Podcast: Download (Duration: 40:56 — 47.1MB)
AI cannot supply context it was never given, judge subjective work the same way twice, or hold your instructions across a long task. Everything with a verifiable right answer, from code to campaign reporting, it now does faster and better than a person.
Toni Herrbach and I compared how much of our businesses we have handed over. I rebuilt the Bumblebee Linens website in one weekend, work that took a human design team four weeks last time.
Toni’s copywriter needed five minutes to catch a voice shift in 20 AI-written video scripts. Hours of AI-assisted review had missed it entirely.
This post covers what to hand off today, the four ways AI fails in a real business, how to structure batch work so it stops drifting, and where a human still earns their fee.
Get My Free Mini Course On How To Start A Successful Ecommerce Store
If you are interested in starting an ecommerce business, I put together a comprehensive package of resources that will help you launch your own online store from complete scratch. Be sure to grab it before you leave!
Table of Contents
Key takeaways
- AI wins outright on tasks with a verifiable right answer, such as code, math, and report analysis.
- AI turns accurate data into bad recommendations when one piece of context is missing.
- Quality decays across long sessions, and corrections quietly revert on the next pass.
- Give each item in a batch its own agent with a clean context window.
- AI ratings of creative work flip between runs, so a human makes the final call.
- Generate a best practices guide with deep research before starting any AI project.
- Generated images and video need a close check for anatomy, spelling, and physics errors.
- Experts get the most from AI because they know what to ask and can spot a wrong answer.
What can AI do better than a human?
Anything with a definitively right or wrong answer, which means code and math above everything else. Nothing beats it when the output can be verified.
Report analysis is the highest-value handoff for most sellers. Toni pulls and analyzes hundreds of Klaviyo campaign reports through AI instead of exporting, sorting, and clicking through each one.
The time savings are large. Two months of campaign analysis that would take her three hours to do well now runs while she works on something else.
It reaches conclusions you would soften. Claude told me flatly to stop sending an entire category of campaigns because they do not convert.
Post-sale reviews scale the same way. A seasonal promotion with 30 emails comes back as an answer key instead of a spreadsheet project.
What AI cannot do for your business
AI cannot fill in context nobody gave it, judge subjective work consistently, hold instructions through a long task, or catch its own small visual errors. Those four gaps account for every failure Toni and I have hit.
Some tasks also still need hands on software. My personalized embroidery orders flow automatically from the website into machine files, and a person still adjusts letter spacing for certain fonts before stitching.
Automating that last step means AI controlling the embroidery software directly. It is on my list and has not been the priority.
Here is how the division of labor looks in our businesses today.
| Task | Hand to AI | Keep a human |
|---|---|---|
| Pulling and analyzing reports | Yes, entirely | Read the context behind any surprising number |
| Writing code | Yes, entirely | Know enough to ask the right questions |
| Website design and copy | Yes, from a brand brief | Initial direction when starting from zero |
| Generating images and video | Yes | Review every asset for errors |
| Rating thumbnails and taglines | No | Final choice on anything subjective |
| Batch writing 20 scripts | Yes, one agent per script | Fresh-eyes read of the whole set |
| Monitoring 100 email flows | Yes, flags underperformers | Decide whether a drop is seasonal or a problem |
| Accessibility sweep | Yes, given a best practices guide | Nothing, if the guide is loaded |
Why does AI give bad recommendations from accurate data?
Because a single missing piece of context changes the right answer, and AI will not know to ask for it. Toni’s example is an inventory report showing 8,000 units sold over six months.
Roughly 4,000 of those sold during the launch promotion. Real ongoing demand was closer to 1,000 a month.
Inventory got ordered against the inflated number. The report was accurate and the recommendation was wrong.
Campaign analysis hides the same traps. A send AI flags as a failure may have gone out on a Saturday night, landed too soon after another email, or promoted a product that was out of stock.
Promotion fatigue is invisible in the data. An offer that flopped may have run two months earlier and worked.
You have to tell it what to weigh. The nuance lives with the person who made the original decision.
