Layers Were Latency

On May 12, 2026, GitLab CEO Bill Staples published a memo called Act 2. He said the company carried eight layers of management for a workforce of under three thousand. He called that too deep. Three of those layers were coming out. About sixty smaller teams would replace the old structure, each with end-to-end ownership. The company would exit roughly twenty-two of the sixty countries it operated in. Agents would handle the reviews, the approvals, and the handoffs that the missing layers used to broker. Roughly 350 people, near fourteen percent of staff, would not be coming with the company into Act 2.

Read the memo carefully and the popular AI layoffs framing falls apart. The work stayed. The translators left. That distinction is the whole story of AI and organizational design in 2026, and the boards that miss it will spend the next two years cutting the wrong layers.

The Translation Tax

For decades the case for management was coordination. Strategy lives at the top. Work lives at the bottom. The work has to be sliced, scheduled, sequenced, approved, escalated, summarized, and reported on. The slicing and sequencing requires people who can speak both languages. The summarizing and reporting requires people who can compress without losing fidelity. Those people sit between the executive and the engineer. Their output is translation, and translation is the product they were hired to produce.

Translation is expensive. Every layer adds latency. Every layer adds a margin of distortion. Every layer adds a salary, an office allocation, a stock grant, and a benefits load. The eight-layer org Staples described was an org paying for seven translation passes between intent and shipped work.

The strange thing about translation as a job is that it scales with the organization it serves. A team of eight needs no translator. A team of eighty needs one. A team of eight hundred needs a hierarchy of them. Coordination cost rises faster than linearly because every new translator has to be coordinated by another translator. This is the problem Frederick Brooks named in The Mythical Man-Month, and it is the reason every founder who has watched a company cross 200 people feels the company get slower at exactly the moment it should get stronger.

Until 2026, no one had a credible answer. You could push back the curve with better tools. Slack instead of email. Notion instead of Word. Linear instead of Jira. None of those tools reduced the translation tax. They made the translators faster. The number of translators kept climbing.

Then agents got good enough to do the translation themselves.

What Agents Eat First

The popular story about AI in 2026 says agents are coming for the bottom of the labor market. Call center work, paralegal review, claims processing, basic copywriting, code generation. That story is incomplete. The hard data from the layoffs of the last six weeks shows agents eating the middle.

Amazon cut sixteen thousand corporate roles in late May 2026. The official reasoning from CEO Andy Jassy was almost word-for-word a description of removing translators. The phrase he used was "reducing layers, increasing ownership, and removing bureaucracy." Coinbase cut fourteen percent of its workforce around the same time and explicitly flattened to five layers. Intuit announced 3,000 layoffs on May 20, 2026, with seventeen percent of staff exiting, framed as a refocus on AI. None of these cuts hit the front-line engineers building features. They hit program managers, internal coordinators, business analysts, area directors, and the people whose calendars were ninety percent meetings.

That pattern is structural. Coding agents still need software engineers to scope problems and review work. Support agents still need humans for hard escalations. Coordination work, which is most of what a middle manager actually does in a given week, sits inside the strike zone of a language model. Status synthesis. Memo drafting. Cross-team alignment. Reading three documents to write a fourth. Translating an executive directive into work for a team that did not hear the original conversation. Compressing a quarter of engineering output into a slide for an executive who never reads the source. Agents do every one of those things, today, in seconds, for cents.

Salesforce already lived through the customer-side version of this. Marc Benioff confirmed in late 2025 that the support headcount went from nine thousand to about five thousand, with half of customer interactions now handled by agents. The customer service org was, structurally, a translation layer between the product and the customer. It went first because the language was narrow and the volume was high. The internal corporate org speaks a slightly wider language at slightly lower volume. It is going next.

The June Tally

TechCrunch's running tally of major tech layoffs in 2026, updated June 22, shows 267 layoff events impacting around 186,000 workers. Roughly 56 percent of the events cite AI, automation, or machine learning explicitly as a contributing factor. That figure represents more than 156,000 people whose roles were named, in writing, as redundant against a software system.

