Most people think education ends when work begins. That belief is becoming expensive, especially in an age where intelligence is outsourced and commoditized.
Skills now decay faster than careers, and innovation is speeding that up. That’s why one answer caught my attention. In an interview with TBPN, Mark Cuban was asked about the next big job opportunity for students. His answer wasn’t “learn to code” or “get a data science degree.” It was this:

“Companies don’t understand how to implement AI right now to get a competitive advantage… learn to customize a model, walk into a company, show the benefits. That is every single job that’s going to be available for kids coming out of school.”

He’s right. Most companies don’t know how to implement AI, especially small businesses. But I think he’s underselling the problem.

I replied on X:

One major issue that will be amplified with AI is delegation versus control, especially micromanagement. If a manager already struggles to delegate tasks to humans, it is hard to imagine how that will work with a team of AI agents.

Learning to trust your team will be important for long-term relevance and competitiveness.

Most companies still don’t understand why they are not AI-ready. They have many human inefficiencies and fail to document their processes. Implementing AI is not about buying a SaaS tool or a system you simply plug in and expect magic.

This article unpacks why that happens, and what to do about it.

A quick note before the prompt I’m sharing

The long prompt link below is a starter version, meant to show the method, not the full playbook. Real implementations require more depth because some principles, like “correct” changes by workflow, risk level, data sources, and who signs off.
In practice, each step of the method has its own dedicated prompt and checks, tailored to the company context. This example is intentionally generic so anyone can try it, learn the logic, and see where their system breaks, unclear ownership, messy inputs, approval loops, or no definition of correct.
If you want a sector-specific version and the full set of prompts, including how to set risk tiers, build a small eval set, and design monitoring, reach out (my contact at the end of this post), and I’ll adapt it to your workflow.

Universal AI Readiness Audit Prompt

The delegation problem nobody wants to name

If a manager already struggles to delegate tasks to humans, it is hard to imagine how that will work with a team of AI agents.

Read that again!

This is not a tech problem. This is a team spirit and a leadership problem. And it’s the same problem Jocko Willink and his team at Echelon Front have been diagnosing in organizations for years, long before anyone was talking about AI agents.

I highly recommend reading his books or listening to his podcast every Wednesday.

Willink’s core thesis is simple: leadership is about trust, ownership, and decentralized command. You can’t scale a team if every decision has to route through you; that becomes a bottleneck. The leader’s job is to set clear intent, train the team, and then get out of the way so they can execute and come up with solutions, with full trust.

Most managers fail at this with humans. They say they want autonomy, but they design systems that require constant check-ins. They claim to trust their team, but they review every output before it ships.
With a simple fact…Do what I say, but not what I do.

Now give that same manager an AI agent.
What do you think happens?

The same patterns emerge, just faster and more visible.
The agent produces output.
The manager reviews it.
The agent waits.
The manager second guesses.
The agent stands by.
The cycle repeats…

It becomes a loop, an illusion of work, or (un)intentional fake work disguised as progress, where nothing actually moves forward in the big picture.

Just like Cuban salsa (dance), 3-steps forward, 2-steps back, explained in my previous post about finding the right rhythm for innovation without losing our humanity.

As Einstein said, Doing the same things over and over again while expecting different results is the definition of insanity!

Why does everything slow down after the shiny AI demo video you saw?

On Monday, your team is excited. You deploy an agent to handle a workflow everyone hates: sorting out support tickets, creating campaign briefs, cleaning CRM notes,…

By Wednesday, the agent is “working” in the narrow sense. It produces output.

By Friday, the team feels slower.

But wait, it’s almost the weekend…What happened?
The agent produces more drafts, but delivery slows down. Reviews pile up, managers add extra approvals, and people bypass the official tool with private workarounds using their consumer AI licenses with Gemini or ChatGPT just to meet deadlines…Hello! Sensitive info leaks into web searches on Google.

If this sounds familiar, the AI model is not the main problem. The control design is.