The best AI output starts with human input. Toni’s view is that the strongest copy comes from AI iterating on something you wrote, and the strongest analysis comes from AI surfacing everything so you can spot what it cannot.
Why does AI quality drop on long tasks?
Because context degrades as a session runs, and the model slides back to old habits after you correct it. Toni watched the hooks in a batch of 20 founder-facing video scripts go from good to terrible around the halfway mark.
Corrections did not hold. She told it to stop ending every script with the same line, it agreed and fixed the close, then reverted on the next iteration.
Best practices, marketing briefs, and her full content history were all loaded. The drift happened anyway.
Supervising the batch took almost two hours. Every few scripts needed another round of feedback.
Treat it like a small child. You cannot hand off a long task and trust the instructions survive to the end.
How do you stop AI from drifting across a batch?
Give each item its own agent with a clean context window, so nothing accumulates from one script to the next. That is how I produce my YouTube scripts.
Add graders as separate agents. A retention rater scores every script against my guidelines.
Fact checking gets its own agent too. Isolating it keeps the writing agent’s drift from contaminating the check.
The cost is tokens. Clean contexts burn far more than one long session, and the consistency is worth it.
Can AI judge creative work?
No, because its opinion changes between runs on identical inputs. I generate about ten thumbnail variations per video and ask four different AI tools to rank them.
The rankings flip. Sending the same images to the same tool twice produces different winners.
The instability reveals the real problem. There is no correct answer to converge on, so consistency never appears.
Humans break the tie. I ask Toni, I ask Jen, and I ask my editor before choosing.
Taglines needed the same treatment. AI was confident that its Bumblebee Linens tagline beat anything I had written, and I still had to reshape it into something I liked.
Jen decided the final version. She wanted the company represented a specific way, and I would not have gotten there alone.
What did a human catch that AI missed?
A copywriter spotted a voice change at video eight in Toni’s script batch, five minutes into reading. She asked whether there was a reason the delivery switched from one style to another.
There was no reason. AI had drifted mid-batch, and Toni had stared at the content too long to notice.
The copywriter uses AI herself. Her edge was expertise and fresh eyes, which is what caught the switch.
Toni’s design agency made a similar catch. Emojis in announcement bars and heading text register as noise for screen readers, which makes them an accessibility problem she had never considered.
Nearly every AI-generated announcement bar she had seen included an emoji. The agency flagged it as frowned-on practice in seconds.
Can AI replace a graphic designer?
For execution, largely yes, and for direction and taste a designer still adds real value. The common objections to AI images have both collapsed.
Resolution is solved. Nano Banana 2 and Nano Banana Pro generate 4K images.
Editability is solved too. Canva’s magic tool separates a generated image onto layers, so text, faces, and graphics move and resize independently.
Direction is where designers win. A designer knows the spacing, lighting, and composition choices that a prompt never specifies.
Toni’s view is that a designer using AI takes on more clients rather than losing them. The workflow speeds up and the taste stays.
Starting from zero is the hardest case. Someone without a brand, history, or data benefits from a designer setting the initial direction, or at least a branding kit to build from.
Taste is the constraint. People with poor style feed AI poor direction and approve poor output, while AI itself has good taste when told to follow best practices.
How do you check AI-generated images and video?
Go through every generated asset with a fine-tooth comb, because the errors are small and specific. Toni caught a cow with a fifth leg while making images with her grandkids.
Spelling breaks inside images. Words come out rendered wrong in otherwise clean graphics.
Video fails at physics. My AI-generated demo for a pizza cutting board was perfect until the end, when the board tilted vertical and the pizza stayed glued in place.
I generate most of my site images and videos in Higgsfield. Someone still reviews every one before it goes live.
Nobody catches every AI image. Toni has published generated images that no one identified, and she still checks each one.
How much does AI cut the cost of a website redesign?
It compressed a four-week project into one weekend for me, and the result arguably looks better. The new Bumblebee Linens site went live the night before this recording.
Working with a human designer meant a day’s wait for every change. AI made each change in minutes.
