The Conversation You're Probably Having With Yourself
If you're a business owner right now, you've probably had this moment:
Someone tells you that you need to "use AI." Maybe it's an employee, a peer, an article, a vendor. And you think: I know I probably should. But every time I look into it, I have no idea where to start. And honestly? I'm not sure I trust it.
If that's you, you're not alone.
Some owners are afraid—worried AI will make expensive mistakes with their customers or replace the expertise they've spent decades building. Some are skeptical—they tried ChatGPT once, asked it a question about their business, and got back something generic and useless. Some are overwhelmed—there are a thousand AI tools, each promising to revolutionize everything, and they don't have time to sort through the noise.
Here's what I want you to understand:
AI isn't a strategy. It's a capability. And like any capability, it's only useful if you know where it fits—and where it doesn't.
The opportunity most business owners are missing isn't "use AI more." It's understanding that AI is exceptionally good at specific kinds of work—and terrible at others. Once you see that distinction clearly, you can stop chasing hype and start making real decisions about where AI can actually help your business.
The Real Question Isn't "Should I Use AI?"
The real question is: For this specific piece of work, what should be doing it?
Not "can AI do this?" but "should AI do this?" And if so, how much of it?
This is where most of the AI conversation goes wrong. People talk about AI like it's all-or-nothing. Either you're fully automated or you're doing everything by hand. Either you trust it completely or you don't use it at all.
But that's not how work actually operates.
Think about a spectrum:
The Execution Spectrum
↑ High judgment required
↑ Relationships matter
↑ Novel situations
↑ Trust is the product
Zero judgment needed ↑
Same input, same output ↑
High volume, low stakes ↑
Pure rules ↑
On the left: work that requires human judgment, relationships, and craft. The customer who calls upset and needs to be heard. The estimate that requires reading between the lines. The decision about whether to fire a subcontractor.
On the right: work that's purely mechanical. Send a confirmation email when an appointment is booked. Calculate the total on an invoice. Update a status field when a job is marked complete.
Most of your work isn't at the extremes. It's somewhere in the middle.
And that middle is where AI gets interesting.
A decade ago, "the middle" meant spreadsheets and checklists. Today, it means AI can draft the email while you decide whether to send it. AI can summarize the customer's history while you decide how to approach the conversation. AI can flag the anomaly while you decide what to do about it.
What AI Is Actually Good At
Let's be precise. Because the hype makes AI sound like it can do everything, and the fear makes it sound like it can do nothing useful.
Here's what AI genuinely excels at today:
Turning Fuzzy Into Structured
Reading unstructured input—an email, a voicemail transcript, a chat message—and extracting the important pieces. What does this customer need? What category does this request fall into? What's the priority?
Summarization and Synthesis
Condensing a 20-page document into key points. Summarizing a 45-minute call into action items. Taking information scattered across multiple places and pulling it into a coherent brief.
Drafting and Generation
Producing a first pass at something that follows a pattern. A follow-up email. A proposal based on a template. A status update. A response that matches your tone and style.
Pattern Recognition
Spotting anomalies in data. Flagging items that match certain criteria. Noticing trends across large volumes of information. Finding signals humans would miss because there's simply too much to look at.
Notice what these have in common: they're cognitive grunt work. Tasks that require intelligence but not necessarily judgment. Tasks that eat up time but don't require the thing that makes you valuable—your expertise, your relationships, your ability to make decisions in ambiguous situations.
Where AI Fails (And Why It Matters)
Here's the part the AI vendors don't emphasize:
AI is not deterministic.
Give the same AI the same question twice, and you might get different answers. Not slightly different—sometimes meaningfully different. This is the fundamental tension: you want predictable, repeatable outcomes in your business, but AI as it exists today doesn't guarantee them.
This matters because:
→ You can't fully rely on AI for consistency
→ You can't promise customers that "the AI will always do X"
→ You can't audit AI decisions the same way you audit rules
Beyond the consistency problem, AI is also genuinely bad at:
Novel Judgment
Situations it hasn't seen patterns for. Trade-offs without clear right answers. Contexts requiring deep domain expertise.
Relationships
Reading unspoken cues. Building trust over time. Navigating interpersonal complexity. Knowing when a customer needs to vent versus needs a solution.
Accountability
AI can recommend; it can't be responsible. High-stakes decisions need human ownership. When something goes wrong, "the AI did it" isn't an acceptable answer.
Accuracy (Sometimes)
AI will be wrong sometimes—confidently wrong. The question isn't "will it be perfect?" but "what's the cost when it's wrong?"
This is why "let's just have AI handle it" is almost always the wrong framing. The right framing is: "For this piece of work, what combination of human and AI makes sense?"
