Kyndryl's Third Annual AI Readiness Report: 64% of Enterprise Leaders Report Positive ROI — but External Risk Readiness Has Fallen to 24%
A new survey of 2,000+ executives reveals the production-to-profit chasm at the heart of enterprise AI in late 2026: agents are running, but the measurement architecture to prove they're working isn't.
A Genpact survey of more than 2,000 executives, published November 12, 2026, surfaced the defining tension of enterprise AI in the fourth quarter: 79% of surveyed enterprises have AI agents running in production, and only 15% have fully tied their AI agent spending to profit-and-loss outcomes. The gap between those two numbers — 79% deployment versus 15% P&L accountability — is not a lag that closes automatically as deployment matures. It is a structural failure in how most enterprises have approached the measurement problem, and it is becoming more consequential as AI agent budgets grow and CFOs begin asking for the financial evidence that the deployment wave has produced.
The 79% figure is a genuine milestone. Twelve months ago, the dominant enterprise AI narrative was about moving from pilot to production. That transition has happened, broadly and quickly, across a majority of large organizations. The production question is largely answered. The accountability question — what are these agents actually worth, and how do we know — is the question that will define the next 12 months of enterprise AI leadership.
How 79% of Enterprises Got to Production Without a P&L Framework
The deployment-first, measurement-later pattern that explains the Genpact gap is not irrational — it is the natural output of how enterprise technology adoption has worked for two decades. Enterprise software is bought, integrated, and deployed before anyone has built the measurement infrastructure to evaluate it rigorously. The assumption is that the efficiency gains are obvious enough to justify the investment, and that formal ROI measurement is a tool for justifying budget renewals rather than informing deployment decisions.
AI agents entered the enterprise on this same adoption model, but with two features that make the measurement deficit more consequential. First, AI agents are expensive at scale — not in the way that a seat license is expensive, but in the way that compute-intensive, continuous-operation infrastructure is expensive, with costs that compound as usage grows. Second, AI agents are designed to be evaluated on outcomes rather than features: an agent is not valuable because it has the capability to complete a task, but because completing the task produces a business result. The business case for every significant AI agent deployment is built on an outcome claim. Without measurement infrastructure, that claim cannot be validated, and the CFO eventually asks why it cannot.
The enterprises that reached the 79% deployment figure without P&L measurement infrastructure did so because the deployment decision and the measurement design were handled by different functions — typically engineering or IT for deployment, and finance or operations for measurement — with no forcing mechanism to coordinate them before launch. The enterprise AI ROI measurement gap documented in Q4 2026 found that 83% of CFOs were increasing AI budgets while just 31% were satisfied with outcomes visibility. The Genpact data gives that dissatisfaction a sharper frame: the 85% without P&L accountability are the enterprises producing that CFO dissatisfaction at scale.
What "Tied to P&L" Actually Means and Why Only 15% Have Done It
P&L accountability for AI agents is not a reporting exercise — it is a measurement design challenge. The enterprises that have achieved it did not do so by applying post-hoc analysis to agent activity logs. They did so by designing the measurement before the deployment.
The practical distinction is between activity metrics and outcome metrics. Activity metrics track what the agent does: number of tasks completed, queries answered, documents processed, decisions routed, escalations handled. These metrics are valuable for monitoring agent health and reliability, but they do not answer the question a P&L-focused executive is asking: what financial outcome did the agent produce?
Outcome metrics connect agent activity to financial lines. A customer service agent that handles 10,000 queries per week produces an activity metric. A customer service agent that is associated with a 12% reduction in average handle time, a 4-point improvement in CSAT score, and a measurable reduction in repeat contact rate produces outcome metrics that can be translated into cost avoidance and revenue protection. The translation from activity to outcome requires a measurement design — a control group, a baseline measurement, a defined attribution window — that must be established before the agent is live, because retrofitting it after deployment is significantly harder and rarely produces results that satisfy a rigorous CFO challenge.
The 15% who have tied AI spend to P&L built this measurement architecture as a precondition to production launch. They treated it the same way they treat A/B test design for product features: define the hypothesis, establish the control, set the measurement period, then launch. The 85% who did not are now in the uncomfortable position of having expensive systems running in production with strong activity metrics and no defensible financial case.
The 20% Who Operate at Scale: What Broad Application Looks Like
The Genpact survey's third major finding — that only 20% of enterprises apply AI agents broadly and consistently across functions — is a separate dimension of the production story and a signal about the structural ceiling on enterprise AI value creation.
