AI Leader: How to Drive Measurable ROI

Business professional typing on keyboard as AI finance automation dashboards show revenue forecasting, transaction processing, risk analysis and payment gateway control for efficient operations.

Almost every large enterprise now uses AI. Almost none of them can prove it made money.

That is not rhetorical or anecdotal. Nearly nine in ten organizations report using AI in at least one business function [1]. In McKinsey's 2026 global survey of 1,719 executives across 97 nations, 44% of organizations reported scaling AI across the enterprise — up from 38% in 2025. The share attributing any EBIT impact to AI stayed flat at 37%, and the share qualifying as "AI high performers" stayed flat at 6% [1]. McKinsey defines an AI high performer as an organization attributing at least 5% of EBIT to AI [1].

The other interesting aspect of this study is that eighty percent of respondents said AI improved their own productivity. That individual gain did not aggregate into enterprise value.

Summary: AI Adoption moved. AI Value did NOT.

1. Executive Summary

These are the questions asked by the CxOs in Enterprises in 2026:

CEO / BOARD — "What is AI actually doing for us?"

CFO — "I get an aggregate cloud bill that grows but cannot decompose it. How can I get to unit economics including cost per workflow run, per outcome, per avoided FTE-hour?"

COO — "Is there a way to break the link between revenue growth and headcount growth without breaking the customer relationship?"

CIO / CTO — "Can I demonstrate to the CFO and CEO the rate at which the cost of building AI is falling inside the organization while we are shipping more faster?"

Gartner attributes agentic AI project cancellations to three causes [2]: escalating costs, unclear business value, and inadequate risk controls. Model capability is not on the list. Gartner also predicts over 40% will be cancelled by the end of 2027.

Data is another big reason for cancelling AI projects. Gartner predicts organizations will abandon 60% of AI projects unsupported by AI-ready data through 2026 and found 63% of organizations either lack or are unsure of the data management practices AI requires [3].

"Agent Washing" is prevalent among vendors and internally in an enterprise. Gartner has estimated that of the thousands of vendors marketing agentic AI, only around 130 offer something genuinely agentic. Others are just rebranding chatbots, assistants, and RPA as Agentic AI [2]. The same instinct operates inside an enterprise. A model gateway plus a vector database plus a Slack channel will not help move the enterprise to be AI-native. It is three tools and an intention.

2. What Should Enterprises Do to Increase AI Value?

These nine focus areas operationalize the outcomes and map directly to the required AI ROI dimensions:

Strategy and Alignment: Connect the AI strategy, roadmap, and balanced portfolio to business goals and financial targets. Align enterprise leadership. Continuously update as needed.

Leadership & Accountability: Assign executive ownership, decision rights, escalation paths, risk management, and benefits accountability.

Digital Core Platform for AI + Data that's future-proof: Provide AI-ready data, secure platforms, orchestration, ITSM, MLOps/LLMOps, and cost controls. Create Forward-Deployed Engineers teams partnering with business domain and product experts for AI-production-ready solutions.

Responsible AI, Ethics & Governance: Embed fairness, safety, privacy, explainability, and compliance across the AI lifecycle including the AI control spine. Communicate and adapt Responsible AI policies.

Use Case Portfolio Prioritization & Delivery: Prioritize the portfolio by strategic fit, feasibility, risk, and value; enforce idea-intake selection and application kill criteria.

Maximized AI Value: Redesign workflows around baselines, customer impact, reuse, time to value, and ROI. Increased workforce productivity value is a beginning. Workflow automation, augmentation, and business transformation are the key to continuous value realization and significant ROI.

AI Product or Platform Operating Model & Operations: Fund persistent cross-functional product teams with clear owners and human oversight. Organize the workforce within products and platforms that will accelerate the creation of an interconnected enterprise, eliminating organizational friction while increasing speed to value.

Workforce and Culture: Build role-based AI fluency, human-machine practices, and talent pathways that create an AI-fluent organization and culture.

Change Management: Drive transformation and adoption through dynamic change plans, change leadership, continuous learning, and quarterly reassessment.

3. What Should a Future-Proofed Digital Core Platform for AI + Data Look Like?

It should have two layers and a control spine:

Layer 1 — Data, Context & Integration Layer

The focus of this layer is to clearly identify what agents can know and touch, agent decision quality, data quality, and integration abstraction. This layer contains AI-ready data products, a semantic layer, metric definitions, entitlement-aware access, quality, lineage, memory, and an action fabric including MCP.

