Most AI budgets target tasks. The larger expense sits between them.

Every handoff between a planner, a buyer, a controller, and an ERP carries a fee. No one invoices it. No one owns it. Economists have studied that fee since 1937, and McKinsey recently gave it a name: the coordination tax.[2] It may be the clearest path to AI agent ROI.

This analysis traces where the coordination tax accumulates and why agentic AI can reduce it. It also examines why so many agent projects still fail to pay. Platforms such as Zaptiva Automate were built for this layer, where orders, records, approvals, and exceptions move between systems. The argument below rests on published research, cited throughout.

 

Executive Summary

Enterprise AI adoption is broad, yet enterprise returns remain narrow. In McKinsey’s 2026 global survey, 80% of respondents said AI improved their individual productivity. Only 37% attributed any EBIT impact to their organization’s AI use. That share was essentially unchanged from 2025.[1]

The gap has a structural explanation. Most deployments speed up work inside individual workflow steps. Much of the cost, however, accrues at the handoffs between steps. McKinsey estimates that coordination consumes 35% to 60% of total work time in knowledge-intensive organizations.[2] At one Fortune 500 manufacturer, waiting at handoffs exceeded processing time by roughly 9 to 36 times.[2]

Agentic AI can verify, reconcile, and route work at those interfaces, which earlier automation could not do reliably.[2] The return therefore depends less on model selection and more on interface redesign. Organizations that baseline handoff latency, price it with finance, and automate verification first are better placed to prove ROI.

 

Five Key Takeaways

  • The largest AI returns tend to sit between steps. Verification, reconciliation, approval, and waiting can absorb 35% to 60% of knowledge work time.[2]
  • Most of the tax never reaches the income statement. McKinsey found visible coordination headcount was typically about one-third of the total.[2] Latency and foregone revenue made up the remainder.[2]
  • Bolting AI onto old workflows rarely pays. Nearly three-quarters of AI high performers report fundamentally redesigning workflows. About one-quarter of other respondents do.[1]
  • Cancellation risk is material. Gartner expects over 40% of agentic AI projects to be canceled by the end of 2027.[3]
  • Measure the handoff before buying the agent. Latency, touches, exception rate, and rework give finance a baseline it can audit.

 

I. What Is the Coordination Tax?

The coordination tax is the total cost of moving work between people, teams, and systems. It includes the labor of checking, reconciling, approving, and chasing. It also includes the cost of waiting. McKinsey frames it as the cost of human-mediated handoffs between workflow steps.[2]

The idea is older than the label. In 1937, Ronald Coase asked why firms exist at all.[4] His answer was that organizing work internally can cost less than contracting for it in markets.[4] Those costs did not disappear inside the firm. They migrated into approval chains, status meetings, reconciliations, and inboxes.

Thomas Malone and Kevin Crowston later defined coordination as managing dependencies between activities.[5] The definition is useful because it is operational. Every dependency needs a manager, whether a person, a rule, or an agent.

The three layers of the tax

McKinsey’s analysis locates the coordination tax in three places, only one of them visible in SG&A.[2]

  1. Visible overhead. Planning teams, project offices, and review forums. In manufacturing and industrial firms, McKinsey estimates this at 3% to 5% of revenue.[2]
  2. Latency. Each day a handoff waits extends cycle time. Inventory sits longer, receivables arrive later, and capacity idles.[2]
  3. Foregone revenue. Launches slip and deals stall because decisions move too slowly.[2] This layer appears in no financial statement.

Cost programs usually target the first layer because it is the one they can count. By McKinsey’s estimate, the other two layers typically hold about two-thirds of the total.[2]

 

II. Why Coordination Costs Outgrow Headcount

Coordination cost scales differently from labor cost. Frederick Brooks observed that a team of n people has n(n-1)/2 potential communication paths.[6] Ten people share 45 possible channels. Fifty people share 1,225.

Line chart showing how the coordination tax grows as potential communication paths rise quadratically with team size

 

Disclaimer: Illustrative model. Values show the theoretical maximum number of communication paths under Brooks’s n(n-1)/2 formula. Real organizations use a fraction of these paths, so coordination load will vary by structure, tooling, and process design.

In practice, no organization uses every path. They prune them with hierarchy, which is itself a form of coordination overhead. The arithmetic still explains a familiar pattern. Each added team, region, or approval layer can create more handoffs than it removes.

The same math governs software. A company running twelve business applications has 66 possible point-to-point connections. Every needed link left unbuilt becomes a human relay: an export, a re-key, a reconciliation email.

