AI ROI and accountability, but is not an empirically validated measurement scale. Graphic: Unhyd.

## What a credible AI success story would disclose

For a customer-service deployment, useful reporting might include the share of requests handled, resolution accuracy, the number of escalations, customer complaints, cost per resolved case and whether people can appeal a wrong answer. For an internal research assistant, it might include source accuracy, time to a verified answer and the rate at which reviewers catch errors. The numbers should be measured on the same task definition over time.

Organizations should also be explicit about what remains outside AI's authority. A system that declines to automate a sensitive decision may be performing exactly as designed. High adoption is not an ethical objective in itself, and more automation is not automatically better performance.

The practical conclusion from this October research is not that AI is failing. FICO's respondents report favorable returns, Maryland's respondents report widespread modest gains, and Workday's customer data show demand for more sophisticated capabilities. The conclusion is that **AI ROI** answers only one question. Until organizations can demonstrate the quality, accountability and sustainability of the work behind those returns, the most important part of the business case remains unfinished.