The 79% deployment figure counts any enterprise that has AI agents running in production in any context, including those with a single agent deployed in a single business unit. The 20% broad application figure counts enterprises that have moved beyond isolated deployment to something that resembles general-purpose enterprise AI infrastructure: agents deployed consistently across multiple functions, with shared governance frameworks, shared measurement standards, and shared learnings from deployment.
The gap between 79% and 20% is a function of organizational and architectural complexity rather than technical complexity. Scaling an AI agent from one business unit to multiple requires resolving questions that the initial deployment never had to answer: how are data access controls managed across functions with different data governance requirements? How are accountability and escalation paths defined in functions with different operating models? How are measurement frameworks standardized when the outcome metrics that matter vary by function? These are organizational design questions, and the enterprise AI activation crisis documented in earlier Signal analysis suggests that most enterprises have not yet invested in answering them systematically.
IBM and Microsoft's Platform Launch and What 6,000 Prebuilt Agents Signal
The IBM and Microsoft enterprise AI agent platform announcement on November 10, 2026 is the most significant infrastructure development for the enterprise deployment picture since the year began. The platform includes more than 6,000 prebuilt AI agents and assets developed through IBM Consulting, is designed for industry-specific transformation workflows, and is paired with forward-deployed IBM engineering support — consultants embedded within enterprise technology teams during implementation.
The 6,000-agent figure is not a catalog number — it is a strategic statement. It signals that IBM and Microsoft are treating AI agent deployment as an enterprise services market, not purely a technology product market. The forward-deployed engineering model is the key differentiator: it acknowledges that the barrier to production deployment for most Global 2000 companies is not access to capable AI agents but the organizational and technical capacity to implement them within existing enterprise systems at production quality.
For the AI agent analytics and measurement landscape, the IBM-Microsoft platform creates a forcing mechanism. When an enterprise deploys agents through IBM Consulting's forward-deployed model, the implementation engagement includes measurement framework design as a deliverable — it is part of the professional services contract, not an afterthought. This is one mechanism by which the 15% P&L accountability figure could plausibly rise: not because enterprises independently build measurement infrastructure, but because the platform vendors are building it into the deployment engagement as a standard component.
The platform also signals something about where the enterprise AI market is heading more broadly: toward the vertical software model, in which prebuilt agents are optimized for specific industry workflows rather than deployed as general-purpose infrastructure that individual enterprises customize from scratch. The industry-specific design of the IBM-Microsoft platform is a direct response to the finding that broad application is the barrier — agents built for a specific workflow in a specific industry context require less customization, integrate more cleanly with existing systems, and are easier to implement consistently across functions than horizontal agents designed for every use case.
The Agent Accountability Stack: Why Measurement Architecture Must Come Before Scale
The enterprises that have achieved P&L accountability for their AI agents share a common design pattern: they built the measurement architecture before they built the deployment. This is not a philosophical preference — it is a practical constraint. Retrofitting measurement infrastructure onto a production AI deployment is significantly harder than designing it in from the beginning because production deployments create dependencies that make the system harder to modify.
A properly designed agent accountability stack has four layers. The first is the data layer: the agent must emit the events and attributes necessary to connect its activity to the outcome metric. If the agent processes customer service inquiries but does not log the account identifiers or ticket types necessary to link agent interactions to downstream CSAT scores and repeat contact rates, the measurement is impossible without redesigning the data model. The second is the attribution layer: the measurement design that connects agent activity to outcome metrics through a comparison methodology — control groups, matched cohorts, time-series analysis — that is defensible under a statistical challenge. The third is the reporting layer: the dashboards and reports that translate the attribution data into the financial language the CFO and business unit leaders use. The fourth is the governance layer: the ownership model that assigns P&L accountability to named executives, creates incentives for maintaining measurement discipline, and provides escalation paths when agents produce unexpected outcomes.
Most enterprises that have deployed agents without P&L accountability have layer four (governance) in place as a loose organizational structure, layer three (reporting) in place as activity dashboards, and are missing layers one and two entirely. Building layers one and two on top of a live production deployment requires either a re-instrumentation project — modifying the production agent to emit the necessary data — or a measurement redesign that accepts noisier attribution. Neither is impossible, but both are expensive relative to designing them in from the start.
