For administrators and development teams, this changes the question that should accompany a production launch. “Is it running?” remains necessary, but it cannot answer whether the system has reliable information, whether its work reaches completion, or whether the organization can account for its cost and consequences.
Three first-party research announcements—from The Modern Data Company, Talkdesk, and Flexera—provide substantive detail behind that argument. They examine different populations and different parts of the technology stack. Read together, they support a practical distinction: deployment is a milestone; operational readiness requires separate evidence.
Enterprise AI production figures leave the value question open
The broad adoption figures in Grit Daily come from reporting on Plug and Play’s 2026 Enterprise AI Strategy Pulse Survey. According to that coverage, 74% of large enterprises have at least one AI solution in production, while 93% are piloting AI or further along. A separate account by Aiven Consulting Group describes the survey as focused on Fortune 500 and Forbes Global 2000 companies.
That population boundary matters. These findings concern large enterprises, not every employer, small business, or IT department. Having one production solution also establishes a relatively modest threshold: it does not establish that AI spans the organization or completes a business process without substantial human assistance.
The Plug and Play figures reproduced by Grit Daily identify data foundations as a scaling obstacle for 71% of respondents, governance friction for 53%, and legacy-system integration for 26%. They also describe deployments at different levels of breadth, from one business function through several functions to organizations identifying as AI-native. The original survey tables are not available in the inspected evidence, so those category percentages should not be reconstructed into a precise maturity distribution.
The measurement finding deserves similar care. Grit Daily reports that half of respondents were either too early to measure return on investment or were not consistently tracking AI performance. Aiven’s account describes the measurement gap among enterprises running production AI. Those formulations are not identical enough to establish the exact denominator without the primary report, but both describe a substantial unresolved measurement problem.
Being too early to measure an investment is different from having no consistent measurement process. The first may be a temporary condition; the second can persist after deployment. Neither establishes that the investment has produced no value.
Aiven goes further, attributing the problem to missing pre-deployment baselines and arguing that value then becomes permanently unmeasurable. That causal explanation is not established by the primary material available here. The useful management lesson is narrower: recording the process’s starting condition before changing it makes later comparisons more defensible. It does not justify declaring that every deployment without such a baseline can never be evaluated.
The Modern Data survey puts trust behind agent deployment
The Modern Data Company’s August 20 announcement offers a more detailed view of the underlying data problem. Its findings are explicitly interim: the third annual Modern Data Survey was still underway, with more than 540 qualified responses from data leaders and practitioners across 66 countries.
Among those respondents, 57.3% reported piloting or running AI agents in data and analytics workflows. That comprised 23.5% with agents in production and 33.8% conducting pilots. Only 8.4% said the data feeding their AI systems was trustworthy enough for production.
These figures describe different questions within the same research, not a measured failure rate. It would be incorrect to subtract the readiness figure from the deployment figure and claim the difference represents unsafe installations. The more directly relevant result is the production subgroup: even among respondents already running agents in production, only 21.7% were very confident that their data was trustworthy enough for production.
The company also reports that data quality and trust ranked among the top three production barriers for 75.9% of respondents. Missing context and lineage followed at 63.5%, and security concerns at 61.7%. Skills gaps and immature tooling ranked lower, at 25.9% and 19.5%, respectively. Within this sample, respondents were pointing more strongly to the information and controls beneath agents than to a simple lack of tools.
AI data readiness includes meaning, ownership, and policy
The research defines business context as definitions, relationships, lineage, and policy. Lineage is the record of where data came from; business definitions explain what the information means. These are distinct from whether an application can technically retrieve a record.
That distinction explains why making more data accessible may leave an agent’s operational problem unresolved. Access establishes that information can be obtained. It does not, by itself, establish that the information has the right meaning for the decision, that its origin can be traced, or that the applicable policy permits its use.
The Modern Data Company reports that 60.9% of respondents consider a reliable context layer necessary for AI agents, but only 16% deliberately design and engineer that layer as a product. One in four reported no formal context layer. Here, “context layer” means the organized business information that helps systems interpret data and apply it appropriately—not merely additional text placed in a prompt.
Respondents with production agents were more likely to have engineered context layers. However, the company explicitly cautions that its snapshot shows correlation rather than causation. The research cannot establish whether stronger foundations enabled agent deployment or whether deploying agents prompted organizations to improve those foundations.
That limitation also rules out a purchasing shortcut. The findings support examining definitions, relationships, lineage, and policy; they do not prove that buying a particular vendor’s context-layer product will produce a particular business return.
AI actions make traceability a production requirement
The operational concern becomes sharper when an agent can move from interpreting information to acting on it. The Modern Data Company describes agents making determinations across enterprise systems and workflows that previously remained with people. An error at that point can affect the work being performed, rather than stopping at an inaccurate answer.