It read the entire existing site first. I gave it the brand story and what we do well, and it produced the design and all the copy.
Images were the only manual input. I had it write prompts referencing my products, then generated a set of images to use across the site.
The cost gap is enormous. A full agency redesign can run $50,000 against roughly $100 in credits.
Justifying $50,000 means predicting a conversion lift in advance. Few sellers can, and AI gets you 90% of the way there.
I still held back on a more drastic design. Claude Fable researched wedding forums and Reddit, built a brand kit, and mocked up a full redesign that looked fantastic.
Navigation and conversion elements stopped me. A beautiful site is only useful if it converts as well as the old one.
Why generate a best practices guide before an AI project?
Because it loads domain knowledge you lack before any work begins, and it changes what AI catches without being asked. This is my first step on every project.
Deep research produces something genuinely comprehensive. My advertising best practices guide ran about 80 pages and taught me things I did not know, and I teach this material.
The guide triggered an unprompted accessibility sweep during my redesign. AI changed my shade of teal for contrast compliance, then offered a full audit of the site.
Whole-project access is the other half. I ran the redesign where it could read every source file, so it acted across the entire site instead of one page.
Beginners skip this step because nobody told them it exists. If you are new to AI, deep research comes before execution every time.
Does AI mean hiring fewer people?
The first effect is hiring fewer, and the bigger effect is finally reaching work that stayed on the back burner for years. Toni describes it as the difference between affording two people and having ten.
Most small operators cannot hire their way out. They are doing everything themselves and can afford one or two people at most.
Neglected areas get attention. Toni ran a YouTube deep dive for a client that had been postponed because the channel was doing fine.
It analyzed 100 videos in about two hours instead of two and a half days. The recommendation was to update titles and thumbnails on older videos.
She changed three. One two-month-old video sitting at one view a day jumped to 37, with click-through rising from 1% to 3.7%.
Monitoring scales the same way. Toni has around 100 email flows, and AI now flags any that underperform last month so she can decide whether it is seasonal.
Building the system is the human part. Her next step is a weekly process so the review runs without her looking at 100 videos by hand.
Who gets the most out of AI tools?
People who already know the domain, because they know what to ask and recognize a wrong answer when they see it. I write better code with AI than someone who has never coded.
The same holds in every field. A designer gets better images from the same tools than I do, and an experienced copywriter can look at AI output and say try flipping it.
Deep expertise is the protection. If you are an expert in something, AI cannot beat you at it right now.
Refusing to use it is the real risk. People at your level who adopt these tools will out-produce you on sheer volume.
Frequently asked questions
What tasks should you not give to AI?
Anything where a missing piece of context changes the answer, any subjective judgment call, and any long batch you cannot supervise. AI reads data accurately and cannot know that a launch spike inflated it.
Why does AI give wrong recommendations from accurate data?
Because it lacks context nobody thought to supply, such as a product being out of stock or a promotion running two months earlier. The report is right and the conclusion is wrong.
Why does AI get worse the longer you work with it?
Context degrades across a session, and the model drifts back to earlier patterns after being corrected. Quality in a 20-script batch collapsed around the halfway mark despite repeated fixes.
How do you stop AI output from drifting in a batch?
Run each item in its own agent with a clean context window, and add separate grading and fact-checking agents. It costs more tokens and keeps quality consistent.
Can AI replace a graphic designer?
For execution, mostly. Tools like Nano Banana Pro generate 4K images and Canva separates them into editable layers, while a designer still wins on direction, taste, and starting a brand from zero.
Can AI pick the best YouTube thumbnail?
No. AI rankings of the same thumbnails flip between runs because there is no correct answer to converge on, so a human should make the final choice.
What should you do before starting any AI project?
Run deep research and generate a best practices guide for the task, then use it as a reference. It loads knowledge you lack and changes what AI catches on its own.
Will AI replace jobs in a small business?
It reduces immediate hiring while freeing owners to reach work they had postponed. Deep expertise stays the strongest protection, and refusing to use AI is riskier than adopting it.