AI Agent Deployment Patterns and P&L Measurement Maturity
The following table maps the most common enterprise AI agent deployment patterns against their typical P&L measurement maturity level, based on the patterns documented in the Genpact survey data and corroborated by practitioner reporting from enterprise AI programs across industries.
| Deployment pattern | % of production deployments (est.) | P&L measurement maturity | Primary barrier |
|---|---|---|---|
| Single-function pilot extended to production | ~40% | Activity metrics only | No measurement design before deployment |
| Multi-function deployment, centrally governed | ~20% | Mixed: some outcome metrics, no P&L attribution | Attribution methodology not established |
| Vendor-deployed with professional services | ~15% | Partial P&L attribution | Measurement scope limited to vendor-defined metrics |
| In-house platform, enterprise-wide | ~15% | Full P&L attribution possible | Organizational readiness, not technical |
| Broad, consistent cross-function deployment | ~10% | P&L attribution standard | None — this is the 15% that has achieved accountability |
The table illustrates why the 15% P&L accountability figure is not primarily a technology problem. The enterprises with the highest measurement maturity are almost all in the in-house platform or broad cross-function deployment categories — the deployments where the organization has made sufficient investment in AI infrastructure that measurement architecture was included in the investment thesis. The 85% without P&L accountability are concentrated in the first two categories: deployments that reached production via the path of least organizational resistance, without the upfront measurement investment that P&L accountability requires.
Five Steps to Close the Production-to-Profit Gap in 90 Days
The transition from production deployment to P&L accountability does not require a multi-year measurement transformation program. The enterprises that have done it most effectively built their accountability framework incrementally, starting with one agent and one metric and scaling the methodology. The following five-step sequence is designed to produce a defensible first P&L data point within 90 days.
1. Audit all running agents against outcome hypotheses. For every AI agent currently in production, document the outcome hypothesis that justified the deployment — what financial result was it expected to produce? For many agents, this hypothesis was never written down explicitly. Write it down now. If no one can articulate a financial outcome hypothesis for a live agent, that agent's budget is at risk in the next planning cycle, and the audit makes that visible before the CFO does.
2. Set a measurement baseline for the highest-priority agent. Select the agent with the clearest outcome hypothesis — typically the one that is closest to a revenue or cost line with available data — and measure the baseline financial metric for the 60 days before the measurement sprint begins. If the agent is a customer service agent, measure average handle time, CSAT, and repeat contact rate. If it is a sales prospecting agent, measure pipeline creation rate and time-to-qualified-lead. The baseline must be measured before the sprint, not estimated retroactively.
3. Assign an owner to each agent's outcome accountability. P&L accountability does not exist without named ownership. Assign a business unit leader — not an AI engineer or IT director — to own the outcome metric for each major agent deployment. This person is accountable for the financial result, not the technical performance. The measurement design is the AI team's responsibility; the outcome accountability is the business leader's.
4. Build a 90-day attribution cycle with a comparison methodology. Design a simple comparison for the highest-priority agent: a geographic region, a customer segment, or an employee cohort that is not yet using the agent. Measure the outcome metric for both groups over 90 days. The delta between treatment and control is the attribution starting point. It will be imprecise — 90 days is a short window and confounding factors are real — but it will be more defensible than a before-and-after comparison with no control, and it creates the measurement template for subsequent agents.
5. Publish a monthly agent P&L report to the CFO and business unit leaders. A single data point is a curiosity. A monthly report with consistent methodology is a governance artifact that changes how the enterprise thinks about AI agent investment. The Menlo Ventures 2026 enterprise AI report found that enterprises with formal AI ROI measurement processes are significantly more likely to increase AI budgets than those without — not because their ROI is higher, but because the measurement itself creates the organizational confidence to commit capital. The report is the accountability mechanism. Build it before you need it.
The five-step sequence applies to every enterprise in the 85% that has not yet achieved P&L accountability. It does not require a measurement transformation program, a new analytics platform, or a team of data scientists. It requires a measurement design decision made before the next agent goes into production, and a willingness to apply that design retroactively to the most important agent already running. The AI agent accountability gap documented in AI agents news tracking from November 2026 is not a gap that closes by waiting for vendors to solve it. It closes when enterprises decide to treat measurement architecture as a first-class deployment requirement rather than a reporting afterthought.
Takeaway: The Genpact survey's 79%/15% split is the most precise diagnosis yet of where enterprise AI actually stands in late 2026: past the deployment question, stuck on the accountability question. Getting 79% of enterprises to production was the technology industry's achievement. Getting from 15% to something approaching majority P&L accountability is an organizational achievement — a function of measurement design, change management, and business leadership commitment that no vendor platform can substitute for. The IBM-Microsoft platform announcement on November 10 suggests that the platform vendors understand this and are moving toward forward-deployed professional services as the delivery model for the accountability gap. That is the right direction. But the enterprises that close the gap fastest will be the ones that treat measurement architecture as a deployment prerequisite — not a reporting feature — and build it in from the beginning on every new agent launch.
Frequently Asked Questions
What did the Genpact 2026 AI agent survey find?