A July 2025 study from METR, a research organization focused on AI evaluation, found something counterintuitive: experienced open source developers using AI tools took 19% longer to complete tasks compared to working without AI. Even more striking, developers expected AI to speed them up by 24%, and even after experiencing the slowdown, they still believed AI had helped. (METR, July 2025)

The study found that review bottlenecks consumed the gains. AI generated more output, but verifying that output against repository standards, documentation requirements, and testing coverage took more time than the AI saved.

An industry observation shared by Enaiblement put it bluntly: “For every 1 hour AI saves, organizations spend 2 hours verifying it.” That’s not a universal truth, but it’s a recognizable pattern for anyone who has rolled out automation into a team that wasn’t ready. (Enaiblement, LinkedIn, September 2025)

This is not an edge case

Before we build frameworks, let’s establish the scale.
MIT’s NANDA State of AI in Business report (2025) found that “95% of organizations are getting zero return… Just 5%… extracting millions in value.” The success cases aren’t running better models. They’re running better workflows.

According to Emergent Mind’s December 2025 survey, “68% of production agents execute at most 10 steps before requiring human intervention.” These agents aren’t failing because they’re incapable. They’re slowing down because they’re embedded in control architectures designed for manual work.

And when official tools feel too slow, employees route around them. KPMG’s survey of 48,000 workers found “57% say they have hidden their use of AI or presented AI-generated work as their own.”

These numbers describe a system problem, not a technology problem.

Where the human bottleneck lives

A diagnostic lens, not a black magic voodoo framework

For an agent to work, it needs 3 things:
A clear state, a clear target, and a safe, testable path between them.

If any of these is missing, your agent becomes an advisor, helpful, sometimes impressive smart or sounds correct, but stuck!
But the triad is a diagnostic, not a full solution. Each element requires buildable parts.

Think of it like moving money. A bank transfer works not because the cashier behind the desk is smart, but because there are rails, limits, checks, and reversals.
The transfer is reversible within a window. There’s an audit trail. There are transaction limits and rules. The risk of any single action is contained. That’s what a safe, testable path means in practice.

That’s why intelligence is not the limiting factor. Infrastructure is.
AI Agents don’t fix leadership debt. They surface it.

The missing artifact: the Quality Agreement

“Clear target” sounds simple until you try to define it.
Most teams skip this step. They say “we’ll know it when we see it” or “the reviewer will catch problems.” Then week 1, a plausible answer feels like success. Week 3, someone complains the numbers don’t match finance. Week 6, legal says the tone is wrong. Those aren’t tweaks. Those are different products.

Nate B. Jones, an expert in AI strategy and product development, uses a 7-step sequence: Define, Decide, Translate, Specify, Anticipate, Measure, Operate.
It’s a practical checklist for moving from “we want AI” to “AI is running reliably.”

After searching on the web, 2 other frameworks cover similar ground:
NIST AI Risk Management Framework (AI RMF) is a U.S. government standard for managing AI risk. It’s designed for organizations that need to demonstrate responsible AI use to regulators, auditors, or enterprise clients. 4 functions: Govern (cross-cutting policies), Map (understand context and stakeholders), Measure (assess risks and performance), Manage (respond and monitor). If you’re selling to government, healthcare, or finance, or expect to, this is the compliance language your buyers speak.

MLOps Lifecycle is the engineering playbook for deploying machine learning in production. It covers problem definition, data preparation, model training, deployment, and monitoring. If you’re building custom models or working with data science teams, this is their workflow. For most SMBs using off-the-shelf AI tools, you won’t need the full MLOps stack, but the “monitor after launch” principle still applies.

