Cutting the Coordination Tax: Redesign One Workflow First

The demo finishes in minutes. The ticket still takes a week. That gap is usually not model quality. It is the time work spends waiting between people, queues, and systems—status checks, re-entry, approvals that nobody owns, and rework after a silent handoff. McKinsey’s industrials framing calls that friction the coordination tax: the opportunity sits between steps, where work passes across teams and tools, and agents help only when those seams are redesigned rather than decorated.

The useful claim is not that agents will automate jobs. It is that organizations should treat one end-to-end workflow as the product, cut avoidable coordination inside it, and measure cycle time and exception rate instead of demo polish. Deloitte’s agentic-transformation research maps the operating-model conditions that make that redesign possible—or block it. OpenAI Academy turns the same idea into a practical unit of change: one recurring workflow with a clear start, end, owner, and named human controls.

In a redesigned path, verification, reconciliation, and routing can run as agent-owned steps, with exceptions sent to human judgment rather than every interface requiring a person.

Most of the cycle is not the work

Lean and Kanban communities have measured this for years under a different name. David J. Anderson’s synthesis of workflow duration data defines flow efficiency as value-adding touch time divided by total lead time. Best case studies approach roughly 40 percent. Typical ranges sit much lower—often toward 1–25 percent, and without deliberate focus on delay frequently under 10 percent. The corollary is blunt: most of an item’s calendar time is waiting, not doing.

Michael Hammer made the same point for IT-heavy process change in “Reengineering Work: Don’t Automate, Obliterate” (HBR, 1990). Automating an old path, he warned, is paving the cow paths. In the Mutual Benefit Life example he cites, applications typically took five to 25 days—most of that time spent passing work between departments. Another insurer estimated 22 days elapsed against 17 minutes of actual work. Speeding those minutes with software left the days intact until the process itself changed—fewer handoffs, decisions closer to the work, outcomes owned end to end.

Agentic systems that only accelerate a single step inherit the same trap. Parallel execution and shared context matter precisely because sequential handoffs are where the tax accumulates. McKinsey’s broader agentic playbook, Seizing the agentic AI advantage (June 2025), states the mechanism directly: agents reduce cycle time when they remove delays between tasks and run steps concurrently, instead of waiting for the next human inbox.

A 2026 research paper formalizes a related quantity for multi-agent and organizational settings. In “When Coordination Is Avoidable”, the authors classify 65 enterprise workflows across the APQC Process Classification Framework and find 74 percent monotonic under their decompositions—structurally able to proceed without coordination for correctness—then extend the idea to O*NET tasks. Under their stated overhead assumptions, they estimate that up to 24–57 percent of coordination spending may be unnecessary for correctness. Those figures are upper bounds under modeling assumptions, not a surveyed industry average. They still sharpen the operator question: which sync points exist because the work requires them, and which exist because the old org chart does?

Layering agents is a bridge, not the destination

Deloitte’s August 2026 study surveyed 501 U.S. leaders involved in agentic strategy or implementation (April–June 2026) plus 20 executive interviews. Adoption is real and uneven: 42 percent were still testing small numbers of agents or early deployments, 43 percent were expanding across functions, and 15 percent had reached scaled, orchestrated multi-agent adoption. Nearly two-thirds said their organizations were reevaluating the business model because of agentic AI. The blockers were familiar and concrete: lack of a unified data foundation (72 percent), inability to trust and govern agents (70 percent), and integration cost and complexity (67 percent).

Preparedness for the work redesign is weaker still. Fewer than half of surveyed leaders said they were prepared for agentic AI across six aspects of the business; workforce and business processes were the weakest. Only 16 percent said business processes were prepared (5 percent highly prepared). Even organizations with scaled multi-agent adoption reached only 46 percent process preparedness. Just one in five leaders said their organization was prepared to redesign processes to run autonomously with AI agents.

In interviews, Deloitte found many teams “layering” agents on top of existing processes because short payback windows favor incremental ROI. Layering can build fluency and credibility. Deloitte is explicit that it is not the destination: without a longer commitment to redesign, layering tends to stall at local optimization. Expectations reflect that timeline—31 percent expected at least half of processes redesigned around agents within two years, rising to 74 percent within four. Three-quarters of leaders said human–agent collaboration creates more value than pure automation, yet fewer than half had defined the operating models to make that collaboration real. Half said their organizations were not investing enough in the workforce changes agentic adoption requires.

