Leadership is the new management. Or more precisely: leadership is what management becomes when AI handles everything else.

I've spent enough time inside organisations to know that middle management has always been two jobs squeezed into one role. There's the leadership part – coaching people, making judgement calls, providing context and direction. And there's the coordination part – collecting updates, chasing actions, maintaining trackers, writing reports, managing handoffs.

For the last fifty years, we've called both of those things "management" and assigned them to the same people. That's ending. AI is absorbing the coordination work. Which means companies will need fewer middle managers – but the ones who remain will finally get to do actual leadership.

This isn't theoretical. It's happening now, with today's tools. Here's how smart companies are implementing it.

The unbundling of management

If you track what managers actually spend their time doing, you'll find that 70–80% of it is procedural coordination. Collecting status updates. Updating trackers. Pinging people for progress. Checking SLAs. Escalating delays. Writing reports that synthesise information from six different systems. Creating meeting agendas. Routing tickets. Following up on approvals.

None of that is leadership. It's essential operational work, but it's not strategic, not creative, and not particularly human. It's procedural. Which means it's automatable.

The leadership work – the actual value of having experienced humans in the middle – is the other 20%. Making nuanced judgement calls. Coaching people through difficult situations. Providing context that only comes from experience. Spotting patterns that metrics don't capture. Building relationships. Translating strategy into action.

AI can't do that work. But it can do almost everything else.

The economic reality is simple: if you can automate 70–80% of what middle managers do, you need fewer of them. The organisations that figure this out first will run leaner, faster, and with significantly lower coordination costs. The ones that don't will keep paying a tax that compounds every quarter.

What this looks like in practice

Let me be specific about how this works with today's tools, not theoretical future capabilities.

Start with one stable workflow

Pick a single workflow that happens repeatedly, has clear inputs and outputs, and involves multiple handoffs. Common examples: weekly status reporting, incident triage and routing, procurement approvals, pipeline reviews, or onboarding task coordination.

Don't pick your most important or visible workflow. Pick a boring one that nobody will miss if it runs slightly differently.

Build the AI shadow

This is the critical step that most companies skip. Before AI takes over a workflow, it shadows the human manager for 2–4 weeks.

Here's what that actually looks like with current tools:

Using Claude or ChatGPT as a shadow manager:

  • Set up a shared project or custom GPT with context about your team structure, regular workflows, and key data sources
  • Each day, feed it the information a manager would normally collect (Slack updates, ticket statuses, email threads, meeting notes)
  • Ask it to draft the outputs the manager would normally create (status update, action items, nudges to team members, report summaries)
  • The human manager reviews, edits, and sends

You're not automating anything yet. You're documenting the workflow by having AI attempt it under supervision.

Example prompt structure:

You are shadowing me as a manager for the DevOps team. Here's today's information:

- Slack channel updates from the past 24 hours: [paste]
- Current ticket board status: [paste]
- Outstanding action items from Monday's standup: [paste]

Based on this information:
1. Draft a brief status update for leadership
2. Identify any blockers that need escalation
3. Suggest who I should follow up with and about what
4. Draft those follow-up messages

Remember: You're shadowing, not deciding. I'll review everything before sending.

After two weeks, you'll have a clear picture of what the AI gets right consistently and where it needs human judgement.

Connect to your systems

Once the shadow phase proves the AI understands the workflow, connect it to your actual systems so it can read data directly rather than having you paste it in.

Today's practical options:

For Slack, Jira, Asana, Linear, GitHub: Use Make.com or Zapier to pull relevant data into a structured format. Set up a daily or weekly automation that:

  • Pulls ticket statuses, pull request activity, or project updates
  • Formats it into a structured summary
  • Sends it to Claude or ChatGPT via API
  • Receives back a draft status update, list of suggested actions, or risk flags

For custom workflows: Use Claude with computer use or ChatGPT with function calling to let the AI read from your systems directly. The new agentic capabilities (Claude's computer use, OpenAI's function calling) make this significantly easier than it was even six months ago.

Example architecture:

  1. Daily automation triggers at 9am
  2. Pulls data from Jira (current sprint status, overdue tickets)
  3. Pulls data from Slack (DevOps channel activity, incident reports)
  4. Sends structured data to Claude via API with instruction: "Analyse this data and identify: blockers requiring escalation, team members falling behind, incidents that need follow-up"
  5. Claude returns structured JSON with findings
  6. System posts summary to private channel for manager review
  7. Manager approves or edits
  8. Approved items get sent to team

Cost: roughly £3–5 per day in API calls for a 10-person team.

Move from shadow to operator

Once the AI is consistently accurate (usually after 2–4 weeks of shadowing), flip the switch. Let it start taking actions within defined boundaries.

Start small:

  • AI can post routine status updates automatically (with manager receiving a copy)
  • AI can update ticket statuses based on Slack confirmations
  • AI can send gentle nudges about overdue items
  • AI can triage incoming requests and assign them to the right person

Anything that requires judgement, discretion, or handling exceptions still goes to the human manager as an escalation.

The exception protocol

This is the most important design decision. You must explicitly define:

What AI can decide:

  • Routine status updates
  • Standard task assignments based on workload
  • Scheduled reminders
  • Updating trackers based on completed work
  • Closing tickets that meet "done" criteria

What requires human approval:

  • Anything involving budgets over £X
  • Personnel decisions
  • Customer-facing communications
  • Changes to project scope or deadlines
  • Escalations outside the team

What triggers immediate escalation:

  • Critical incidents
  • Team member repeatedly missing deadlines
  • Cross-team conflicts
  • Anything marked as urgent by stakeholders
  • Situations with no clear precedent

Build this as a simple decision tree. The AI checks every action against it before proceeding.

