We are entering the era of AI agents in business, but most companies aren’t thinking clearly enough about what that actually means.
The AI era is not about sprinkling ChatGPT into your workflows. It’s not about automating your email responses and calling it a day.
It’s about documenting the way your business works at a fundamental level so that the robots that will work for you very soon understand "how things work around here".
If you're in leadership you need to be designing your company now for where work is headed—not where it’s been. We are in the early days and weeks of a technological revolution that will be the primary driver for a complete economic one. You need to start getting ready. Here's some thoughts I have on this after having read a few articles recently on the subject of deploying AI Agents (autonomous workers) in a company. I've tried to tailor this with how I see these things coming into play in the manufacturing or industrial distribution space.
Design Your Business for Agents and Humans
Satya Nadella recently said: "What lean did for manufacturing, AI will do for knowledge work—increasing value and reducing waste."
This is a quote from a man with a clear vision of the future. Further, it sounds a lot like a strategic directive to me. It's something that I believe wholeheartedly that we're staring at the genesis of every day we walk through the front doors to the office and put our bag down. Lean manufacturing revolutionized industrial output by eliminating waste, improving cycle time, and structuring work so value could flow unimpeded. Lean introduced disruptive positive change on shop floors and it ultimately changed how we thought about process design at a fundamental level.
Now AI is doing the same thing for knowledge work.
In the same way that lean forced companies to identify bottlenecks, standardize processes, and remove ambiguity, AI is thrusting organizations forward, pushing them to codify their knowledge, clarify decision-making logic, and reduce the friction that comes from people re-solving the same problem over and over again.
If you're leading a company through this next phase of transformation, that quote isn't a metaphor. It's a roadmap.
AI won't just automate busywork. It will reshape how teams are built, how decisions are made, and how fast ideas turn into actions.
But only if you build for it. If you aren't thinking about this yet, people will eventually become the bottleneck.
If you’re familiar with Design for Manufacturing (DFM), you know the principle: make it cheap, easy, and foolproof to build.
The same is now true for process design.
Design for AI means structuring your operations in a way that reduces ambiguity, context-switching, tribal knowledge, and reliance on hallway conversations. What works well for humans (ad hoc collaboration, soft context, social fluency) often breaks AI.
AI excels in a sea of context.
Today's models don't yet have eyes and ears and physical forms like we humans have. They are deprived of experiences that we people have day to day interacting with our co-workers, teams, and physical spaces. Think of an AI model today as a disabled coworker who lacks our 5 senses. They need all of the context about the world around them provided to them in written and/or structured form. That means documenting. On the plus side, what works well for AI (clear rules, codified knowledge, modular outputs) can make human teams a hell of a lot more productive, too.
The companies that win in this next era will treat AI not like a bolt-on tool—but like a first-class operator.
That means:
- Documenting processes the way you write code or quality procedures: clean, versioned, accessible
- Creating structured data and taxonomies and investing in that as soon as possible.
- Using clear inputs and expected outputs in every recurring task
- A long-term commitment from a personnel and financial perspective to build your business for the AI-powered future.
Put differently: the companies that will move fastest are the ones who stop relying on Teams chats, Outlook emails, archaic sharepoint sites, and shared intuition.
Write It Down; Write It All Down.
Your company is only as smart as what you have documented.
This has always been true, but it becomes mission-critical when you introduce AI agents into your workflows. If your business runs on tacit knowledge, side conversations, gut feel, inherited tribal norms you’re building your house on sand.
Humans are incredibly good at reading a room, inferring intention, or asking Jim in purchasing what "normal" looks like. AI can’t do that. And that’s why codified documentation is no longer a nice-to-have.
It’s now become your organizational infrastructure.
Natural-Language Programming for Your Business
Austin Vernon says it plainly: creating wikis, documentation, and SOPs isn’t busywork—it’s natural-language programming for your business.
Your agents don’t have intuition. They don’t guess. Every task you want them to perform needs a complete, unambiguous instruction set.