McKinsey’s June 2025 report describes the same pattern as the “gen AI paradox”: nearly eight in ten companies report using gen AI, and roughly the same share report no significant bottom-line impact; more than 80 percent still report no material earnings contribution, and only 1 percent view their gen AI strategies as mature. Fewer than 10 percent of deployed vertical use cases make it past pilot.

A copilot assists a person inside a step that person already owns—it drafts, summarizes, or suggests—while the person still moves the work to the next team and the handoff still waits. An agent takes a bounded action on the work itself: verify, reconcile, or route, escalating only what falls outside its limits. Bolting an agent onto the old path and redesigning the path so the agent owns the interface are both agents; the difference is whether the old handoffs remain.

On a call-center impact ladder, McKinsey illustrates modest gains from copilots on the existing process, larger but still bounded gains from agents bolted onto the current workflow, and much larger cycle-time effects only when the process is reimagined around agent autonomy. Those upper rungs are illustrative redesign scenarios, not industry averages—but they match the Deloitte diagnosis: overlay and redesign are different products.

Consulting peers land in the same place. Bain’s 2025 technology report says AI leaders delivered material EBITDA gains by scaling across core workflows, while most organizations remain stuck in experimentation and microproductivity—and that the important work is process redesign plus data and application cleanup, with human-in-the-loop remaining the pragmatic posture for years. BCG (January 2026) likewise argues that step-change value requires end-to-end redesign, not incremental automation of discrete steps. An Accenture unpublished analysis, reported in a January 2026 HBR article (Accenture reprint), estimates that more than four in ten working hours relate to transactions inside business processes—coordination friction by another name.

If the success metric is still “the agent demo looked good,” the organization is still layering.

Start with one recurring workflow, not a platform tour

OpenAI Academy’s Activator Labs worksheet (August 21, 2026) is deliberately narrow. Choose a recurring workflow the team owns or can influence. Describe the work, not a product, feature, or prompt. Name who does it, who depends on the result, the observable start event, the observable completion, the business priority it supports, the accountable owner, and what sits outside scope. The scope check collapses into one sentence: when X happens, these people need to reach Y so that Z—and this redesign does not cover W.

Mapping current state is where the coordination tax becomes visible. Trace the real path, including handoffs. Mark where a person interprets, judges, or approves. Mark where work waits, repeats, breaks, or becomes inconsistent. Friction is a clue to investigate; it does not prove AI is the fix. Then define a better outcome for the people doing or receiving the work—not “use AI”—plus a success signal, a failure signal, and a safeguard that cannot be compromised.

Responsibility comes next, before tooling. Academy separates three modes: AI completes a bounded, reversible step; AI supports by drafting, comparing, summarizing, or flagging for a named person; people own sensitive, ambiguous, consequential, policy-required, or hard-to-reverse work. Every AI contribution needs a human control: who reviews what, what interrupts the routine path, who receives the exception, and who remains accountable for the overall result. The readiness recommendation is Proceed, Revise, or Stop. Stop is not failure of imagination. It is the honest finding that ownership, decision criteria, usable information, permissions, or review capacity are missing—and that AI cannot invent those prerequisites.

That unit of change aligns with Bain’s “redesign entire workflows, not siloed use cases” and with McKinsey’s shift from “Where can I use AI in this function?” to reinventing the process so agents can own a large share of the routine path while humans supervise exceptions. One workflow is enough to learn whether the tax moved.

Put judgment and accountability where they already belong

Autonomy without named accountability fails governance tests. A September 2026 MIT Sloan Management Review piece, drawing on a BCG Responsible AI panel, reports that 72 percent of a 29-expert panel agreed or strongly agreed that responsible governance treating agents as autonomous decision makers will fail. Operational autonomy is not moral or legal accountability. The practical recommendations match the Academy worksheet: calibrate autonomy to stakes and reversibility; enforce limits in architecture—permissions, gates, hard stops of the kind explored in Incident Lab—not only in policy; name a human accountable for every consequential decision before deployment; govern the sociotechnical system, not the model card.