The accompanying story across most of these cuts is record financial performance. The four largest cloud and consumer-AI buyers, Amazon, Microsoft, Alphabet, and Meta, have committed roughly seven hundred billion dollars in capital expenditure for 2026, almost all of it tied to AI infrastructure. They are profitable. They are growing. They are still cutting people. The cuts are a reallocation from headcount to compute, because compute is the cheaper translator now.

The reallocation is structural. Capital expenditure is durable. Once Amazon has committed thirty billion dollars to a campus in Indiana with its own substation and a private power purchase agreement, the people whose work that campus now does are not coming back. The translator role has been moved off the payroll and onto the balance sheet.

This is the part that makes most boards uncomfortable. A capex line is depreciated over years. A salary is reset every twelve months. Boards that grew up on labor-elastic businesses have rarely underwritten decade-long bets on machine cognition. They are about to learn.

Microsoft's Number

The 2026 Work Trend Index, Microsoft's annual research drop on enterprise AI use, landed in late May and contained one number that should reset every operating plan in the Fortune 1000.

Active agents in the Microsoft 365 environment grew fifteen times year on year. At large enterprises the number was eighteen times. Microsoft also reported that thirty-three percent of senior executives globally said they will consider using AI to reduce headcount in the next twelve to eighteen months, and another forty-five percent said they will maintain headcount but use AI as digital labor. Combined, more than three quarters of executives are restructuring the relationship between people and software inside their company on a horizon shorter than most product launches.

The most useful line in the report was about constraint. Microsoft's data showed that organizational factors, by which they meant culture, manager support, and talent practices, accounted for sixty-seven percent of the variance in AI's real business impact. Individual mindset and behavior accounted for only thirty-two percent. The person is ready. The system around the person is the bottleneck.

That finding is the practical heart of AI and organizational design in 2026. The capability is here. Model performance is sufficient for most knowledge work in most companies. What is missing is the institutional shape that lets agents actually do the work. A standard hierarchy with weekly status meetings and quarterly planning rituals cannot absorb an agent that completes a week of work in a day. The agent shows up with output and there is no one whose job is to merge that output into the company. So the output sits in a Slack thread until it goes stale, the executive concludes the agent was overhyped, the company concludes the AI thing was overhyped, and the layoffs at the competitor that did rewire start to look mysterious.

They are easy to explain. The competitor changed the shape.

What a Five-Layer Company Actually Does

The companies that are removing layers do it because eight-layer companies cannot run agents, regardless of what the press release says about costs.

In an eight-layer company, work moves up through summaries and down through directives. The intermediate layers exist to handle the impedance mismatch between a thousand-person workforce and a small executive team. Each layer adds context, removes context, rewords, prioritizes, defends. The system is tuned to a human cadence. A directive issued on Monday is reflected in the work by Friday. A status report compiled on Friday is reviewed by the executive the following Monday. The clock speed is roughly one week per layer.

Agents run at the clock speed of the underlying model, which is closer to seconds than weeks. An eight-layer org will lose every agent it deploys to the impedance mismatch between machine cadence and human bureaucracy. The agent will produce a memo on Tuesday afternoon that needs to be approved by Thursday by a director who only reads memos on Monday, and the company will conclude the agent does not understand the business.

A five-layer company can move agent output into production fast enough that the agent looks like a teammate. Sixty small teams with end-to-end ownership, the GitLab structure, can each absorb agent output directly into shippable work. The shape is flatter because decisions have to land at the speed agents produce them.

This is also why Asha Sharma, Microsoft's AI platform product lead, has been telling reporters that the org chart is on its way to becoming what she calls a work chart. An org chart describes who reports to whom. A work chart describes which capabilities, human and machine, are composed to deliver which outcomes. The work chart is a different governance object. It moves the unit of analysis from the role to the workflow.

You can already see early work charts at the operating level inside companies like Anthropic, GitLab, and Coinbase. The pattern is small pods, durable around a business outcome, with mixed composition of human staff and standing agent services. The reporting line up is short. The autonomy down is deep.

What Has to Change at the Top

This is the part of the AI and organizational design conversation that most boards skip. Removing layers does nothing if the strategy still gets handed down as long-horizon plans against fixed quarterly milestones. A flatter company that still runs on a slow planning ritual produces underemployed executives staring at a backlog of decisions they could have made in real time.

