The presentation took twenty minutes instead of two hours. The developer finished a first draft before lunch. The product manager turned a pile of interview notes into a useful summary before the next meeting.
Then everything waited for approval.
Welcome to a familiar stage of enterprise AI: faster work, same queues.
The individual gains are encouraging. In McKinsey’s 2026 survey, 80% of respondents say AI has improved their productivity. But only 37% report a positive contribution to their organization’s earnings before interest and taxes—a share essentially unchanged from the previous year. These are reported experiences, not a controlled experiment, but the gap is worth paying attention to. McKinsey, The State of AI in 2026
It raises a question that another software purchase probably won’t answer:
What happens to all that saved time?
A faster task is not a faster business
Imagine a product team using AI to summarize research, write specifications, and produce code.
Each task gets faster. Everyone has a convincing demonstration. The usage dashboard looks excellent.
But releases still wait for the same architecture review. Customer questions still bounce between product, sales, and support. Nobody can resolve a priority conflict without finding thirty minutes in three executives’ calendars.
The team has accelerated production. It has also accelerated the arrival of work at the next bottleneck.
This is how genuine improvements can coexist with disappointing business results. The time saved in one part of the system gets absorbed by waiting, rework, or coordination somewhere else.
There is an economic step missing, too. Saved hours create capacity. That capacity becomes financial value when the organization uses it to serve more customers, avoid costs, improve retention, or reduce expensive mistakes.
If none of those things changes, the spreadsheet does not award a productivity bonus.
The constraint may have moved
Early in an AI rollout, the obvious questions are about tools. Does the model perform well enough? Can it access the right information? Will people use it?
Those questions still matter. Reliability, integration, cost, and skills can all limit progress.
But once people are getting useful results, leaders need to look beyond the individual task. The next constraint may sit in how work is organized: who decides, how teams hand work to one another, what information they share, and what happens when something goes wrong.
That is the operating model. Less mysterious than it sounds, and considerably harder to change than a software subscription.
McKinsey’s research associates stronger AI performance with fundamental workflow redesign. BCG’s research similarly emphasizes connecting AI deployment to changes in how work gets done and tracking the resulting financial impact. These findings support redesign as a serious management priority, rather than proving it is every organization’s single biggest problem. McKinsey, BCG, How Leaders Build an AI-First Cost Advantage
For organizations already seeing task-level gains, the bottleneck may have migrated.
Follow the work, especially across the handoffs
Return to that product team.
Making specifications faster may help. But if the longest delay comes from unresolved customer questions, a better specification generator will have limited effect.
The useful questions are more specific:
Where does work spend the most time waiting?
Which decisions get reopened because the right people or information were missing?
Who can resolve a disagreement across teams?
Does anyone own the outcome from beginning to end?
Departments usually have their own targets. Customers experience the combined result.
Sales may be rewarded for closing quickly, delivery for controlling scope, and support for resolving tickets. Each team can meet its targets while the customer repeats the same problem three times.
AI can help each department move faster. Someone still needs to make the whole journey work.
Give AI the context your people keep in their heads
Organizations run on a surprising amount of unwritten knowledge.
“That customer has a special agreement.”
“We stopped using that definition last quarter.”
“Check with Priya before promising a delivery date.”
People learn these exceptions through experience and conversation. An AI system needs a reliable way to retrieve the relevant information when it matters.
A larger document collection does not automatically solve that problem. Conflicting policies, outdated records, and unclear definitions can make a system confidently inconsistent.
Start with the workflow you want to improve. Identify the information required to make its decisions, which sources are authoritative, and who keeps them current.
This work helps people as well as AI. Your newest employee would also appreciate knowing which of the six documents titled “Final Process” is actually final.
Name the owner—and make oversight workable
Someone needs to own the business result, with enough authority to change the workflow that produces it.
That does not mean the CEO approves every AI-assisted action. It means responsibility should survive the moment work crosses a departmental boundary.
When AI moves from assisting people to taking actions, the design needs another layer. A system that drafts a customer response and one that issues a refund require different permissions and controls.
Be explicit about what the system may do, when it must ask for help, and who can stop it.
Human review needs design, too. A person asked to approve hundreds of outputs with little context and no time to investigate is being given a queue, not a meaningful control.
Useful oversight gives reviewers the relevant evidence, clear escalation rules, and the authority to intervene.
Start with one workflow, not the whole company
“Redesign the operating model” can sound like the opening slide of a very expensive project.
The first step can be much smaller.
Choose one workflow that matters to customers or business performance. Establish a baseline: elapsed time, rework, cost, or another relevant outcome. Find the main constraint. Name an owner who can change the necessary handoffs or decisions.
Then test a specific improvement.
For the product team, that might mean resolving customer questions earlier, simplifying a review, or giving someone clear authority to settle scope. AI may support the change. The test is whether the entire workflow performs better after accounting for review effort, errors, and operating costs.
A faster draft is useful. A shorter path to a successful customer outcome is more valuable.
Many organizations are becoming AI-rich: more tools, more usage, more impressive demonstrations.
Becoming AI-powered requires a further step. The organization has to turn those capabilities into a better way of working.
Before buying the next tool, follow one piece of work from beginning to end.
Pay particular attention to where it stops.
Sources
BCG / KPMG 2026 enterprise AI ROI surveys (via Value Add VC synthesis); KPMG Q1 2026 Global AI Pulse
McKinsey 2026 AI productivity / EBIT findings (via synthesizer reporting in research pack IR-2026-09-26-001)
MIT NANDA 2025–26 Enterprise AI Survey
EY Work Reimagined Survey 2025; EY, "How agentic AI can help unlock enterprise value at scale" (2026)
Deloitte, State of AI in the Enterprise (2026)
Teradata, Arrested Automation: 2026 Agentic AI Report (vendor-sponsored)
Anthropic, The 2026 State of AI Agents Report
BCG, "The Authorization Gap in AI Agent Governance" (2026)
Cloud Security Alliance / Strata AI agent governance reporting (2026)
Thorsten Meyer AI, on human-in-the-loop oversight and governance theater (2026)
Berkeley CMR, "Governing the Agentic Enterprise" (March 2026)