That means your org needs to:
- Document key workflows and repeatable tasks
- Define terminology, decision-making criteria, and escalation paths
- Clarify who does what, when, and why
This is where most companies struggle.
They assume "writing documentation" means some massive knowledge management system staffed by ex-consultants and corporate librarians. But if you use Microsoft 365, SharePoint, Teams, or Dynamics in the cloud—you already have what you need to start.
Pick a tool. Pick a starting point. Don’t overthink it.
What Belongs in Your Wiki?
The goal isn’t to write a perfect manual.
The goal is to create a usable operating layer that makes your human teams and your AI agents faster, clearer, and more accurate. Here are some examples of what to include:
1. Recurring Tasks
How do you onboard a customer? Submit a quote? Generate a part number? These are high-friction workflows that agents can own if they have access to the rules.
2. Reference Data
Glossaries, naming conventions, approved vendor lists, lead scoring rules—anything that removes ambiguity should be documented.
3. Decision Logic
When do we escalate a support ticket? How do we approve pricing exceptions? These criteria should live in writing, not in someone’s head.
4. Examples and Edge Cases
The fastest way to train your agents (and your team) is through real-world examples. Store the best proposals, most effective email sequences, and annotated call transcripts.
From Camp Games to Command Structures
MIT robotics has a challenge that they play with their students which asks them to explain how to make a peanut butter and jelly sandwich. The students might say:
- Take a slice of bread
- Put peanut butter on the slice
- Take a second slice of bread
- Put jelly on that slice
- Press the slices together
But if the teacher follows these instructions literally? MIT's lesson description says this might happen:
"would result in you taking a slice of bread, putting the jar of peanut butter on top of the slice, taking a second slice of bread, putting the jar of jelly on top of that slice, then picking up both slices of bread and pushing them together. After this, tell the students that their peanut butter and jelly sandwich doesn’t seem quite right and ask for a new set of instructions."
The overarching theme being introduced in this exercise is that computers do what they are told and nothing more. The ability to read between the lines and determine what was meant rather than what was said is a skill computers lack. Additionally, students are introduced to the concept of debugging their processes through iterative attempts to program a computer to make a peanut butter and jelly sandwich.
- Open the jar of peanut butter by twisting the lid counterclockwise
- Pick up the knife by the handle
- Insert the knife into the jar, scoop out a portion of peanut butter
- Withdraw the knife from the jar of peanut butter and run it across the slice of bread
- Spread it evenly across the top face of a single slice of bread, using a horizontal motion
- Take a second slice of bread and repeat steps 2-5 using Jelly in place of peanut butter
- Press the two slices of bread together such that the peanut butter and jelly meet
You get the idea.
The distance between what a human understands and what a machine can execute is called tacit knowledge. And AI doesn’t have it. That’s why you have to fill the gap with precision, documentation, and shared context.
Stop Using People like Google Search
Too many companies treat their coworkers like a database:
- "Go ask Marcy how to process that request."
- "Jim always knows what to do in those situations."
That works when your org is small. Or when it’s all humans. But, even then it's fundamentally flawed. What happens if Marcy quits tomorrow, or when Jim retires? The implicit system collapses, thats what. Jim & Marcy are great at what they do, and that expertise should belong to your company so that their excellence can be used to educate the next generation. Succession planning and AI deployment share a lot in common.
AI is fast. But only when the data and logic it needs is already written down.
Codifying your org’s operations is how you unlock AI’s speed and create a workplace where new hires can onboard faster, errors happen less, and ops become scalable.
It starts with documentation. And it scales with agents.
Take the Shackles Off: How to Trust Your AI Agents Without Losing Control
Letting AI agents operate autonomously shouldn’t feel like handing your toddler a chainsaw. But for most companies, it does. And it’s usually because they’ve skipped the most important step: training the agent not just to perform tasks—but to understand the rules of the business.
You don’t have to be afraid of what your agent might say in an email to a key account or publish in a blog post. Not if you’ve set it up right.