Survey and telemetry data point in the same direction. Asana’s 2023 index found knowledge workers spend 58% of their day on “work about work.”[7] Microsoft’s 2025 telemetry found its most heavily messaged users interrupted about every two minutes in core hours.[8] These studies measure different things with different methods. None isolates coordination cost cleanly. Their direction, however, is consistent.

 

III. Where Do AI Agents Create ROI?

AI agents tend to create ROI where three conditions overlap. Handoffs are frequent. Judgment is bounded by rules. Delay carries a measurable cost.

McKinsey describes three functions agents can perform at an interface.[2] They verify that upstream output meets downstream requirements. They reconcile competing constraints. They route work onward and escalate what exceeds their confidence.[2]

Interfaces that typically meet those conditions include:

  1. Order to cash. Checking orders against credit, pricing, and inventory before an invoice posts.
  2. Procure to pay. Matching purchase orders, receipts, and supplier invoices, then routing mismatches.
  3. Demand to production. Reconciling forecasts against capacity and supplier positions.
  4. Record to report. Reconciling subledgers and intercompany balances ahead of close.
  5. Partner and EDI exchange. Validating inbound documents against master data and trading terms.
  6. CRM to ERP. Turning closed deals into accurate orders, projects, and billing schedules.

The composition of interface work is revealing. At one manufacturer’s demand and supply interface, McKinsey found roughly 65% routine verification.[2] About 25% was rule-bound reconciliation, and about 10% was genuine exception handling.[2] If that pattern generalizes, most interface work is codifiable. A smaller residue still needs human judgment.

 

Chart comparing interface latency and processing time at six workflow handoffs, showing where the coordination tax concentrates and where AI agents can create ROI

Disclaimer: Approximate ranges adapted from a single disguised manufacturing case reported by McKinsey (2026). Figures are directional and show where delay tends to concentrate. Every organization’s handoffs differ, which is exactly why measuring your own baseline comes first.

 

What the field evidence says

Task-level gains are real. In a study of 5,172 customer-support agents, AI assistance raised resolved issues per hour by about 15% on average.[9] Gains were largest for less experienced workers.[9]

Interface-level gains appear larger, though the evidence is younger. An INSEAD field experiment randomized 515 start-ups that received identical AI tools. Firms that redesigned work across handoffs generated about 1.9 times more revenue than firms using AI within existing steps.[10] The sample is early-stage, so effects at enterprise scale may differ.[10]

The manufacturer McKinsey studied offers an enterprise data point. After it redesigned its planning interfaces, the planning cycle fell from 30 days to three. About 80% of planning became touchless. That result came from one disguised case and included broader platform consolidation.[2] It is best read as an illustration of scale.

 

IV. Why Do Most Agentic AI Projects Stall?

Gartner expects over 40% of agentic AI projects to be canceled by the end of 2027.[3] It cites escalating costs, unclear business value, and inadequate risk controls. Gartner also estimates that only about 130 of the thousands of agentic AI vendors are real.[3] It attributes much of the gap to “agent washing,” the rebranding of assistants, chatbots, and RPA tools.[3]

Other research points the same way. BCG’s 2025 study classified only 5% of companies as achieving AI value at scale.[11] Another 60% reported no material value at all.[11] An MIT NANDA report found 95% of studied organizations saw no measurable P&L return from generative AI.[12] It drew on interviews, a leader survey, and a review of public initiatives. The figure is best treated as directional.[12]

Four failure patterns recur across this literature.

  • Automating the box while ignoring the arrow. A copilot drafts a forecast faster, and the forecast then waits days for approval.[2]
  • No one owns the interface. Each function owns its step. The gap between steps belongs to nobody, so nobody funds its redesign.[2]
  • There is no baseline. Teams struggle to prove ROI when handoff latency was never measured before launch.
  • Governance arrives late. Controls bolted on after deployment can consume projected savings. Run costs matter too: about one in five McKinsey respondents said AI operating costs constrained their use.[1]

All four sit in the operating model, where a better language model offers little help.

 

V. How Do You Measure AI Agent ROI at the Interface?

Measure the coordination tax before automating it. The sequence below gives operations and finance a shared, auditable baseline.

  1. Choose one workflow with a clear start and end. Order to invoice or receipt to payment works well.
  2. Map every handoff. Record who passes the work, what gets checked, and which system holds the record.
  3. Baseline four measures per interface. Track waiting time, touches per transaction, exception rate, and rework rate.
  4. Classify the interface work. Separate routine verification, rule-bound reconciliation, and true exceptions.
  5. Price the delay with finance. Convert latency into carrying cost, idle capacity, and revenue at risk.
  6. Automate verification first. Set confidence thresholds, keep people on exceptions, and log every decision.
  7. Re-measure at fixed intervals. Compare results with the baseline at 30, 60, and 90 days. Report the change to finance.

