AI Replaces Middle Management: Building Mycelium Organizations for the AI Era

AI-Driven Organizational Evolution: Beyond Middle Management Compression

The article opens with Block's 2026 announcement to cut staff from over 10,000 to under 6,000, linking the reduction to intelligent tools changing how the company operates. Block's new arrangement lets systems provide work context previously passed by managers, makes problem owners organize cross-team collaboration, and assigns hybrid business-and-coaching roles to handle professional development and people growth. The author cautions that not all middle-layer cuts are AI-driven; P&G's 2025 plan to cut up to 7,000 non-manufacturing roles also involves brand, marketing, and supply-chain changes.

Why Middle Management Faces Replacement Pressure

As organizations scale, executives lose sight of team details and frontline staff lack full information and support. Middle managers historically aggregated status, explained decisions, tracked progress, and coordinated interfaces — moving information up and stitching collaboration across. AI now changes part of this equation: when it can access reliable data and integrate with business systems, it can assemble task context from meeting notes, customer interactions, and business records, flag conflicts and gaps, and deliver evidence directly to the person doing the work. Stable-rule tasks like data verification, status tracking, and approved actions can be handled continuously by humans and AI together. Roles built around repetitive reading, context stitching, and layered retelling may merge, shrink, or shift.

However, the article emphasizes that surfacing a resource conflict (e.g., two projects needing the same expert) does not decide priority; categorizing a complaint does not resolve the loss; alerting a newcomer to risk does not teach judgment. The core question: when the middle layer is compressed, who picks up these remaining responsibilities? Removing the old structure without a new mechanism creates a vacuum, not agility.

Mycelium Organization Theory: Strong Core, Agile Frontline

The mycelium organization framework summarizes the direction as "strong core, agile frontline": the organization continuously supplies trusted facts, professional capabilities, shared rules, and risk guardrails, while people close to customers and the field combine capabilities and act within effective authorization. The focus is on turning functions once concentrated in management roles into stable support that teams can reliably access.

Evolution, Not Revolution: Laying New Pipes in an Old City

Enterprises cannot pause customers, orders, or cash flow. Building new organizational capability resembles laying new pipes in an old city: shops above keep operating while underground connections and supply gradually improve. Professional specialization, quality control, financial management, and talent development are accumulated order that retains value. What needs re-examination are situations where information can now reach directly yet still passes through layers; where the frontline has judgment conditions yet still seeks repeated approval; where each department completes its own actions but the customer's problem remains unowned.

Reform tends toward two extremes: full decentralization that dumps resource conflicts and professional disputes onto the frontline, or buying tools and adding software while old waiting and escalation habits persist. The former removes the catcher; the latter just adds a tech layer to old relationships.

Start small: a lead who can coordinate relevant businesses takes a blocked order or complaint, works with the handler and specialists to see what the customer is waiting for and where the work is stuck. Missing data? Add trusted information. Unclear ownership? Clarify handoff. A single interface can solve the problem? Fix the interface. If multiple departments are each correct but together fail the customer promise, rearrange the combination of participants, capabilities, authorities, and responsibilities — this is business-scenario reconstruction, working from the customer outcome backward to decide what to keep and change, not starting by deleting departments.

Stable, repetitive, high-risk work may still need defined processes and central scheduling. When frontline experience is insufficient, necessary management support should stay. The mycelium organization does not eliminate hierarchy; it turns hierarchy from a "signal relay station" into a "capability amplifier."

Let Signals Flow Natively

In traditional organizations, problems travel far: frontline discovers, supervisor compiles, manager filters, director explains. Necessary filtering reduces noise, but when every layer only reports good news, filtering becomes distortion. By the time a capable person sees the issue, details and timing are lost.

Letting signals flow natively means reducing repeated copying and retelling, preserving source, caliber, and changes, and reaching the authorized handler promptly. AI's value is not just faster weekly reports, but helping relevant people know what happened without waiting for the report.

