In October 2024, Gartner issued a prediction that quietly circulated through boardrooms: by 2026, 20% of organizations will use AI to flatten their structure, eliminating more than half of today's middle-management positions. It isn't a forecast about layoffs. It's a forecast about which tasks change.
In April 2025, Microsoft described the emerging role in its Work Trend Index: the "agent boss" — someone who builds, delegates to and manages AI agents to amplify their impact. This isn't science fiction. 71% of workers at organizations following this model say their company is thriving, compared with just 37% globally. The number that matters here isn't the share that's thriving. It's the fact that 71% of people surrounded by AI agents didn't say "I'm losing autonomy." They said "we're thriving."
The angle is clear: AI doesn't eliminate the manager. It eliminates the manager who manually assembles reports and passes along requests. What survives — and gains strategic weight — is the manager who reviews, approves, provides context, sets priorities and answers for what the team and its agents deliver. That only works with a process behind it.
The middle-layer work that disappears
Much of what a middle manager does today is moving information around: reading a report, pulling out the number that matters, passing it up with a note for the CEO, bringing the answer back down, translating it for whoever produced it in the first place. Each of these layers takes time, and the information changes, loses context or gets lost between levels.
AI moves that information in seconds, without losing any of it. An agent connected to the sales system can read the quarter's numbers, cross-reference the customer base, pull the complaint history for that period and hand the executive a ready summary: which customers complained most, about what, and whether they renewed anyway. What used to take a week of meetings between analyst, coordinator and manager becomes a structured answer, ready overnight.
The consequence is twofold. First, structures genuinely flatten: when information arrives at the top already assembled, you need fewer people piling up data in the middle. Second, the manager who did that moving-around work loses their reason for being. But the manager who knows what to ask, thinks in context and says "this number is wrong because it doesn't account for the July holiday" gains relevance.
Why most companies get this change wrong
Three standard traps:
Fully unrestricted agents, with no trail. You connect an AI model to every system, let anyone ask it for whatever they want, and assume "it'll work out." Three weeks later, an agent gets a critical number wrong in a contract because it used an outdated file, and nobody knows why. Who approves this? Who pays if it goes wrong? The manager is there to review, but review what, if there's no record of how the answer was produced?
Everything goes through a human, so nothing changes. The opposite reaction: require a person to review every single response before any action, because "security." Result: you keep the manager's old workload, add the responsibility of reviewer on top, and the agents sit idle waiting for approval. That isn't flattening — it's slowing down. The manager still spends an hour-long meeting approving three documents when they could be thinking about strategy.
Data and credentials scattered everywhere. An agent taps into the HR system because it needs one number, but that integration exposes everyone's salary. Another agent builds a customer report using a generic account because "it was easier to set up." Now anyone chatting with these agents can ask for a number they shouldn't see. The manager who's supposed to review can't, because they don't even know the gap exists.
Most companies fall into one of these traps because they're the most obvious ones. But neither solves the real problem, which is simpler than it looks: who authorizes what, in which context, when and why.
What has to be in place
This checklist isn't for the anxious manager — it's for the one who wants things to actually work:
Corporate identity tied to the company's Active Directory. When someone is let go from the directory, they lose access to every AI agent along with it — with no dependence on someone remembering. Someone joining the company inherits the permissions of their role in the directory; there's no need to register them in a separate system. This protects two things: you know exactly who asked for what, because identity is verified, and you don't carry around a parallel list of accounts that turns into a mess.
Access matching each person's role. A sales agent connected to your CRM can't query recruiting's calendar just because it's sitting right there. Someone on the support team can see a customer ticket, but not the salary of whoever closed the sale. Each role sees only what it's authorized to see. Change roles, and access changes with it.
Automatic approval for low risk, human approval for high risk. An agent that drafts a standard reply to a customer sends it on its own — but it lets the manager know it did. An agent that wants to schedule a meeting with a client pauses the conversation and asks for confirmation: "I'll book it for Thursday, okay?" An agent about to answer something involving pricing policy, which could turn into a contractual promise, waits for approval from an authorized person, and only acts afterward.
Real-time audit trail. Every action is logged: which agent did what, with which information, who approved it, when. When an audit or incident investigation comes along, you don't reconstruct it from memory. You open the system and show it, step by step: who authorized what, and when.
Reviewed knowledge before it becomes an official answer. New documents, updated policies, changing processes — none of it becomes "knowledge the agent uses" for free. A person, or a pair of people, review it: is this version correct? What scope does it apply to? Who owns it? Only then does it enter the base the agents consult. Critical documents can require two signatures — whoever writes doesn't approve, whoever approves didn't write.
That's how Skyller was designed: identity coming from the company directory, role-based access, risk-based approval and a trail that stays.
The gain that actually matters to managers
When these layers work, what changes is the manager's time. They don't leave meetings behind. They change what the meetings are about.
Instead of spending 4 hours a week in an "operations alignment meeting" — where someone reads a number, someone asks a question, someone translates it into context — the manager spends 1 hour actually thinking: "the numbers moved this way; what do we do now?"
The company benefits too, because the information reaching the executive comes with full context. It isn't an isolated number that can be misread. It's: "we sold this much; the cost was this; the complaint was about this; the renewal happened anyway." The manager coordinating all of that doesn't disappear. They become the person who makes better decisions because they have the complete picture.
There's a budget effect too. Without clear governance, every manager requests "their own" AI tool: some sign up for ChatGPT Pro, others for Copilot, someone else tries an open-source option nobody's heard of. Spending is scattered and invisible. With governance, credits are shared across the team: whoever needs more uses more, whoever needs less uses less, and the director sees consumption by area every month. A shared, transparent budget takes the politics out of who deserved more.
Three questions to take into your next meeting
Before drafting another security policy or calling in a consultant, it's worth answering these three questions with IT and with whoever runs operations:
-
If an employee is let go today, how many AI tools can they still get into tomorrow? If the answer is "depends on someone remembering" or "we deactivate it in two weeks," the problem isn't policy — it's identity. There's no security in "we hope it works."
-
When an agent delivers an answer about something critical — a contract, a price, a regulation — where did that information come from, and who authorized it to be used that way? Without a review step beforehand, any outdated document can become the official answer. Without "four eyes," you don't know whether the agent got it right or just got lucky on that particular question.
-
What did you learn about how AI is actually used here last quarter? If the answer is "nothing, that would be nice to know," usage isn't lower. It's just invisible — and invisible means there can be no real governance.
These three questions measure whether your organization is ready for the manager to shift from "task executor" to "decision reviewer." If the answers are "yes, we have that," the manager's job is already different, and the AI agent is just the tool that makes it visible. If they're "no," the change isn't about technology. It's about process.