Six Ways Work Gets Done
In my book Deliberate Work, I describe six "modes of execution"—different ways that any piece of work might get done. Understanding these modes is the key to figuring out where AI fits in your business.
Here's a simplified view:
| Mode | Who Does It | When to Use |
|---|---|---|
| Human | A person, fully | Relationships, craft, novel situations, high-stakes judgment |
| Guided Human | A person with system assistance | Repeatable work that needs consistency across people |
| AI-Assisted | AI produces, human reviews | AI can draft; human judgment still needed |
| AI Agent | AI executes a sequence | Multi-step work with bounded decisions; human supervises |
| Autonomous | System, no human touch | Pure rules; same input always produces same output |
| External | Someone outside your org | Customer, vendor, or third party does the work |
The magic happens in modes 3 and 4—AI-Assisted and AI Agent. These are where AI handles the cognitive grunt work so your people can focus on judgment, relationships, and craft.
Let me show you what this looks like in practice.
Four Patterns That Actually Work
These aren't hypothetical. They're patterns I've seen work for plumbers, electricians, accountants, consultants, and lawyers. Small teams. Real constraints. Actual results.
Pattern 1: Intake and Classification
The problem: New requests come in through multiple channels—calls, emails, texts, website forms. Someone has to read each one, figure out what it is, and route it to the right place. It's tedious, it's error-prone, and it eats up time your office manager could spend on higher-value work.
The AI pattern: AI reads the incoming request, extracts the key information (what service, how urgent, new customer or existing), and proposes a classification. A human confirms or corrects.
Example: HVAC Company
Before AI:
Office manager reads every email and voicemail transcript. Manually enters data into the system. Decides if it's urgent. Routes to scheduler or tech.
With AI:
AI reads the message, extracts customer name, service type, urgency signals ("no AC" + "elderly parent" = urgent). Proposes category. Manager confirms with one click.
The human role shifts: From "read and categorize every request" to "confirm AI's work and handle the edge cases."
Pattern 2: Context Before Conversations
The problem: Before an important call or meeting, you need to know the customer's history. What jobs have you done for them? Any issues? When did they last call? What did they complain about? This information exists—scattered across your CRM, email, invoices, notes. Gathering it takes 15 minutes you don't have.
The AI pattern: AI pulls the relevant information and synthesizes it into a one-page brief. You review it in 2 minutes and walk into the conversation prepared.
Example: Accounting Firm
Before AI:
Partner opens three tabs—CRM, email search, last year's return. Scrolls. Scans. Tries to remember. Walks into annual review half-prepared.
With AI:
AI generates brief: income changes, new deductions mentioned in emails, questions from last year, current engagement status. Partner reads it on the walk to the conference room.
The human role shifts: From "gather context" to "receive synthesized context and decide how to use it."
Pattern 3: Drafting Follow-Ups
The problem: After every job, someone needs to send a follow-up. After every estimate, someone needs to send a thank-you. After every call, someone needs to send a summary. These emails follow patterns, but writing them takes time—and if they don't go out, customers feel forgotten.
The AI pattern: AI drafts the communication based on what happened. Human reviews, personalizes if needed, and sends.
Example: Electrical Contractor
Before AI:
Tech finishes job at 4pm. Follow-up email should go out same day. Tech is already at next job. Email goes out three days later—or never.
With AI:
Job marked complete triggers AI to draft follow-up: thanks for the business, summary of work performed, reminder about the warranty, invitation to leave a review. Office manager reviews and sends—5 minutes instead of never.
The human role shifts: From "write every email from scratch" to "review and personalize drafts."
Pattern 4: Document Transformation
The problem: Creating proposals, reports, or documentation requires taking information from one format and turning it into another. Notes from a site visit become a formal estimate. Meeting notes become a project brief. Technical findings become a client-friendly summary.
The AI pattern: AI transforms the raw material into a draft document following your template and style. Human reviews for accuracy and completeness.
Example: Consulting Firm
Before AI:
Consultant spends 3 hours turning discovery call notes into a proposal. Most of that time is formatting, organizing, and writing boilerplate. The unique insight takes 30 minutes.
With AI:
AI takes call notes and generates proposal draft following firm's template: background, objectives, approach, timeline, investment. Consultant reviews, adds the strategic insight, personalizes. 45 minutes total.
The human role shifts: From "write the whole document" to "add the insight and judgment that only you can provide."
The Non-Negotiables: How to Keep AI From Hurting Your Business
Every AI pattern needs guardrails. Without them, you're not using AI deliberately—you're just hoping it works.