The article illustrates with a manufacturing damage complaint: the customer requests a replacement; customer service verifies actual need, then organizes quality and logistics to judge; a digital employee pulls order, batch, and transport records per approved data. Customer service doesn't wait for a weekly report; quality and logistics don't make the customer repeat the story.

After information arrives, judgment continues. A sign-off record alone doesn't prove no quality issue; a model classification doesn't replace damage-cause verification. When records conflict, each must be examined for what it actually states, and the relevant professional must verify — not auto-accept the newest record. Similarly, finding an expert doesn't mean they're available; seeing inventory doesn't mean it can be reallocated. The network must surface these constraints (expert's existing tasks, parts already committed) or coordination simply shifts to the frontline searching for people.

OA, CRM, ERP continue maintaining their professional facts. Getting necessary information around a specific job and being able to trace back to the source for verification matters more than dumping all company data into a new lake. "One network" does not mean everyone sees everything.

Technology also faces a harder distortion: employees fear bad news will be used for blame, so they keep a parallel shadow system. Whether signals arrive truthfully depends on whether the organization helps handle the problem or first looks for someone to blame.

Different Nodes, One Formation

Direct information access does not mean everyone has equal judgment. People, digital employees, business systems, and external partners can collaborate around an outcome because different capabilities have defined places and know when to seek others' support.

An intelligent agent is a technical form; a digital employee is its organizational identity, requiring clear permissions, guardianship, and exit arrangements. The same digital employee can organize materials, compare options, and execute routine actions within authorization. Humans bear responsibility, but that doesn't mean re-clicking confirmation at every step; machines can participate in judgment but cannot self-expand authority.

Work with stable rules, verifiable outcomes, and easy remediation can run continuously per approved plans. Major resource trade-offs, professional disputes, and key customer commitments need capable, authorized people to decide. The division basis is task and consequence, not simply "humans judge, machines run errands."

Middle management's value reappears here: identifying problems worth solving, organizing support, handling exceptions, developing successors, and gradually turning what only they could coordinate into what the team can deliver.

Judgment is trained, not innate. Employees first learn to verify AI analysis, distinguish fact from inference, then compare options and own choices within clear boundaries, receiving mentor feedback. Business acumen, data literacy, systems thinking, and judgment in ambiguity only become real through actual work.

Such practice requires organizational time investment. Mentors cannot only coach after hours; newcomers cannot be demanded to be independent while denied key data and professional help. Good-faith mistakes are judged by facts, authority, and due diligence at the time; concealment, self-dealing, or knowingly crossing lines are different matters.

Hubs won't disappear, but hubs must be governable. The more work depends on one expert, the more the enterprise should provide backups and develop successors, making judgment criteria inspectable. Enabling others to not need you should be a contribution, not a reason to lose your position.

Set Rules First, Then Give Freedom

"Let go and chaos; tighten and stagnation" often stems from treating authority as a single master switch: either headquarters decides everything or the frontline bears everything. Effective authorization simultaneously provides decision space, professional support, and deliverable resources.

The enterprise maintains shared standards for quality, finance, and qualifications — not deciding every detail for every team. The frontline arranges services and drives projects within clear scope; cross-department resources or broken commitments escalate to the appropriate authority. Once-granted authorization cannot be arbitrarily overturned by a leader's personal preference.

Returning to the damage complaint: the customer service lead organizes all parties to fulfill the customer promise; quality verifies the batch; logistics checks transport; parts owner confirms quantity and timing. Each owns their professional judgment and support commitment. Customer service can demand the organization honor promised support, but cannot bypass quality release or unilaterally claim arbitrary budget authority.

If support fails, the lead organizes remediation; if the issue exceeds authority, it escalates promptly to someone with coordination power, while continuously updating the customer. The customer must not lose their contact point due to internal handoff. Before the new owner confirms acceptance, the original lead keeps organizing — a group chat or forwarded message does not equal handoff.

