The Manager Bottleneck: AI Doubled the Output. The Review Layer Did Not Move.

On May 25, 2026, Harvard Business Review published a piece by Liz Fosslien and Mollie West Duffy that names the pattern most CHROs and Chief Transformation Officers can already feel in their organizations but have not yet figured out how to measure. The headline finding sits in a quote from a manager interviewed for the piece. "Every 30 minutes, someone creates something I have to look at."

That sentence is the entire problem in seventeen words. AI has dramatically increased the rate at which individual contributors can produce reviewable work. The management layer above them, the people who approve, give feedback, sequence, and make decisions, was sized for a world where the output cadence was three or four times slower. The workflow architecture has not changed. The number of managers per individual contributor has not changed. In several large organizations, the count is being actively cut. The queue grows. The queue is now the bottleneck, not the AI.

Two things are true at once. AI is producing genuine throughput gains at the individual contributor level. And those gains are not converting into organizational throughput because the review and decision layer was never instrumented to absorb them. Both being true is what makes the next four quarters expensive for any CHRO who confuses individual productivity reports with organizational performance.

Three independent data points landed inside a thirteen-day window in May 2026, and together they describe the same phenomenon from three different angles.

The HBR piece, May 25. Managers are becoming the bottleneck because AI has accelerated the production cadence of the people they manage. Fosslien and Duffy describe a category they call feedback debt that accumulates inside teams when reviewable output outpaces review capacity. The work product gets produced. The decision about whether to ship it, refine it, or kill it does not get made. Velocity at the individual layer turns into stalled inventory at the organizational layer.

The Gartner press release, May 13. By 2027, half of enterprises without a people-centric AI strategy will lose their top AI talent to competitors who prioritize workforce enablement over basic adoption. Gartner introduced a useful phrase for the gap: the "enablement illusion." The illusion is that giving people access to AI tools is the same as enabling them to do better work. The data says it is not. Access without a management layer that can absorb the resulting throughput produces frustration on the way out the door.

The Forrester future-of-work predictions, continuing into May. HR departments may see staffing cuts of up to 50 percent as executives test how aggressive AI substitution can be inside the function that owns the workforce. Forrester also predicts that more than half of AI-attributed layoffs will be quietly reversed. Both predictions are pointing at the same thing. The review layer is being thinned at the moment the throughput it has to absorb is widening.

The Aon 2026 Human Capital Trends Study, drawn from 2,361 board directors and senior leaders across 62 geographies, frames the macro version of the same gap. AI implementation has gone mainstream. AI readiness has not. The instrumentation that would let an organization tell whether it is in the readiness cohort or the unready cohort is exactly the layer that does not exist in most workforce stacks today.

The Levos team has been having conversations with people leaders across the Fortune 500 to 2,000 since 2024. The same pattern keeps surfacing in different vocabulary.

A senior people operations leader at Google described an environment with manual narrative entry into GRAD, Google's performance tool. The platform is sophisticated. The input remains human, text-based, and synchronous. When the underlying work product an engineer or PM creates accelerates by 40 percent. They get behind. The system cannot tell anyone they are behind, because it was designed to measure roles and ratings, not throughput.

That is a 156,000-person organization with arguably the deepest internal AI capability on the planet. If the review layer is bottlenecked there, it is bottlenecked everywhere.

The same pattern shows up at a 50,000-person defense contractor with no integration across the four. The platform's headcount, budget, and brand vary. The bottleneck does not.

The instrumentation problem is structural. The systems most organizations are paying millions of dollars per year to operate were designed in an era when work output was reviewed in cycles measured in weeks, not in thirty-minute increments. They are excellent at storing roles, ratings, compensation, and survey responses. They are functionally blind to the rate at which work is produced and the rate at which it gets reviewed.

A CHRO who wants to see the manager bottleneck today has to assemble it from four data sources that do not talk to each other: AI tool telemetry (Copilot, Gemini, Glean, internal Claude), work product telemetry (Git, Jira, Linear, Figma, Salesforce), review signal from the performance system, and qualitative signal from manager 1:1 notes. Five vendors, five schemas, no joinable identifier on the people side. Assembling the picture takes a quarterly analytics project that finishes after the bottleneck has already done its damage.

Vendor-supplied dashboards do not close this gap. The dashboards inside any individual workforce tool reflect that tool's view of the worker. They have structural conflicts of interest in surfacing data that suggests their own customer should be running fewer of their own product. A measurement layer that aggregates across all of those tools is, by definition, not a workforce tool. It is the intelligence layer above them, the one that surfaces signal none of them sees alone. That is exactly the position the [Levos platform](https://levos.ai/platform) occupies in the stack.

| Layer | What the workforce stack measures today | What task-level instrumentation measures |

| Individual contributor | Role, rating, tenure, compensation band | Throughput delta vs baseline, AI tool depth of use, output quality trace |

| Manager | Span of control, review cycle completion | Review latency, feedback debt per team, queue depth |

| Team | Engagement survey scores, headline attrition | Throughput vs review capacity gap, strain concentration |

| Executive | Aggregated headcount, summary KPIs | Where the bottleneck is forming this week, and why |

 

A task-level layer does not replace the workforce stack. It joins the operational signal across the tools where work actually happens, normalizes it into comparable units, and surfaces the manager bottleneck inside a working week rather than inside a quarterly analytics review. The [AI Impact signal family](https://levos.ai/ai-impact) is the one that maps directly to throughput delta and depth-of-use measurement. The [Human Capital Optimization](https://levos.ai/platform#human-capital-optimization) view is where strain concentration, retention risk, and the rising attrition signal show up before they hit the exit interview.

Enterprise platforms are not going to solve this for the mid-market in time. Sales cycles are too long, price points are too high, and the integration scope is built for organizations with dedicated workforce analytics teams. Visier's May 2026 Glean integration is a thoughtful step.

 

Mid-market CHROs and Chief Transformation Officers platform build. They need an intelligence layer that sits above the tools they already own and surfaces the bottleneck this quarter. That is what Levos is: a Human Capital Operating System that gives mid-market organizations a single line of sight on throughput, review capacity, and the gap between them. The full approach, including confidence scoring and the audit trail, is described in the Levos measurement methodology https://levos.ai/measurement#methodology