A worked example: weekly status reporting

Let me walk through a specific implementation I've seen work.

The old way:
Manager spends 3–4 hours every Friday collecting updates from five systems, synthesising them, writing a status report, formatting it, and sending it to three different stakeholders. The work is tedious but necessary.

The new way with AI:

Monday: Manager sets up a Make.com workflow:

  • Pulls ticket data from Jira (completed this week, in progress, blocked)
  • Pulls deployment data from GitHub Actions
  • Pulls incident data from PagerDuty
  • Pulls team capacity info from Forecast or Float
  • Formats everything as structured JSON

Tuesday–Thursday: Manager runs the workflow manually, reviews AI-generated draft, makes edits, learns what needs adjustment.

Friday: Manager switches workflow to automatic:

  • Workflow runs at 4pm every Friday
  • Sends data to Claude with this prompt: "Generate weekly status report for leadership. Include: key completions, current sprint status, blockers and risks, next week's priorities. Format for email. Keep it under 400 words. Flag anything that needs my attention before sending."
  • Claude generates draft report
  • Draft posts to private Slack channel
  • Manager reviews, edits if needed, approves
  • Approved version auto-sends to stakeholders

Time saved: 3 hours per week. Cost: £2 per week in Make + Claude API calls.

The manager now spends Friday afternoon doing actual leadership work – one-on-ones, strategic planning, or coaching – instead of compiling reports.

Another example: onboarding task coordination

The old way:
Manager manually tracks 47 onboarding tasks across IT, HR, and team-specific items for each new hire. Sends reminder emails. Chases people. Updates spreadsheet. Takes 5–7 hours per new hire spread over two weeks.

The new way:

Setup (one-time, 2 hours):

  • Create structured onboarding task list in Airtable or Notion
  • Set up Make.com automation that checks daily for overdue tasks
  • Connect Claude via API

Daily operation (automated):

  • System checks for tasks due today or overdue
  • Sends task list to Claude with prompt: "Review these onboarding tasks. Draft polite reminder emails for any overdue items. Flag any tasks blocked for more than 3 days for manager escalation."
  • Claude generates reminder emails, identifies blockers
  • Manager gets daily summary: "3 reminders sent automatically. 1 blocker flagged: IT equipment still not ordered for new hire starting Monday."
  • Manager only handles exceptions

Time saved: 4 hours per new hire. The coordination happens automatically. The manager intervenes only when something's actually wrong.

What this means for headcount

Let's be direct about the economics.

If a manager spends 25 hours per week on coordination and 5 hours on actual leadership, and AI can handle 20 of those coordination hours, then one manager can effectively cover 4x the scope. Or you need one-quarter as many managers to handle the same operational load.

This doesn't mean firing people immediately. It means:

  • Natural attrition isn't backfilled
  • Span of control increases
  • Managers are promoted based on leadership ability, not coordination skills
  • New managers aren't hired to handle coordination – they're hired when there's a genuine need for more leadership

The managers who remain will be different. They'll need strong coaching skills, good judgement, and the ability to provide context and make ambiguous decisions. Pure coordinators won't survive.

What goes wrong when companies rush this

Three common failure modes:

Skipping the shadow phase: Companies jump straight to automation without documenting what the AI needs to know. It makes bad decisions, breaks workflows, and everyone loses confidence. Recovery is hard.

Automating the wrong workflows: They automate high-stakes, visible processes instead of boring, repetitive ones. High-stakes workflows need human judgement. Boring workflows need AI. Get that backwards and you create risk instead of value.

No exception protocol: The AI keeps operating in situations where it should escalate, creating small disasters. Or it escalates everything, defeating the purpose. You need clear boundaries, defined in advance, before letting AI take actions.

The trick is starting small, shadowing first, and automating genuinely tedious work that nobody will miss.

Where this goes in 3–5 years

Right now, this requires deliberate effort to set up. You need to connect systems, define workflows, establish protocols. It's doable with today's tools, but it's not trivial.

Within 3–5 years, this becomes productised. Companies will offer "AI middle management layers" that plug into your existing systems – Slack, Jira, GitHub, your CRM – and handle standard coordination workflows out of the box. You'll configure them like you configure Zapier today, but they'll be agentic enough to handle complex multi-step workflows.

At that point, the companies that haven't figured out workflow automation will be running at half the efficiency of their competitors. The coordination tax will be obvious and measurable.

The takeaway

Management is unbundling. Coordination is becoming automated. Leadership is becoming the job.

This means fewer people in middle management roles, but the ones who remain will be doing fundamentally different – and more valuable – work. They'll be coaches, strategists, and decision-makers, not traffic controllers.

You can start implementing this today with current tools. Pick one boring workflow, shadow it with AI for two weeks, connect your systems, define the exception protocol, and flip the switch. Start small. Learn what works. Scale what proves valuable.

The businesses that figure this out first will run faster, leaner, and with a fraction of the coordination costs. The ones that don't will keep paying a tax that compounds every quarter.

This isn't an AI story. It's an organisational design story. And the window to get ahead of it is closing fast.