You don’t need to memorize these frameworks. What matters is the pattern they share:

  1. Define what you’re solving and what “correct” means
  2. Build with guardrails and checkpoints
  3. Measure before and after
  4. Monitor in production, don’t assume launch day quality lasts

Quality is not a model property. It’s a team agreement. Without a written agreement for what “correct” means, you get moving goalposts, then blame the model for being unreliable. The model didn’t change. Your definition did.
A Quality Agreement answers:

  1. Who uses this output, and what decision does it drive?
  2. What must be true? (3-5 rules)
  3. What must never happen? (3-5 rules)
  4. What uncertainty is allowed? (When can the system say “I don’t know”?)
  5. Examples: 2-3 correct outputs, 1-2 unacceptable outputs, in the real format
  6. Pass/fail checks: Thresholds when possible, human judgment criteria when not
  7. Escalation rule: When does this route to a human, and who is that human?

If your team can’t agree on these 7 items, you don’t have a quality problem. You have a requirements problem. And no model will fix that.

This is why pilots stall. Not because AI is unreliable, but because teams resist defining quality, because defining it forces tradeoffs they’d rather avoid.

The double check loop

Most teams argue about control in moral terms. “People don’t trust the AI.” “Leadership is micromanaging.” Those arguments create heat, not progress.

The real issue is that reviews grow when reviewers are forced to “judge” instead of “verify.” Without pre-written pass/fail checks and representative examples, every review becomes a negotiation. That’s exhausting. So people add more reviewers to spread the blame. You want measurements, not opinions.

Micromanagement Metrics Dashboard

A May 2025 study showed that companies save about 25 percent in time and money by not checking every AI result in the same way. Outputs that could cause problems are reviewed by experts, while normal ones are checked only sometimes.
The results stay just as good while maintaining quality. (WJAETS, May 2025)

This is factory QA (Quality Assurance) versus QC (Quality Control) logic in the Manufacturing industry. You don’t inspect every (sub)step of the process with a microscope. You set tolerances, sample, and escalate exceptions.
The same principle applies to AI output.

Safety that scales without paralysis

The simplest control design that scales: 2 lanes.
Fast lane for low-risk work.
Slow lane for high-risk work.

This is how airports handle passengers. Not everyone goes through the same process. They separate risk, then enforce the right checks.

2 practical rules

First, the fast lane must be easier than the shadow path. If your official tool takes 4 clicks and ChatGPT takes 1, you’ve already lost.

Second, the slow lane must be specific, not vague. If the slow lane is vague, it grows. Then everything becomes slow lane.

Safety scales when quality is explicit, and risk is proportional. That’s why monitoring exists: tests miss real world edge cases, so operating needs ongoing checks, not just pre-launch gates.

A December 2025 research paper from Google tested this idea directly in multi-agent systems. Researchers ran 180 controlled experiments across 5 coordination architectures and 3 major model families. They found that independent agents, those operating without shared checkpoints, amplified errors roughly 17 times faster than a single agent baseline. Adding a coordinator role with validation gates reduced that amplification to about 4 times the baseline. Still imperfect, but a major improvement. (Kim et al., arXiv:2512.08296, December 2025)

In plain language, here is the nuance in centralized coordination versus decentralized command (remember Jocko): centralized coordination is having one shared point, human or automated, that routes work and catches mistakes before they spread.
It’s not about hierarchy or control. It’s about where you place checkpoints.

A 14-day pilot that produces signal

If you want results, don’t start with 10 workflows. Start with one.

Pilot deliverables:

  • One page that clearly defines what “finished and acceptable work” looks like
  • A small set of examples, around 20 to 50, that show what counts as good or bad output
  • A simple repeatable check you can run again after changing prompts or workflows

After 14 days, look at two things:
First, how often work needed to be checked by a person before it could move forward, compared to before the pilot

Second, whether the amount of work that had to be fixed later stayed about the same

How to read the result:

  • If the amount of checking the tasks workflows goes down and the amount of fixing later stays the same, the problem was how control was designed.
  • If checking goes down but fixing later goes up, the rules for what gets checked are wrong.
  • If nothing changes, the problem is probably missing data, unclear state, or poor tool access.

Limitations

Some work should stay slow. Legal matters, safety-critical decisions, and irreversible financial actions require strong controls.