Bake in the Boundaries, Not Just the Capabilities
The fear that an agent might leak pricing, reveal secret IP, or make a wild public claim is justified—if you’ve given it no guardrails. But a properly trained agent isn’t operating on vibes. It’s executing on your business logic.
Autonomous agents can be incredibly powerful, but only when they understand:
- What’s public vs. private
- What requires review vs. what is pre-approved
- What tone, formatting, or phrasing your executive team would green-light
- What business rules constrain decisions around pricing, timelines, and approvals
You don’t need AI lawyers. You need context programming.
Replace Endless Reviews with Requirements + Surveillance
Traditional orgs rely on human checkpoints for a reason: they don’t trust people—or systems—to know when something is off the rails.
But here’s where AI gives us something new: the ability to unit test every output.
In software, unit testing is the process of validating that individual components of a system do exactly what they're supposed to—no more, no less. You define test cases, expected outcomes, and failure conditions. Then every code change gets evaluated automatically.
Apply this to AI agents, and it changes everything.
Before an agent sends an email, publishes a post, or hands off a pricing recommendation, a unit testing bot in its agentic workflow checks it against your:
- Tone and voice guidelines
- Regulatory requirements
- Style and formatting rules
- Product or pricing constraints
This isn’t just grammar checking. This agent is trained on your business rules to ensure compliance, quality control, and brand governance running in real time.
In an agentic operation, every output goes through these checks automatically. If it passes, it ships. If not, it gets flagged, logged, or escalated.
That means you don’t need a review committee or a final approver hovering over every single action.
You have agents for that.
But human review doesn’t scale. And worse: it stalls everything. In our agentic future, humans are our biggest bottleneck. Human intervention should only occur in pre-defined and well-designed scenarios. We'll cover that a bit more in the section on stop-work authority.
Here’s the better model:
1. Pre-approved standards
Codify what “good” looks like:
- Approved messaging frameworks
- Style guides
- Pricing bands
- Legal disclaimers
- Escalation criteria
2. Real-time surveillance
Set up your agents to:
- Cross-check output against these standards
- Measure confidence scores before publishing or sending
- Pause and flag anything that falls outside expected parameters
Now you can move fast, generate at scale, and still stay inside the lines.
Agents Are Good Soldiers, Not Rebels
It helps to remember: agents don’t improvise unless you train them to. They don’t decide what’s right. They evaluate based on the instructions and context you give them.
You’re not building another marketer or sales rep. You’re building a synthetic operator with a system prompt full of your values, rules, and redlines.
The cost of setting this up isn’t astronomical. In many environments (especially those already using tools like Microsoft Copilot or Dynamics 365), your system prompt can live in an orchestration layer or act as a reusable asset across agents.
Train once. Instruct often. Iterate quickly.
Think Like a Regulator (So Your Agent Doesn’t End Up in Trouble)
The smartest companies will think more like internal regulators when rolling out autonomous agents. That means:
- Pre-defining boundaries with clarity and redundancy
- Using whitelists and blacklists
- Adding logging and traceability to every decision
- Building “stop work authority” directly into agent architecture
Your agent shouldn’t feel like a freelancer with a GPT login. It should feel like a process you built, tested, and hardened—one that runs itself with a high degree of reliability, and notifies you when it doesn’t.
The Payoff: Confidence and Speed
The result of all this isn’t just safety. It’s velocity.
When agents are trusted to act, create, respond, and adapt—within your boundaries—your humans stop doing low-value supervision and start doing high-value strategy.
You don’t want your best talent approving subject lines. You want them designing markets.
When you build agent architecture with guardrails baked in, you get to scale insight, not just output, and most importantly, you finally get to take the shackles off your AI.
Give Your Agents Stop Work Authority
The Toyota Production System changed the world. Not because it introduced automation, but because it introduced autonomy.
One of its most powerful concepts that I learned of recently while getting my Six Sigma Green Belt with P4 Lean is that of "Stop Work Authority".
Anyone on the shop floor, from the most seasoned engineer to the newest hire, has the right (and responsibility) to pull the Andon cord when they spot an issue. It halts the entire production line.