Documenting the result, boundaries, key judgments, evidence basis, and exception escalations for this work creates a Task Contract (TC). It can live in existing tickets, reference existing rules, and need not be a long form each time. When agreements change, participants get the current version; when authorization expires or is withdrawn, corresponding actions are immediately restricted. Mutual dependents confirm support and continuity through this, giving the responsibility network real substance.

The article describes a scenario where a batch slated for replacement is put on hold: affected shipments stop; customer service seeks alternatives; the guardian accesses key evidence, understands risk, pauses actions within authority, and organizes takeover. Technical recovery, professional re-release, and customer acceptance of new delivery are separate decisions. Restoring a system configuration does not auto-restore all action permissions or erase incurred customer loss.

When these relationships span multiple systems, they can break again. Mido's MOOS (Mycelium Organization Operating System) aims to link orders and professional conclusions in original systems with the effective agreements, owners, and resource commitments for the same work. After a batch hold, it should find affected shipments and customer agreements, route items needing re-judgment to the right people, and capture actual handling results. Original systems remain the authoritative fact store; humans still make professional and business trade-offs; existing tools that work reliably and economically stay in use.

Responsibility cannot fall solely on the last clicker. Result organization, professional judgment, digital capability operation, and authorization each have their own owners. When things go wrong, investigate who introduced it, who should have caught it, who had actual control, and whether the organization lacked necessary guardrails.

Rules must ultimately connect to supply. When an expert supports another team, their original task needs adjustment; when a manager absorbs more exceptions, they need shift cover and backup. If supporters cannot sustain, frontline freedom has no foundation. Set conditions first, then delegate action — conditions must include these people's time.

Change Is Not Sentimental

Organizational evolution redistributes information rights, decision rights, and rewards. Losing a familiar position creates anxiety that a single mindset session cannot erase.

Middle managers were asked to aggregate, coordinate, and escalate; now they're asked to discover problems, organize improvements, and develop teams. If the enterprise only says "transform" without training, time, and new evaluation criteria, the risk falls solely on the employee.

Whether to continue leading a team while also driving improvement, or step back from leadership to focus on improvement, depends on real workload and business need. Faster reports don't mean scheduling, coaching, and exception handling have decreased; if improvement duties only fit into evenings, that's added burden. When necessary responsibilities have no catcher, keeping the management role is justified.

Veterans' fear of experience devaluation has basis. Customer understanding, professional judgment, and nuance in complex situations remain valuable; familiarity with approval paths and workarounds may lose some value. Both types of experience often coexist in one person — cannot erase years of contribution with a label like "information intermediary."

Sharing needs reciprocation. When veterans document methods and mentor others, they may be consulted less. The enterprise should recognize this contribution, allocate maintenance and coaching time. Otherwise, developing successors feels like weakening one's own position.

Transition-period income protection, role trials, mentor support, or alternative work must be arranged per business conditions with credible commitments. The enterprise cannot guarantee everyone lands in an ideal new role, but should clarify choices, requirements, and transition terms so people can learn and voice dissent. What's replaced is function, not human dignity.

The founder's struggle is fear of losing control after delegation. Having relied on personal judgment through crises, they must now allow some decisions to happen without their awareness. New arrangements only build trust when they hold risk, catch exceptions, and respect granted authority in real work.

Such restraint needs leadership-team and governance backing. Whether leaders honor authorization, whether bad news gets response, whether cross-team support and talent development enter evaluations — these must deliver in daily operations. Culture cannot be bought, but it can be built by repeated real management choices, and destroyed by them.

Mycelium Grows — It Is Not Designed

Mycelium is not drawn; it grows. Org charts, appointments, and documents only propose arrangements; capability forms through repeated work. Rollout can have three iterative phases, not a uniform deadline for all departments.

