Helping Clients Turn AI Investments Into AI Returns

Enterprises are racing to deploy AI agents and applying them to workflows that were never redesigned for such optimization. RAND found that more than 80% of AI projects fail, roughly twice the rate of ordinary technology projects, and that the causes are overwhelmingly operational rather than technical. Gartner now predicts that over 40% of agentic AI projects will be canceled by the end of 2027. Those two findings together validate that technology is not the constraint to positive ROI for agentic adoption. The implementation approach is. Drop an agent into a process built around human limitations, and you do not transform the process. You automate its dysfunction, at speed and scale.

That implementation challenge begins before workflow automation. Part 3 explains why AI pilots that succeed under controlled conditions often fail to scale into production, particularly when real data, users, governance requirements and operating ownership enter the environment.

Your Workflows Were Designed for People. That Is the Problem

Every workflow in your company carries hidden design assumptions. Work moves in batches because people work in shifts. Queues and handoffs exist because attention is scarce or capacity limited. Approval layers are stacked because supervision used to be expensive. Agents void those assumptions. They do not batch, they do not wait, and they do not get tired. So, the first question is never “which steps can an agent do?” It is “what should this work look like now that those constraints no longer apply?” That is a current-state to future-state gap imperative that forces explicit choices about what deserves automation at all. In our work, three criteria consistently identify the right candidates:

• High Manual Effort With Variable Outputs: The same inputs producing inconsistent results is a signal that rules-driven automation, not headcount, is the fix.
• Recurring Audit Findings and Control Gaps: Where consistency itself is the value, agents outperform willpower and reminder emails.
• Handoff-Dense Processes: If work spends more time waiting between people than being worked, the redesign opportunity is the queue, not the task.

Just as important is what to leave alone: judgment-heavy exceptions, low-volume edge cases, and any process where you cannot define what a good outcome looks like. If you cannot specify it, you cannot automate it.

Redesign First, Automate Second

An industrial manufacturer we worked with faced sustained inflation and tariff pressure and wanted automation to protect margin. The tempting move was to buy tools and point them at existing processes. Instead, we help leadership redesign the operating model and the underlying processes so the organization could actually consume automation with clarified decision rights, standardized inputs, and workflows rebuilt around what technology does well and what people must still own. This sequence also depends on governance being established before deployment. Part 1 explains how upfront AI governance clarifies decision rights, accountability, and risk guardrails before implementation. The automation that followed was tied to a specific business objective rather than a general ambition to “use AI.” That process sequence: Redesign first, automate second, is the difference between a capability and a science project.

The Human in the Loop Is a Design Decision, Not a Disclaimer

Effective human oversight also depends on preparing reviewers for the judgment their role requires. Part 2 explains how role-based training and workforce readiness translate AI adoption into confident, disciplined use.

Here is the part most redesigns skip. Nearly every enterprise deploying agents claims a human in the loop. Almost none have designed one. Deloitte’s 2026 State of AI in the Enterprise found that 74% of organizations plan to deploy AI agents within two years, while only 21% report a mature model for governing them. And decades of human-factors research on automation bias tells us exactly what happens when review is bolted on rather than designed in: Reviewers drift into approving by default. The gate exists, the log shows a human approval, and no real review ever happened.

If your reviewers approve everything, you do not have oversight. You have paperwork.

Designing the loop means asking, step by step: Where does human judgment actually change the outcome? What does the reviewer need to see to make a real decision rather than a fast one? Which conditions escalate automatically? Then instrument it. Track override and escalation rates like the controls they are; a review step that never rejects anything is not a control. The payoff runs both ways: you concentrate human judgment where it changes outcomes, and you earn the confidence to automate fully where it does not.

Where to Start

You do not need an enterprise program to begin. Pick one workflow that matters to a number your CFO watches. Map it objectively, including the workarounds nobody documents. Then ask two questions: Would we design this workflow this way if agents had existed when we built it? And where, specifically, does human judgment change the result? If either answer makes you uncomfortable, that discomfort is data. It means the opportunity in front of you is not automation. It is redesigned, and the returns belong to the organizations that do it deliberately, before the agents arrive rather than after. We can help you get those returns and would welcome the chance to discuss it.

Continue the Series

This is Part 4 of a four-part series on turning AI investment into measurable business value.

Part 1 explains how upfront governance establishes the decision rights, accountability and guardrails needed to accelerate AI implementation while managing risk: “Upfront AI Governance Accelerates Value. Do Not Bolt It on Like a Brake.”

Part 2 examines how leadership alignment, role-based training, manager readiness and change management prepare the workforce to adopt AI confidently and translate investment into business results: “Workforce Readiness Is the Missing Link Between AI Investment and AI Returns.”

Part 3 explains why successful AI pilots often fail to scale into production and how governance, ownership and deployment planning help organizations move from proof of concept to enterprise deployment: “Your AI Pilot Was Designed to Succeed. That Is Exactly Why It Will Not Scale.”

Sources

1. RAND Corporation, The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed (2024).
2. Gartner, press release: over 40% of agentic AI projects to be canceled by end of 2027 (June 2025).
3. Deloitte, State of AI in the Enterprise (2026).

^{© Copyright 2026. The views expressed herein are those of the author(s) and not necessarily the views of Ankura Consulting Group, LLC, its management, its subsidiaries, its affiliates, or its other professionals. Ankura is not a law firm and cannot provide legal advice.}