This isn't done as punishment. It's done as protection.
That pause allows the team to swarm the problem, fix the root cause, and resume operations with higher confidence. Bill Greider at P4 Lean explained to me that this is how you bake quality into the process as it happens rather than inspect for defects later.
This isn’t just a principle. It’s an operating philosophy. One your AI agents need to adopt.
AI Confidence Thresholds are Your New Andon Cord
AI systems don’t get anxious. They don’t wonder if something feels "off." But they do run on probability.
Every output an agent generates is accompanied by confidence scores. This internal signal can tell you whether the agent knows what it’s doing—or whether it’s guessing.
When that score drops below a certain threshold? That’s your moment to pause.
Just like in Toyota's system, that pause isn’t a failure. It’s a trigger for:
Flagging uncertainty
Routing to a human for review
Logging the case for future learning
AI with Stop Work Authority doesn’t blindly produce. It knows when it needs help.
Build the Andon Cord Into Your Agent Framework
To make this real, you need to:
- Define confidence thresholds for critical tasks -- Anything that involves pricing, public output, or legal language deserves a tighter trigger.
- Create a response workflow -- Who gets notified? -- Where does the flagged output live? -- How is it resolved and closed?
- Log and resolve the root cause -- Was the issue caused by a lack of context? A formatting edge case? An outdated rule?
Fix it once, and every agent improves from it.
Why This Matters to Manufacturers
If your company is steeped in lean, this is going to sound familiar.
Just like a production issue gets fixed at the source, a content or decision error from an agent should trigger a process improvement—not just a patch.
And just like Toyota empowered its people to stop the line, we must empower our agents to halt a workflow, escalate uncertainty, and learn from it.
This isn’t a metaphor. It’s a one-to-one translation of lean principles into digital operations.
AI isn’t just your assistant. It’s your junior operator.
And it needs the authority to protect your work, your customers, and your brand.
Agents Are a Creativity Prosthetic
One of the most compelling use cases for agentic AI in marketing and sales is creative testing. A recent research paper, Agentic Multimodal AI for Hyper‑personalized Advertising, outlines how autonomous agents can generate and optimize B2B and B2C campaigns in real-time using persona-based retrieval-augmented generation (RAG).
This isn’t just for building ad copy. It opens the door to:
- Simulating entire buyer journeys
- Unit testing sales messages against target personas
- Running large-scale creative tests before a single dollar is spent on distribution
- Automatically adjusting cadence, tone, and offer style based on known behavioral patterns
Persona-based RAG gives your agents a way to think like your buyer before ever reaching them. You’re no longer just drafting messages—you’re pressure-testing them against a high-fidelity mental model.
It’s like having a focus group for every message you send, every campaign you run, and every outreach you launch.
Combined with surveillance-based QA and pre-approval standards, this tech won’t just accelerate your go-to-market strategy—it'll make it smarter.
There’s a toxic myth in a lot of manufacturing companies: "I’m not a creative."
Guess what? You are now.
AI agents unlock a new kind of creativity for ops leaders, engineers, and even frontline staff. They can:
- Explore design permutations and edge cases
- Simulate outcomes based on different constraints
- Help draft proposals, plans, or documentation for a half-formed idea
In the past, the cost of chasing a wild idea was too high. Too much coordination. Too much training. Too much time.
AI drops that cost to near zero.
Suddenly, the person with the idea can prototype it. Fast. Quietly. Cheaply.
This means your organization doesn’t just get smarter. It gets braver.
Agentic Sales & Marketing: The Next Frontier
As McKinsey notes in their article, Why Agents Are the Next Frontier of Generative AI, we're at the edge of a shift even more significant than the initial generative AI boom. Autonomous agents aren't just GPT with a wrapper. They are dynamic systems capable of executing multi-step processes, making decisions, and improving with feedback.
This matters in sales and marketing because agents can begin to own—and optimize—entire slices of your go-to-market function.