Get one class of work flowing. Pick recurring, checkable-result, controllable-impact problems. Compare customer outcomes, quality, wait times, and total burden using similar-difficulty tasks. If customer service's saved time becomes unpaid extra verification for quality and logistics, don't just report efficiency gains. This phase must clarify: what middle-manager work is now reliably automated? What responsibilities are picked up by qualified people? What conflicts still need central adjudication? If manager follow-ups just become expert overtime or approvals shift to the boss, fix the arrangement first — don't rush to remove layers. Observe under real business and normal support: confirm coordination burden truly drops, quality and risk hold, and new support/maintenance costs are acceptable. Only when necessary responsibilities are caught and overall investment makes sense, adjust roles and levels. One smooth case is not enough for structural decisions. Pilots may also conclude to narrow scope or stop: fixing the interface may suffice; some exceptions may still need human catch. The pilot's job is to inform choice, not prove reform must succeed.

Let success leave the original experts. Make applicable conditions, judgment basis, required support, and failure modes explicit. Ask another team with relevant business capability to handle similar work under normal scheduling, effective authority, and committed support. The original team can hand over experience but must not shadow-operate.

Independent verification. People not involved in design or execution, with relevant capability, verify: did the customer truly get the result? Is the professional bottom line held? Is the investment sustainable? Based on this evidence, the authorized person decides scope and arranges maintenance. Normal professional help can stay; key is that support is real, obtainable, and cost-bearable; organizational capability is not measured by driving human help to zero.

Customer issues meeting agreed conditions should close normally, not wait for experience rollout completion. Worthy methods keep accepting test; when conditions change they can be restricted, corrected, or withdrawn. Low invocation count may mean poor supply, or low-frequency but critical expertise — not a simple reason to cut.

Expand collaboration scope last. When suppliers, partners, and customers join, each has business goals; resource usage, data purpose, fees, disputes, and post-exit obligations must be clear. More nodes add value only when customer results are more reliable and support costs stay acceptable.

As technology, people, and suppliers change, yesterday's working methods will fail. Who maintains, who detects change, how affected work continues — these must be implemented alongside use. Leaving slack for exploration, training, backup, and drills is also an investment in continuous operation.

Unbuyable Underlying Protocol

Culture is not slogans on the wall; it's the organization's default way of working: how decisions are made when no one watches, whether bad news is spoken, what guides action when rules don't cover.

Whether authorization truly works shows in how leadership handles the first good-faith mistake. Encouraging tries beforehand but punishing only by outcome afterward quickly teaches staff to re-escalate. New facts can adjust decisions, but reasons must be explained and consequences owned.

An employee reports "this deadline may not be met." What follows — verification, added support, re-trade-off — or first questioning why they didn't try harder — affects when the next problem surfaces. "I can't judge, need help" should be a responsible action, not a shame marker.

Frontline counter-examples must influence shared rules. If staff see current methods repeatedly cause errors but can only execute, not question, judgment atrophies. Let them report issues; qualified people verify; authorized people decide changes — then frontline experience can improve the whole organization.

The top leader still faces three gates: the power gate — allowing authorized decisions to happen without their knowledge; the cognitive gate — willing to test AI and team capability in real work; the behavioral gate — resisting the urge to do the old things, not making every judgment for subordinates.

This restraint needs leadership-team and governance support. Whether leaders respect authorization, whether bad news gets response, whether cross-team support and talent development enter evaluations — all must deliver in daily operations. Culture can't be bought, but can be built by repeated real management choices, and can be destroyed.

Conclusion: Start Right Here

If action starts tomorrow, don't begin by debating which level to cut. A lead who can coordinate relevant businesses takes a stuck piece of work, works with the handler to ask: what is the customer waiting for? Who has already committed? Which fact, authority, or support is still missing?

First make it work more reliably in a small scope, then let another batch do it under normal conditions. Enterprises can even exchange such desensitized business cases to see whether to fix interfaces, add professional support, or rearrange commitments and decision rights.

Win first, then battle — not wait until everything is clear, but build the most critical conditions before scaling action. New capability can grow from existing systems and mature practices, without tearing down the past as the price.

One complaint handled well, still handled well after personnel change, and when the method fails it's detected, adjusted, and re-verified — only then can the organization sustain growth. Building a mycelium organization is not the endpoint; gaining the "ability to keep building it" is what this transformation must leave behind.

Start, right here.