Three things have to change at the top.

First, the planning horizon collapses. Quarterly plans are still useful for capital allocation and external communication, but the internal cadence has to be much shorter. The agent loop is daily or weekly. The human loop should match.

Second, the executive role shifts from translator to composer. Most executives spend most of their week reading what people wrote, asking what it means, and telling people what to do next. Agents now do the reading, the summarizing, the drafting, and the asking. The executive's actual job is choosing which capabilities to compose, which agents to authorize, and which workflows to underwrite. The job is closer to portfolio construction than to people management.

Third, the company has to take ownership of its own AI stack. The companies losing the most ground in 2026 are the ones who paid a vendor to install agents on top of an unchanged org. The vendor cannot redesign the company. Only the company can redesign the company. The Microsoft data is unambiguous on this. Sixty-seven percent of AI's impact is organizational. The other thirty-two percent is what a vendor can deliver. If you only buy the thirty-two percent, you will get thirty-two percent of the result, which on most enterprise AI budgets is a deeply negative ROI.

The Decision in Front of You

If you run a company with more than five layers of management today, you are running a company optimized for a coordination tax that no longer has to be paid. The question is whether you remove layers with intention, around a real operating model that uses agents as first-class participants, or whether you let the market remove them for you through three rounds of cost-cutting layoffs and a slow loss of talent to competitors who already moved.

The companies that move with intention have a few things in common. They start by mapping the actual flow of decisions in the business, not the chart on the wall. They identify which steps in that flow are translation steps. They build or buy agents to handle those steps. They redesign the org around the new shape, with smaller pods, end-to-end ownership, and short reporting lines. They retrain the executives whose job is no longer translation into a job that is composition. They underwrite the compute as capex and treat the model providers as long-term infrastructure partners.

The companies that move with reaction also have a few things in common. They wait. They watch a competitor announce layoffs. They run a pilot, declare it a success, then leave the chart alone. The agent output piles up in a Notion folder. The next quarter, growth misses. The board asks why. The CFO recommends a workforce reduction. The reduction hits the front line and leaves the middle alone, because the middle writes the reports the board reads. Eighteen months later the company quietly sells for parts.

I have seen both versions inside client work this year. The architectural version takes ninety days to plan and about a year to install. It costs less than the reactive version and produces a company that actually runs faster. The reactive version costs morale, brand, and the people you most wanted to keep.

Why This Is Not a Tool Problem

The instinct of most boards is to treat this as a procurement decision. Buy a copilot license. Sign with one of the agent platforms. Hire a fractional chief AI officer. Wait to see results.

That instinct will lose you the next two years.

AI and organizational design is a single problem. The tool does not deliver the result. The shape of the company delivers the result. The tool is necessary and almost completely insufficient. You can hand Claude or GPT or any of the agent runtimes to an eight-layer company and watch it produce no business impact, because the org cannot consume what the agent produces. You can hand the same tools to a redesigned five-layer pod-based company and watch the operating margin move three points in two quarters.

The consulting market has been slow to catch up. Most of the large firms are still selling AI as a technology engagement: choose a model, integrate it with your data, train your people. That work is real and necessary. The harder, higher-leverage work is the org redesign that runs alongside the tooling. Without it, the tooling is a sunk cost. With it, the tooling is the moat.

A Final Argument

The right board memo to write this quarter answers one question: which version of the company do you want to operate in 2027? The five-layer version, designed around composed agents, with executives whose job is portfolio choice rather than translation, is achievable for almost any organization willing to do the architectural work. The eight-layer version, with agents bolted on, is the version that pays for three more rounds of headcount cuts before it figures out what GitLab figured out in May.

This work is architectural. No procurement can substitute for it. You cannot buy your way out. You can only design your way through.

Agor AI Advisory builds these designs with operators who run real companies. We map the decision flow, identify the translation work, install the agents that absorb it, and rewire the org around the new shape. We do this in ninety-day cycles, with executive sponsorship from the top, and we leave behind a company that runs at machine cadence with human judgment in the right places.

If you can see your own company in the GitLab memo, do not wait for your board to read the same news in a quarterly earnings release. Schedule a strategic consultation with us today.