Here are the non-negotiables:
Human Review Before Customer Contact
Never auto-send anything to a customer without a human reviewing it first. AI drafts; humans approve and send. This is not negotiable.
Confidence Thresholds
For classification tasks, set a threshold. Above 90% confidence: AI routes automatically. Below 90%: human reviews before routing. You'll calibrate this over time.
Escalation Paths
Every AI mode needs an "I'm not sure" path. When AI hits something it can't handle, there must be a clear escalation to a human. And that escalation should be easy—not buried in a menu.
Kill Switch
You can always turn AI off for a step. Your workflow should function (slower, but functioning) without AI. Never be fully dependent on AI for a critical path.
Audit Trails
Log what AI produced, what the human changed, and what went out the door. You'll need this to improve the AI over time—and to understand what happened when something goes wrong.
The goal isn't to remove humans from the loop. It's to put humans in the right part of the loop—where their judgment actually matters.
The Difference Between Useful and Useless AI Output
If you've experimented with AI and found the results underwhelming, there's something worth considering:
AI output quality is almost entirely dependent on input detail. Vague instructions produce vague results. Specific, detailed, and relevant instructions produce specific and valuable results.
This isn't intuitive. When you delegate to a person, they fill in gaps with context, experience, and judgment. They ask clarifying questions. AI doesn't do that—it takes what you give it and runs.
The difference shows up immediately in practice:
Vague Instruction
"Summarize this customer's situation and suggest next steps."
AI has to guess what matters to you—so it produces something generic.
Specific Instruction
"Given the last 3 support tickets and current contract status, produce: (1) primary issue pattern, (2) sentiment (positive/neutral/negative), (3) recommended action from: [schedule call / send resources / escalate / no action]."
AI knows exactly what to look at and how to structure the output—so it produces something useful.
The specific instruction defines what inputs AI should use, what outputs it should produce, and what format those outputs should take. Nothing is left to interpretation.
This is the same discipline required for designing any good business process. In Deliberate Work, I describe the "5×5 Method"—a framework for specifying exactly what goes into a step and what comes out. The clarity you need for a human to execute a step consistently? You need even more of that clarity for AI.
When "done" isn't well-defined, a human will improvise and probably get close. AI will produce something confidently off the mark.
Where to Start (Without Breaking Anything)
Don't try to "implement AI" across your business. That's how you create expensive failures and frustrated teams.
Instead, run one small experiment:
- 1 Pick one piece of work that happens frequently, follows a pattern, and doesn't require deep judgment. A follow-up email. A request classification. A meeting summary.
- 2 Write down exactly what goes in and what should come out. Be specific. What information does AI need? What format should the output take? What would "good" look like?
- 3 Use ChatGPT or Claude to produce 5 examples. Don't integrate anything yet. Just generate outputs and see how they look.
- 4 Have the person who usually does this work review the AI drafts. Ask them: "Is this good enough to edit? Or would you rather write from scratch?"
- 5 Document what you learn. What worked? What didn't? What would need to change to make this real?
That's it. One experiment. Low stakes. Real learning.
If it works, expand to a second piece of work. If it doesn't, you learned something about what AI can't do for you—and you learned it without breaking anything.
The Real Opportunity
Here's the thing about AI and your business:
Your competitive advantage isn't "we use AI." That's not an advantage—everyone will use AI eventually.
Your competitive advantage is using AI in exactly the right places, with humans in exactly the right role.
AI handling the cognitive grunt work. Your people handling the judgment, the relationships, the craft.
That means you need to understand your work well enough to know which parts are which. You need to be able to specify what goes into a step and what comes out with enough precision that AI can actually produce something useful. You need to design the human-AI collaboration deliberately, not accidentally.
The business owners who thrive in the next decade won't be the ones who automated first. They'll be the ones who learned to distinguish between work that benefits from AI and work that doesn't—and built systems accordingly.
AI is changing work. That's real.
But the risk isn't AI taking your job. It's letting fear keep you from learning how to work alongside it—while others figure it out.
Further Reading
- On execution modes and AI patterns: These concepts are explored in depth in Deliberate Work, particularly in Chapters 15 ("Modes of Execution"), 16 ("AI in the Loop"), and 17 ("Designing Steps for the Future"). Sign up for early access.
- On the 5×5 Method: A framework for specifying exactly what goes into any piece of work and what comes out—the foundation for designing steps that can be executed by humans, AI, or both. Covered in Chapter 13 of Deliberate Work.
- On AI limitations: The non-deterministic nature of large language models is well-documented. For a technical perspective, see: Zhao, W. X., et al. (2023). "A Survey of Large Language Models." arXiv preprint arXiv:2303.18223.