Yet the survey reports a sizable gap between governance expectations and supporting records. While 65.1% said AI-enabled decisions must be explainable, traceable, and defensible, only 39% maintained either an audit trail for AI inputs and outputs or a link from decisions back to data sources. Just 10% maintained both.
Those two records answer different questions. An input-and-output trail helps show what information entered the system and what it produced. A decision-to-source link helps explain which underlying information supported the decision. Keeping one does not automatically provide the other.
Accountability is another unfinished part of the picture: only 17.7% reported a clear, documented AI accountability framework. For an IT team considering broader agent permissions, the practical implication is to establish who answers for the outcome alongside what the agent is allowed to do. Expanding capability while leaving responsibility unresolved would carry the reported governance gap into a larger workflow.
Talkdesk finds customer-service AI stopping at departmental boundaries
Talkdesk’s August 25 report examines a narrower but operationally revealing setting: customer experience, or CX. NewtonX conducted the commissioned survey in April 2026, gathering responses from 252 director-level-and-above decision-makers and significant influencers responsible for CX, IT, operations, or AI strategy. Respondents worked at mid-market and enterprise organizations across multiple regions and industries.
Talkdesk reports that 98% had deployed AI somewhere in the customer journey. Only 15%, however, combined agentic AI with cross-departmental orchestration to resolve customer needs from beginning to end. In the report’s terms, orchestration means coordinating AI agents, human teams, data, and workflows across enterprise systems.
Those findings measure substantially different accomplishments. Introducing AI at one point in a customer journey is a deployment decision. Completing a request across departmental boundaries requires the participating systems and people to carry the work forward together.
The more specific context figures help explain the gap. Talkdesk reports that 64% used specialized AI agents, but only 35% retained customer context as work moved between systems. Disconnected systems were a technical barrier for 45%, while legacy infrastructure was a barrier for 44%.
This gives enterprise teams a concrete way to evaluate an automation proposal. The relevant boundary is not simply the agent’s own task. It is the transition to the next system or team: what information travels with the request, whether the next participant can continue the work, and whether someone must reconstruct the context.
Human integration work belongs in the AI business case
Talkdesk reports that human agents spent an average of 28% of their time switching systems, re-entering information, and searching for customer context. This is a survey finding, not an independently observed time-and-motion measurement by WindowsForum. It nevertheless identifies a form of work that a narrow AI usage dashboard could miss.
An organization could increase the number of AI-assisted interactions while leaving that manual coordination burden largely untouched. The research does not prove that this happens in every deployment. It does show why counting interactions or specialized agents cannot, on its own, establish an improvement in end-to-end service.
Only 5% of respondents could quantify AI’s impact on business outcomes, according to Talkdesk. That number should remain attached to this customer-experience sample and its survey question. It is not interchangeable with the Plug and Play measurement finding or Flexera’s separate measure of AI software value.
Talkdesk also reports that organizations combining agentic AI and cross-departmental orchestration were four times more likely to report major customer-satisfaction or Net Promoter Score gains. That is an association in commissioned survey research, not experimental proof that orchestration alone caused the improvement. Talkdesk sells customer-experience automation, so its commercial interest belongs beside its interpretation of the results.
The useful conclusion does not require accepting a vendor’s entire maturity framework. For customer-facing AI, evaluate the completed request and the remaining human workload, rather than treating each automated interaction as a finished business outcome.
Flexera shows AI spending becoming visible before AI assets do
Flexera’s June 24 announcement of its 2026 State of ITAM Report examines the management side of the same problem. ITAM means IT asset management: accounting for the technology an organization owns or uses and managing its associated obligations and costs. The report draws on 512 technology professionals worldwide.
Nearly half of organizations were already tracking AI as part of software spending, according to Flexera. Only 31% reported accurate visibility into AI software. Separately, 59% of respondents said wasted AI spending had increased year over year.
The wording of that last result is important. It does not mean that 59% of AI spending was wasted. It means 59% of respondents reported an increase in spending they considered wasted; the announcement does not supply an aggregate dollar amount or waste percentage.
The difference between expenditure tracking and asset visibility is the practical issue. A spending record can establish that money went to an AI-related service without providing a complete account of the software in use. Flexera describes AI as crossing existing categories—cloud, software as a service, data centers, and devices—while adding models, agents, data, and platforms to the management picture.
This is also broader than AI. Flexera reports that complete visibility across the IT estate had fallen to 36%. The AI findings sit inside an existing asset-management challenge, rather than establishing that AI alone caused the visibility decline.