In Europe, and likely soon in the USA, the EU AI Act requires human oversight for high-risk systems. The goal is not speed everywhere.
The goal is risk-proportional control.

We will likely see significant liability cases, with law firms and insurance companies expanding rapidly to cover damages caused by AI agents.

Artificial Intelligence is also creating jobs, right?

Surveys can mislead. Self-reported data can exaggerate or hide reality.
The 57% “hiding AI use” figure is directional; validate in your own context.

Control is not always the bottleneck. Data quality, integration debt, model capability, and workflow fit can be the real blockers. Use the 5 metric dashboard to separate control bottlenecks from integration bottlenecks.

Quality definitions drift. What counts as “correct” in week 1 may not match week 12. Build a review cadence, monthly at minimum, to check whether your Quality Agreement still matches reality.

If this felt familiar, it should

Human-to-human delegation problems get copied into human-to-agent delegation, but amplified by speed. The manager who couldn’t let go of approvals for their human team won’t let go of approvals for their AI agent either. The organization that never documented its processes will find that AI has nothing clean to amplify. The workflow that required 3 sign offs before won’t suddenly require fewer just because the work is automated.

Prescription #1:
Shift from 100% Review to Risk-Based Control

The most powerful redesign is to make the level of human oversight proportional to the risk of the task. Instead of treating every Al output as a high-stakes decision, create “fast lanes” for low-risk, easily reversible actions and reserve intensive review for what truly matters.

Prescription #2: Start at the “Boring” Edges, Not the Core

The instinct is to automate the most complex, high-value part of a workflow.
This often fails due to hidden complexity and a lack of organizational trust.
A more successful strategy is to automate the “edges” first-the repetitive, mechanical tasks that surround the core expert work. This builds momentum, demonstrates value, and earns the trust needed to eventually approach the core.

If your organization struggles to delegate to humans, it will struggle to delegate to agents. Fix the human system first.

AI agents are not the bottleneck. The bottleneck is the human system around them: trust, context, decisions, and the willingness to define what “quality” actually means.

AI agents stall when teams refuse to define quality, because defining it forces tradeoffs. The teams that succeed are the ones who have that conversation early, write it down, and treat it as a living document, not a one-time exercise.

Sources

  1. METR Study, July 2025: Experienced developers 19% slower with AI tools https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
  2. Enaiblement observation, LinkedIn, September 2025: “2 hours verifying for every 1 hour saved”https://www.linkedin.com/posts/enaiblement_ai-productivity-business-activity-7368731307050196994-4yu5
  3. Kim et al., “Towards a Science of Scaling Agent Systems,” arXiv:2512.08296, December 2025 https://arxiv.org/abs/2512.08296
  4. WJAETS Paper, May 2025: Human in the Loop LLMOps, tiered review reduces compliance costs 25% https://journalwjaets.com/sites/default/files/fulltext_pdf/WJAETS-2025-0643.pdf
  5. MIT NANDA State of AI in Business 2025 https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf
  6. Emergent Mind, December 2025: 68% of agents require human intervention within 10 steps https://www.emergentmind.com/papers/2512.04123
  7. KPMG Shadow AI Survey, 2025: 57% hide AI use (n=48,000) https://kpmg.com/kpmg-us/content/dam/kpmg/pdf/2025/shadow-ai-already-here-take-control-reduce-risk-unleash-innovation.pdf
  8. IAIS, “Global Insurance Market Report 2025”, December 2025. That jump in AI risk disclosure (4% → 43%) shows regulators and companies are treating oversight and liability risk seriously, which is multi-billion dollars today with +30%+ growth rates and forecast expansions into double-digit billions by 2033.
    https://www.iais.org/uploads/2025/12/Global-Insurance-Market-Report-2025.pdf

Author: El Capitano - Otman

AI Educator & Automation Strategy Advisor |
Teaching teams to avoid AI slop & over-automation