The Two AI Traps Costing Small Business Owners Real Time and Real Money
By Warren Schuitema, Founder | Matchless Marketing | The AI Dad
Most people using AI tools right now are making one of two mistakes. Either they're throwing every task at the model and wondering why the output is inconsistent, or they're measuring their AI use by how much they use it instead of what it actually produces. Both of these traps will cost you.
I was listening to a Lenny's Newsletter episode this week featuring Dianne Penn, the person who joined Anthropic in 2023 as their first technical product manager, when the entire product team was five engineers. She's since helped ship every model from Claude 2 through Fable and helped incubate Claude Code, MCP, and computer use. She knows how these models get built from the inside. Two things she talked about landed hard for me, because I've watched small business owners fall into both of them.
Here's what they are, why they matter to your operation specifically, and what I do differently because of them.
The Jagged Edge Is Real, and It's Not Going Away
The jagged edge is one of the most useful mental models I've come across for understanding AI capability. The idea is this: AI systems aren't uniformly good or uniformly limited. Their capability looks more like a wall with towers and recesses. Brilliant in some spots, surprisingly bad in others, and there's no obvious pattern that tells you in advance which is which.
What makes this genuinely confusing is that the tasks AI handles well don't always look harder than the ones it fails at. Claude will write a thoughtful 1,200-word proposal draft faster than I can outline it. Ask it to remember a specific piece of context from three hours ago in a long session, and it sometimes drops it entirely. It'll nail structured writing. It stumbles on certain types of multi-step reasoning where the logic depends on tracking several variables at once.
For a small business owner with no developer and no AI team, this matters in a very specific way. You're not testing models at scale. You're relying on a handful of tools to help you get through your week. If the jagged edge catches you on a task you were counting on, that's not a minor inconvenience. That's a deliverable that doesn't go out, or a client-facing output that goes out wrong.
The practical move here isn't to avoid AI. It's to map the edge on your own tasks before you depend on it. Try the task. See what comes back. Run it a few times. You're not testing the AI abstractly, you're building a personal reliability map for your specific workflow. Once you know where Claude or ChatGPT reliably performs and where it gets shaky, you stop being surprised by the misses.
Token Maxing Is the New Productivity Theater
Token maxing, or "tokenmaxxing," is the practice of maximizing AI usage, specifically consuming as many tokens as possible through autonomous agents and extended runs. It's emerged as a workplace status signal in parts of the tech industry, where high AI utilization gets treated as a proxy for productivity regardless of what that usage actually produced.
I want to name this clearly: it's the AI version of looking busy.
For enterprise teams, token maxing shows up as burning through agent loops, stacking context windows, and reporting high usage numbers to leadership as evidence of AI adoption. For solo operators and small business owners, the version is subtler. It looks like spending two hours "working with AI" to produce something you could've drafted in 40 minutes. It looks like running five different AI-generated versions of a thing when you only needed one. It looks like using AI for tasks where the overhead of prompting, reviewing, and correcting eats the time you were trying to save.
The question I ask before I delegate anything to Claude or ChatGPT is simple: what's the actual output I'm comparing this against? If I can produce a decent version of this in 20 minutes on my own, and the AI route takes 30 with iteration, the AI route isn't faster. It just felt more high-tech. That's not a business advantage. That's theater.
How I Actually Build With This in Mind
I run a fully automated content and prospecting operation at Matchless Marketing. Multiple AI agents handle research, outreach prep, content writing, and CRM updates. That setup didn't happen because I threw everything at AI and hoped. It happened because I tested each piece of it on real tasks first and only automated what consistently worked.
Here's the actual process I use when I'm adding something new to my AI stack:
Step 1: Run the task manually first. Before I involve AI at all, I do the task myself once. That gives me a baseline I can measure against. If I can't describe what "good" looks like, I can't evaluate what the AI produces.
Step 2: Test the specific task, not the tool generally. Don't ask "is Claude good at this?" Ask "is Claude good at this specific task, with this specific context, for this specific audience?" Those are very different questions, and the jagged edge means the answer to the second one is the only answer that matters.
Step 3: Count your time honestly. Time in the prompt, time reviewing the output, time correcting it, time reformatting it. All of that counts. If the total is longer than doing it yourself, you haven't found a workflow yet. You've found a practice run.
Step 4: Only automate what you've verified. This is the one I see people skip most often. They automate a task after one or two successful test runs, then wonder why the agent output is inconsistent three weeks later. I don't automate anything until I've run it enough times to trust the pattern.
What the Insider View Confirms
The reason the Dianne Penn episode was useful to me isn't because she revealed something nobody knew. It's because someone who helped build these models from the inside at Anthropic is describing the same characteristics that every serious AI practitioner has encountered in practice.
Penn explains the decisions that shaped Claude, including how Anthropic built an eval-driven development loop that let them iterate quickly based on hard numbers. That's exactly the approach every business owner should take with their own AI tools. Not vibes. Not impressions. Evals. Run the task, measure the output, adjust.
Token maxing is the practice of maximizing one's AI usage, specifically by consuming as many tokens as possible with autonomous agents, and it has emerged as a workplace trend with ultra-high AI utilization being treated as a signal of productivity, regardless of the output. When that's happening in enterprise companies with AI budgets and teams, you can imagine what it looks like in a solo operation where every dollar and hour is yours personally.
The jagged edge isn't a bug that's going to get patched. It's the uneven boundary of current AI capability, where systems are superhuman at some tasks and surprisingly poor at others of seemingly similar difficulty, producing erratic performance that requires careful oversight. Knowing that going in changes how you build.
One Thing You Can Do Today
Pick one AI task you currently rely on in your business. Something that's part of a real workflow, not a test. Run it three times in a row, with slightly different inputs or contexts, and compare the outputs. You're not looking for the AI to fail. You're building your personal reliability map for that task.
If the outputs are consistent and strong, you've confirmed something you can build on. If they're inconsistent, you've just found the specific edge you need to account for before you trust it with anything important.
That's not pessimism about AI. That's how you actually use it well.
Warren Schuitema is the founder of Matchless Marketing and the creator of The AI Dad, a brand and platform helping small business owners and solopreneurs implement AI tools without hype, overwhelm, or a developer on retainer. He builds, tests, and documents real AI systems live so his audience can follow what actually works inside a running business. He is the operator behind a fully automated AI agent workforce managing content, leads, research, and client onboarding at Matchless Marketing. The eval-first mindset he describes here is the same one behind every system he's built and handed off to clients.