AI inventory must connect spending to operational responsibility
For enterprise administrators, the decision is whether the organization’s AI records connect financial visibility with the systems actually being operated. Finance can ask what was purchased. The operating team needs to know what is running and who owns it. Governance needs enough information to connect that system with the data and actions under its control.
Connecting those questions is an editorial recommendation drawn from the surveys, not a tested remediation program. Flexera establishes the visibility problem; The Modern Data Company establishes gaps in data traceability and accountability. Together, they make a stronger case for linking asset records with operational ownership than either finding provides alone.
They do not establish that a single management platform will solve the problem. Nor do they support attributing these gaps to a particular Microsoft or third-party product. The evidence concerns organizational management capabilities across AI deployments, without identifying a defective version, configuration, or service.
That boundary keeps the purchasing decision grounded. Before adding another management tool, an organization should identify which question it cannot currently answer: what AI it uses, what it costs, what information it accesses, what work it performs, or who is responsible for the result. A new tool should be evaluated against that specific gap.
Enterprise IT should make expansion conditional on evidence
Teams planning to expand an AI deployment should review the workflow’s data, handoffs, accountability, and measurement before approving broader scope. The surveys support that decision more directly than they support either a blanket pause on AI or a faster rollout.
A bounded production system may already have the necessary controls. Conversely, a widely used system may remain difficult to evaluate. The useful review unit is therefore the particular workflow, with its actual information sources and operating responsibilities, rather than the organization’s overall claim to be “AI-ready.”
The following checks translate the reported weaknesses into a practical review. They are questions for planning and governance, not product-specific configuration instructions or a survey-validated checklist.
- Record which AI applications and agents support the workflow, who owns them, and how their spending is accounted for, so that a budget total does not stand in for an operational inventory.
- Identify the data sources, business definitions, lineage, and applicable policies the workflow depends on, and document the remaining trust gaps before expanding its use.
- Establish what information the AI can use, what actions it can take, and who is accountable for the outcome, keeping those responsibilities attached to the specific deployment.
- Follow the work across system and departmental handoffs, checking whether context survives and where people must re-enter information or reconstruct the request.
- Define the business outcome and its starting condition where available, then assess completed work and remaining manual effort alongside usage and cost.
Use workflow evidence to decide what can expand
A favorable expansion case would answer those questions with records rather than general assurances. The team would be able to identify the information supporting the workflow, explain responsibility for its actions, and show how the work reaches completion. It would also have a measurement approach appropriate to the deployment’s age and scope.
Where an answer is missing, the next investment should address that missing capability. A data-trust problem calls for work on the data and its context. Lost information at handoffs calls for integration work. An incomplete inventory calls for asset visibility. None of those findings, by itself, establishes a need for a more capable model.
Measurement should likewise preserve distinctions. A system that is too new to have produced a reliable financial result needs a defined evaluation point. A system with no consistent performance tracking needs a measurement process. Treating both as “ROI not yet available” risks leaving the second problem unattended.
These surveys do not supply universal rollback instructions, permission settings, or compatibility guidance. Those depend on the actual software and deployment. Their contribution is to identify the operational evidence worth requiring before a team makes a system more widely used or more consequential.
The AI readiness gap is a management finding, not a global failure rate
The studies align in direction, but their percentages cannot be pooled into one enterprise AI readiness score. The Modern Data Company surveyed data leaders and practitioners through an ongoing community research initiative. Talkdesk’s research concerns senior customer-experience and technology decision-makers. Flexera’s report examines technology asset management. Their samples, questions, and definitions differ.
The timing also deserves attention. Talkdesk published in August but fielded its survey in April. The Modern Data Company’s August results were interim rather than the completed annual study. These are useful observations about 2026 enterprise practice, not synchronized measurements of the entire market on September 22.
Vendor involvement does not erase the findings, but it limits how far their implications can be taken. The Modern Data Company has an interest in data foundations, Talkdesk in orchestration, and Flexera in technology visibility and spending control. Their announcements provide direct evidence of what their research reports; they do not independently demonstrate that their products resolve the reported problems.
Grit Daily also draws a cybersecurity parallel, reporting that an Omdia study found 93% of surveyed leaders considered absolute immutability a critical backup-storage requirement, while only 16% said their environments met it. That underlying Omdia record is not available in the inspected evidence, so the figures remain attributed to Grit Daily, rather than independently corroborated here. At most, the comparison illustrates the same conceptual distinction between recognizing a requirement and implementing it; it does not strengthen or quantify the AI findings.
The defensible conclusion is practical rather than dramatic. Enterprise AI programs should make their next expansion decision on evidence of trusted data, accountable actions, connected workflows, visible costs, and measurable outcomes. The organizations able to produce that evidence will have a firmer basis for deciding where another agent helps—and where the next useful investment is in the systems and processes already beneath it.