One insight McKinsey emphasizes is that agents introduce the possibility of delegation at scale. Sellers are no longer just using AI to draft emails. They’re handing off territory planning, account research, and opportunity scoring to autonomous systems. The sales role becomes more about judgment, relationship building, and trust—supported by real-time, always-on expertise.
In short: agents shift humans from operators to orchestrators.
This also means your data ecosystem, sales processes, and marketing architecture need to be ready. Your agents are only as good as the context you give them. That makes CRM hygiene, call intelligence, campaign analytics, and knowledge base development non-negotiables.
It also changes the economics of experimentation. With agents, you can spin up new campaigns, iterate outreach sequences, and test messaging variations without tying up your team.
The companies that structure their systems, data, and orgs to take advantage of this now will move faster than their peers because their learning cycles will collapse from months to hours.
Sales is one of the oldest professions, but it's entering a brand-new phase. As McKinsey puts it, the seller’s core role—building trust, creating value, reducing friction—hasn’t changed. But the tools to deliver that experience are now radically different.
Generative AI, combined with agentic design patterns, enables sales and marketing teams to scale their insight, responsiveness, and planning to levels previously impossible. McKinsey estimates this shift could unlock $0.8 to $1.2 trillion in productivity. That’s wild.
Your agents could be running market simulations for your sales leaders and providing dynamic strategies based on their insights.
Here’s what that means:
- Insight extraction is continuous. Agents aren’t waiting for someone to analyze performance data. They’re reading it, clustering it, pattern-matching it, and flagging trends the moment they emerge.
- Campaigns become adaptive. An agent can monitor a message’s click-through rate across verticals, then rewrite the subject line for aerospace while leaving the medtech version untouched. No marketer in the loop. No weekly review.
- Sales motions are pre-tested. Using persona-based retrieval-augmented generation (RAG), agents can simulate outreach messages against a library of real buyer conversations and successful closes. This means you're unit testing sales copy like a programmer might unit test code. If the message triggers a failure (low fit, tone mismatch, misalignment), it gets caught before it ships.
- Failure modes get surfaced early. Adversarial agents which are designed to break assumptions and flag gaps, can pressure test strategy and outputs before they hit the market. No more waiting weeks for a campaign to flop in the wild. You can crash test it up front.
This is your NFL coaching staff for B2B.
But here’s the hard truth: you can’t layer this on top of a broken system. You need to reengineer some core business functions first.
1. CRM Becomes Your Single Source of Truth
Your CRM needs to move from optional to mission-critical. That means:
- Required fields aligned to pipeline stages
- Clean, structured opportunity data
- Tight integration with conversation intelligence and marketing automation
If it's not in the CRM, your agents can't use it. Period.
2. Pipeline Becomes a Shared Language
Sellers, marketers, and agents need to operate on a common sales architecture. That means:
- Clearly defined funnel stages
- Pre-approved exit criteria for every stage
- Automatic progression or alerts based on activity signals
The better your pipeline logic, the more reliably agents can run plays.
3. Forecasting Goes from Backward to Forward
AI agents don’t just summarize the past. They simulate the future.
With enough historical data, they can:
- Predict which accounts are warming up
- Forecast close probability based on behavior, not gut feel
- Suggest actions to pull forward stalled deals
This turns sales management from reactive into strategic.
4. Knowledge Becomes an Operating Asset
You must:
- Record sales calls
- Transcribe and tag interactions
- Build playbooks and competitor matrices in shared wikis
- Feed that content back into agent prompts
Think of every meeting, deck, objection, and win as training data.
The New Build Order
Want to go all-in? Here's how the best AI-native companies will build from the ground up:
- Every workflow is documented and written in natural language.
- Agents are trained on those documents and can access them as context.
- Surveillance systems catch errors and feed improvements back into the loop.
- Human operators stand by to handle exceptions and retrain agents.
- Creativity becomes a team sport.
That’s not five years out. That’s available right now.
The only thing stopping companies from doing this is the same thing that holds them back from everything else:
Organizational inertia.
If you want to move faster, out-innovate, and ship with confidence, you need to start building the infrastructure today.
Not just to work with AI.
To work at the speed of it.