Genpact's survey, published November 12, 2026, polled more than 2,000 executives across enterprise organizations to assess the state of AI agent deployment and measurement maturity. The two headline findings are in direct tension with each other: 79% of surveyed enterprises report having AI agents running in production, a figure that confirms agentic AI has crossed the deployment chasm from early adoption to mainstream enterprise deployment. Simultaneously, only 15% of those enterprises have fully tied their AI agent spending to profit-and-loss outcomes — meaning the vast majority of enterprises with live AI agents cannot demonstrate, through their standard financial accounting, what those agents are actually worth to the business. A third finding adds context: only 20% of businesses apply AI agents broadly and consistently across functions, suggesting that even the enterprises that have reached production have done so unevenly, with agents concentrated in specific use cases or business units rather than operating as general-purpose enterprise infrastructure.
Why do enterprises have AI agents in production but can't connect them to profits?
The disconnect between production deployment and P&L accountability is a measurement architecture failure, not a performance failure. Most enterprise AI agent deployments were launched under a pilot-to-production model in which the primary success criteria were technical: could the agent complete its assigned tasks reliably, with acceptable error rates, and within the latency and cost budgets set for the pilot? Those criteria are legitimate for a technical deployment decision, but they are insufficient for a business accountability decision. An agent that reliably completes its assigned tasks is a successful technical deployment. An agent whose task completion can be traced to measurable business outcomes — revenue influenced, costs avoided, time-to-close reduced, customer satisfaction improved — is a P&L-accountable deployment. Building the second type of accountability requires measurement infrastructure that must be designed before deployment, not retrofitted after. The 85% of enterprises that cannot tie AI spend to P&L deployed agents using the first framework and are now trying to build the second framework after the fact, which is significantly harder.
What does it mean to tie AI agent spend to P&L outcomes?
Tying AI agent spend to P&L outcomes means establishing a causal or at minimum statistically defensible connection between the investment in a specific AI agent and a change in a revenue, cost, or margin line item on the enterprise's financial statements. This is different from activity reporting, which tracks what the agent does — number of tasks completed, queries answered, documents processed, decisions made — without connecting those activities to financial outcomes. P&L accountability requires a measurement design that specifies, before deployment, what financial outcome the agent is expected to influence, establishes a baseline for that outcome measured before the agent is live, defines a measurement period after deployment, and uses a comparison methodology — a control group, a matched cohort, or a time-series analysis — that makes the agent's contribution distinguishable from other factors affecting the same metric. The 15% of enterprises that have done this did not achieve P&L accountability by accident. They designed it into the deployment from the beginning, treating measurement architecture as a prerequisite to production launch rather than a post-deployment reporting exercise.
How does the IBM and Microsoft enterprise AI agent platform change the production deployment picture?
IBM and Microsoft's joint enterprise AI agent platform, announced November 10, 2026, is the most significant infrastructure development for enterprise AI agent deployment since AWS launched Bedrock AgentCore earlier in the year. The platform includes more than 6,000 prebuilt AI agents and assets from IBM Consulting, is designed for industry-specific transformation, and comes with forward-deployed engineering support — meaning IBM consulting teams help enterprises implement the platform within their existing technology stacks. The forward-deployed engineering model is the most important element for the production deployment picture because it addresses the deployment gap that has historically separated enterprises with the in-house AI engineering capacity to build production-grade agents from those without. By pairing a prebuilt agent library with on-site implementation support, the platform substantially lowers the deployment barrier for the Global 2000 companies that have IT organizations but not dedicated AI engineering teams. The implication for the Genpact survey finding is double-edged: the platform makes it easier to deploy agents into production, but it does not automatically solve the P&L measurement problem. Getting to 79% deployment was not the hard part. Getting to 15% P&L accountability — and raising that number — requires measurement architecture that the platform does not provide out of the box.
What is the fastest way to build a P&L attribution framework for AI agents?
The fastest path to P&L attribution is a 90-day measurement sprint designed around a single agent and a single outcome metric. Start by selecting the agent that is closest to a measurable revenue or cost line — a sales prospecting agent that influences pipeline value, a customer service agent that affects handle time and CSAT, a procurement agent that influences supplier pricing. Define one financial outcome metric for that agent, establish a baseline for the metric measured over the 60 days before the sprint begins, and design a simple comparison — either a geographic control group running without the agent or a matched cohort of accounts or employees not yet using the agent. At the end of 90 days, compare the outcome metric between the treatment and control groups. The delta, adjusted for confounding factors, is your first P&L attribution data point. It will not be perfect, and it will require refinement, but it will be defensible enough to present to a CFO, and it creates the measurement template that can be applied to every subsequent agent deployment. The enterprises that have reached 15% P&L accountability almost all did it this way: one agent, one metric, one 90-day sprint, then scaled the methodology.