Layer 2 — Execution Layer

The focus of this layer is to get the work done, tracking model changes and regression, as well as controlling costs. This layer consists of orchestration & runtime, multi-provider model routing, tool registry, scoped autonomy, evals, per-workflow cost, and human-in-the-loop review & override.

Control Spine

Not a third layer but runs alongside the other two layers because evidence needs to be captured from day one. The focus of the control spine is to make sure we have proof of the work done and build an audit trail as part of the platform. This spine consists of AI and data governance, agent / non-human identity, decision logs, evidence packs, bias, drift and eval testing, and adversarial defense.

4. What Are the Major Accountability Roles and Responsibilities in an AI-Native Enterprise?

Organizations with a clear accountable function for responsible AI average a maturity score of 2.6, against 1.8 for those without — a wider gap than between industries or between regions [5]. Naming these owners is an org-chart decision, and it is free.

AI Governance Council — a cross-functional group from Data, Security, Legal, Risk, and business owners who set policy, approve autonomy tiers for AI agents, and own the exception process. This is a strategic group of executive leaders with accountability to make decisions and have budgets.

Platform Owner — accountable for the two layers and the control spine of the AI + Data platform, roadmap, SLAs, and cost model. This is a product role, not an infrastructure role.

Domain Data Owners — one person per domain based on how the enterprise domains are structured, accountable for data KPIs, quality, definitions, and entitlements of what their domains publish to the semantic layer.

Workflow / Application Product Owners — accountable for the outcome, the pre-AI baseline, the eval suite, and the override policy for the workflow.

5. What Is a Good AI Platform / Product Operating Model?

The operating model must move at the speed of AI. In practice that means a workflow reaches production in days and a new MVP in about three weeks — but those are outputs of a structure, not targets you can mandate. The structure that produces them is a product operating model rather than a project one.

Fund products, not projects. A project has a scope, a go-live, and an end date. An AI workflow has none of those: it needs continuous evaluation, drift monitoring, prompt and model refresh, and periodic re-baselining as the domain moves. A project-funded initiative orphans its workflow at go-live, which is precisely when the maintenance burden starts. Persistent teams funded against outcomes, with quarterly reprioritization and an agreed kill criterion, keep the workflow owned for as long as it runs. Fund review, override, and escalation capacity as an explicit line item in that team's budget — unfunded review is the most common reason adoption stalls after a successful launch.

Organize in cross-functional squads, not a relay of handoffs. Each squad should carry the capabilities a workflow actually needs: data engineering, AI and platform engineering, a domain expert who can define what "correct" means, an embedded risk or security reviewer, and change enablement. The relay model — business writes requirements, engineering builds, risk reviews at the end — is exactly what produces a control spine bolted on in month nine, and evidence you did not capture in month one cannot be reconstructed later. Keep the platform central and the domain semantics federated: the platform team owns the runtime, tool registry, eval harness, identity, and cost telemetry, while domain squads own their data products, definitions, and workflows.

Make decision rights the speed lever. Most delivery time in AI programs is spent waiting, not building. The AI Governance Council should approve autonomy tiers once, per action class, so squads can ship inside a pre-agreed envelope without queuing for a review on every release; anything outside the envelope goes through the exception process. This is where the operating model and the platform meet — DORA's 2025 research across nearly 5,000 technology professionals found a direct correlation between the quality of an organization's internal platform and its ability to unlock value from AI, and describes AI as an amplifier that magnifies both the strengths of high-performing organizations and the dysfunctions of struggling ones [4]. A weak operating model does not stay weak under AI. It gets worse faster.

6. What Are the Risks / Challenges in This AI Transformation? How Would You Mitigate Them?

Major risks / challenges along with mitigation are:

Data is not AI-ready: AI-ready data has a higher bar than analytics-ready data, due to the quality at the cadence the AI agents and models consume the data. If you don't have data products with named owners and quality gates, do not work on the AI workflow using that data.

Cost escalates invisibly: Token, tool-call, and compute spend are difficult to decompose. Focus on per-workflow tokenomics from day one. Report cost per successful outcome. Kill AI workflows that don't create value.

Model churn hides regression: Models change very regularly. Without version-pinned evals, decisions made by AI and AI agents can change or drift, impacting customer deliverables. Treat evals as versioned assets that travel with the workflow. Create a quarterly portability test against a second provider.