A simple expression keeps the business case honest:

Interface ROI = (latency savings + labor savings + error savings – run cost) / implementation cost

Run cost belongs in the formula because agents consume compute, monitoring, and human oversight. McKinsey reports that successful interface redesigns devote roughly 70% of investment to people and process.[2] About 20% goes to technology and data integration, and about 10% to model development.[2] A budget that inverts that ratio deserves a second look.

 

VI. Start With the Costliest Handoff

The coordination tax rarely appears as a budget line, so it rarely receives a budget. That invisibility is also the opening. One measured handoff gives finance a baseline and gives operations a controlled test.

Pick the interface where work waits longest, and errors cost most. Automate its routine checks, keep people on its exceptions, and measure the change. If latency falls and the exception queue shrinks, the case for the next interface is already on paper.

This is the layer Zaptiva builds for every day: the work between systems. One Zaptiva customer, CBG, reports cutting a seven-day reporting process to less than one day.[13] Bring us one stubborn workflow, and we will map its handoffs with you. Schedule A Demo to find where your coordination tax runs highest and where AI agents can earn their keep.

 

VII. Key Reference Sources

  1. Source [1]: Tinkoff, D., Van der Veken, L., Chui, M., and Balakrishnan, T., “The State of AI in 2026: On the Road to ROI,” McKinsey & Company, August 25, 2026. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
  2. Source [2]: Sankur, A., Mühlreiter, B., and Fuchs, S., “Cutting the ‘Coordination Tax’: How Agentic AI Can Reshape Workflows,” McKinsey & Company, September 18, 2026. https://www.mckinsey.com/capabilities/tech-and-ai/our-insights/cutting-the-coordination-tax-how-agentic-ai-can-reshape-workflows
  3. Source [3]: Gartner, Inc., “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027,” press release, June 25, 2025. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
  4. Source [4]: Coase, R. H., “The Nature of the Firm,” Economica, Vol. 4, No. 16, 1937, pp. 386-405. https://doi.org/10.1111/j.1468-0335.1937.tb00002.x
  5. Source [5]: Malone, T. W., and Crowston, K., “The Interdisciplinary Study of Coordination,” ACM Computing Surveys, Vol. 26, No. 1, 1994, pp. 87-119. https://doi.org/10.1145/174666.174668
  6. Source [6]: Brooks, F. P., Jr., The Mythical Man-Month: Essays on Software Engineering, Addison-Wesley, 1975 (Anniversary Edition, 1995).
  7. Source [7]: Asana, Inc., “Asana Anatomy of Work Global Index 2023,” press release via Business Wire, March 8, 2023. https://www.businesswire.com/news/home/20230308005268/en/Asana-Anatomy-of-Work-Global-Index-2023
  8. Source [8]: Microsoft WorkLab, “Breaking Down the Infinite Workday,” Work Trend Index Special Report, June 17, 2025. https://www.microsoft.com/en-us/worklab/work-trend-index/breaking-down-infinite-workday
  9. Source [9]: Brynjolfsson, E., Li, D., and Raymond, L., “Generative AI at Work,” The Quarterly Journal of Economics, Vol. 140, No. 2, 2025, pp. 889-942. https://doi.org/10.1093/qje/qjae044
  10. Source [10]: Kim, H., Kim, D., and Koning, R., “Mapping AI into Production: A Field Experiment on Firm Performance,” INSEAD Working Paper 2026/20/STR, March 30, 2026, as reported in Source [2].
  11. Source [11]: Boston Consulting Group, “The Widening AI Value Gap: Build for the Future 2025,” BCG, October 2025. https://media-publications.bcg.com/The-Widening-AI-Value-Gap-October-2025.pdf
  12. Source [12]: Challapally, A., Pease, C., Raskar, R., and Chari, P., “The GenAI Divide: State of AI in Business 2025,” MIT NANDA, July 2025, as summarized in Let’s Data Science, 2025. https://letsdatascience.com/news/mit-report-documents-genai-pilot-roi-gap-e0924d7d
  13. Source [13]: Zaptiva, “CBG Customer Story,” Zaptiva, 2026. https://www.zaptiva.com/customers/cbg/

 

Disclaimer: This article is provided for general informational and educational purposes only. It does not constitute legal, financial, accounting, tax, or investment advice. Statistics, forecasts, and case examples come from third-party research current at the time of writing. They may change and may not reflect results any specific organization will achieve. Charts present illustrative or approximate figures adapted from the cited sources. References to third-party organizations, studies, or products do not imply endorsement or affiliation. Customer results described here reflect individual experiences and are not a guarantee of future outcomes. Readers should consult qualified professionals before making business decisions based on this content.