NIST’s AI Risk Management Framework (AI RMF 1.0) encodes the same discipline in GOVERN, MAP, MEASURE, and MANAGE: define human–AI configurations and oversight, document human oversight processes, and measure trustworthiness in deployment-like conditions with feedback and appeal paths—not only in demos.

On this site, that design question shows up in regulated claims work as where the human signature belongs: a loop is a trigger, a qualified reviewer, and an artifact—not a slogan. Pattern-level multi-agent claims and benefits workflows are documented in the agentic claims case study at disclosure-safe depth. Across domains, the test is identical. If two design docs disagree about who decides, the gate was never built.

McKinsey’s industrials essay on cutting the coordination tax puts harder numbers on the same seam. Coordination—aligning, verifying, reconciling, waiting, and handing work between steps—consumes 35 to 60 percent of total work time in knowledge-intensive organizations. That share is distinct from Anderson’s flow-efficiency ranges above: it measures coordination as a portion of work time, not touch time divided by lead time. At one Fortune 500 industrial manufacturer, a demand-to-production workflow had seven steps and six interfaces. Interface latency totaled 9–18 days; processing inside the steps totaled 12–24 hours; latency exceeded processing by about 9 to 36 times. The estimated coordination interface tax for that single workflow was $140 million to $240 million a year, while visible people dedicated to coordination accounted for under 7 percent of that total. A later redesign of demand-to-production planning compressed the planning cycle from 30 days to 3, made about 80 percent of planning touchless, and cut reaction time from weeks to less than one day.

Measure cycle time and exceptions, not demo quality

Academy’s success and failure signals are the evaluation contract for the redesign. Useful operator metrics track the tax directly:

• End-to-end lead time from start event to completion
• Time in wait or blocked states versus touch time
• Handoff count and re-entry rate
• Exception rate and time-to-resolution for exceptions
• Rework rate after a “done” signal
• Share of volume on the routine path versus human-owned path—with the residual always owned by a named role

MIT Sloan’s overview of agentic AI notes a clinical deployment where roughly 80 percent of the work was data engineering, stakeholder alignment, governance, and workflow integration—not prompts—and warns that reclaiming 20 percent of someone’s time is not the same as a 20 percent labor-cost saving. NIST’s MEASURE function pushes the same point: performance has to hold in production-like conditions, with monitoring as ongoing operating cost. Production evaluation patterns for agents—trajectory checks, approval binding, failure interfaces—are covered separately in how to evaluate production AI agents.

Demo accuracy without those metrics is a local maximum. The organization can look sophisticated while the calendar time of the workflow barely moves.

What redesign around the tax looks like

When a recurring workflow is redesigned around coordination rather than around a model, several moves recur:

Collapse avoidable handoffs. If two teams only exchange status and retyped fields, the agentic design should share one structured context and one owner of the next state—not add another messenger between them.

Separate routine path from judgment path. Let AI complete or support low-stakes, reversible steps. Keep ambiguous, irreversible, or regulated decisions with people. Publish the stop conditions that divert work to a human, and name who receives that diversion.

Prefer concurrency where correctness allows. Sequential approval chains that do not depend on each other are pure tax. Parallel retrieval and drafting under a single merge-and-review gate often beats a longer chain of partial approvals.

Instrument the seams. Logging model tokens is not the same as logging wait states. If the dashboard cannot show where time accumulates between steps, the redesign cannot prove it worked.

Treat layering as scaffolding. Use it to learn data gaps, review capacity, and exception types—then retire the bolt-on once the end-to-end path is rewritten. Deloitte’s bridge metaphor only helps if someone owns the far bank.

McKinsey’s reinvention indicators still apply: high coordination overhead, rigid sequences, frequent human intervention for decisions that could be data-driven, and processes where exceptions dominate. Simple, low-variability tasks may not need full redesign. Complex, cross-functional, exception-heavy workflows usually do.

The operator’s unit of change

Agentic AI does not erase coordination. It changes which coordination is necessary. The organizations most likely to move cycle time are the ones that pick one recurring workflow, map the handoffs honestly, redesign the path so agents remove waiting instead of decorating it, keep judgment and accountability human by design, and score the work on lead time and exceptions. Platform tours and prompt libraries are optional. One owned workflow with measured seams is not.