Agent authority sprawls: Machine identities are created faster than they are reviewed, and an agent can act wrongly at machine speed. Register agent identities with a revocation SLA. Create scoped autonomy per action class. Test for safe write paths, idempotency, and compensating transactions.

Adoption stalls: This is a lack of proper scoping of review, override, and escalation capacity as part of the prioritization / business case review. Fund review as a capacity line item. Override should never be penalized. Review and learn from Shadow AI.

7. How Can You Measure and Maximize AI?

The Five Measures That Matter

VALUE

EBIT contribution per workflow trend. Cost or revenue delta against a pre-AI baseline and % of workflows with a measured baseline.

UNIT COST

Fully loaded cost per run and per successful outcome. Token cost per unit of work over time. Variance vs forecast.

SCALE

Workflows in production vs pilot. Marginal engineering days to ship use case N+1. Reuse rate of registered data products and tools.

CONTROL

% of agent actions with a complete decision log. Time to revoke an agent credential. % of workflows portable across at least 2 model providers.

PRODUCTIVITY

Revenue per employee trended monthly against headcount. Output per FTE-hour and cycle time per case in the specific workflows AI touched. Measure at the process or team level, never from self-reported individual gain, and always net off the review load created.

What Each Executive Should Measure

For the CEO / BOARD — Percentage of EBIT attributable to AI, with the credibility denominator being the percentage of production workflows with a measured pre-AI baseline, and the operating metric being cost per successful outcome trended monthly against the count of production workflows. Two supporting numbers: revenue from AI-enabled products and services as a share of total revenue, and time from idea to production for a new workflow.

For the CFO — Fully loaded cost per run and per successful outcome, with 100% of AI spend attributable to a named workflow and owner, tracked as variance against forecast. Supporting number: token cost per unit of work over time. If volume grows while unit cost stays flat, the spend is an investment; if both grow together, it's a run rate.

For the COO — Revenue per employee trended against headcount, paired with cost per resolved outcome. Supporting number: containment paired with reopen rate, because a deflection figure quoted alone is quoted without its control.

For the CIO / CTO — Marginal engineering days to ship workflow N+1. If the eleventh workflow costs the same as the first, there is no platform. Supporting numbers: reuse rate and portability across at least two providers.

What NOT to Measure

Tokens used per month (as an absolute, rather than per unit of work), seats provisioned, sessions run, prompts sent, pilots launched, models deployed. These are the metrics most AI dashboards report, and they are easy to collect precisely because they measure activity rather than outcome. Activity metrics also fail in the direction that hurts. They look best right when a program is spending the fastest and proving the least outcome.

If a number would still go up in a quarter where nothing reached production, it does not belong in front of an executive.

8. How to Make This Change Stick

Change management is not the soft edge of an AI program; it is where the value is won or lost. 47% of mid-level managers and individual contributors report an AI-related negative effect, against 31% of executives [1]. The obstacle is not fear. It is ambiguity, and ambiguity is a leadership failure rather than a workforce one.

The benefit of getting this right shows up in roles, not just in adoption curves. A workflow with no human review step removes a task and leaves nothing behind. A workflow with a well-designed review, override, and escalation surface converts an execution role into a judgement role which is more valuable, more durable, and more defensible to the person holding it. That is the difference between roles shrinking and roles evolving, and it is an architectural choice made months before anyone announces it. The same design decision protects the organization's expertise. When a specialist corrects an AI output, that correction is either lost in a comment field or captured as a labelled signal that improves the system. Labelled signal is the difference between expertise being extracted from people and expertise being credited to them, and it is the golden dataset the enterprise produces.

Five moves do most of the work:

Name what is being automated this year and what is not, because specificity is the antidote to speculation and no amount of reassurance substitutes for a list.

Fund review and escalation as an explicit capacity line item, since unfunded review decays into rubber-stamping within a quarter.

Appoint business mentors who are respected practitioners in each domain who use the tools first and present their own results, because a peer showing what they did converts a room faster than any rollout communication.

Build role-based AI fluency in the flow of work rather than as a course catalog, tied to the workflows being deployed.

Reassess quarterly against retention rather than sign-ups.

9. What Is the Next Step?

Start with one workflow that has a countable baseline, in a domain where data is already governed, and build a thin slice through both layers and the spine.

Ready to Turn AI Adoption into Measurable ROI?

Project Brilliant helps leaders connect AI strategy with meaningful business outcomes through diagnosis, strategy, and disciplined execution.

Let's Talk

As an AI Leader in an Enterprise, How Can I Drive Measurable ROI?