AI Won't Automate Your Problems Away

There's a story being told about AI right now that goes something like this: adopt the right tools, automate the right workflows, and your team shrinks while your output grows. It's clean. It's compelling. It fits neatly into a pitch deck.

It is also, for most ecommerce operations in mid-2026, substantially wrong.

Not because AI is useless — the tools are genuinely powerful and improving at a pace that's hard to overstate. But there's a widening gap between what AI automation is being sold as and what it actually delivers when you point it at the messy, consequential, context-dependent work of running a commerce business. That gap is where reputations get damaged, money gets wasted, and operators who trusted the pitch find themselves worse off than if they'd just done the work themselves.

This one's about that gap — what causes it, where it shows up, and how to think about it without losing your mind.

The Confidence Problem (Or: Why Your AI Lies Like a Job Candidate)

The single most dangerous thing about current large language models is not that they make mistakes. Every tool makes mistakes. The problem is they make mistakes with the exact same confidence as correct output, and they genuinely do not know the difference.

Here's what that looks like in practice. Ask a model to research 100 companies matching a specific buyer profile, and it will cheerfully fabricate 30 of them — inventing names, domains, and social media handles for businesses that do not exist. Ask it to analyse a Search Console export, and it will invent page names that aren't in the file while completely ignoring the highest-traffic content sitting right there in the data. Ask it to transcribe an audio file, and it may hand you a confident transcript of something that was never said.

These aren't hypotheticals. They're Tuesday.

And when you challenge the output, the model doesn't say "my bad." It doubles down. It invents explanations for why the wrong answer is actually right. It claims to have performed verification steps it never performed. It produces revised output labelled "100% verified" that contains the same category of errors. (It's like a new hire who, when caught making up a sales number, responds by making up a more detailed sales number.)

For anyone running an ecommerce operation, this isn't abstract. A fabricated competitor in a market analysis leads to a flawed positioning strategy. A hallucinated product spec in a listing generates returns and tanks your seller metrics. An incorrect data analysis sends your PPC spend chasing the wrong keywords for three weeks before someone notices.

Inconsistency Is the Real Villain

If AI were consistently wrong, we could work with that. You'd learn which tasks to avoid and move on. The far more insidious problem is that it's inconsistently wrong.

The same model, given the same prompt, can produce flawless output on Monday and fabricated nonsense on Tuesday. A product description workflow that generated clean, on-brand copy for three weeks will suddenly start making claims that have no basis in the product data. A supplier enrichment process that nailed 200 entries last month hallucinates half the dataset when you run it again.

This is the part that genuinely breaks people's workflows. You can't build reliable processes around a tool when the boundary between "trustworthy" and "making things up" moves every time you use it. And the review burden this creates is a different kind of hard — because AI output looks polished on the surface. The errors aren't typos. They're structural fabrications wrapped in fluent, professional-sounding language. It's the difference between catching a misspelled word and catching a confidently stated fact that doesn't exist.

The Paradox Nobody Warned You About

Here's where the economics of AI adoption get genuinely counterintuitive, and it's worth understanding because it explains why so many teams feel busier after adopting AI tools.

When AI makes previously expensive tasks cheap, more people attempt those tasks. An operations manager who would never have written a product listing now generates 50 of them in an afternoon. A marketing coordinator who would never have attempted a competitive analysis now produces one in an hour. An Amazon seller who would never have built A+ content for their entire catalogue now has drafts for every ASIN.

The volume goes up. But so does the volume of output that needs someone who actually knows what "good" looks like to review it — whether the listing reflects the brand's positioning, whether the competitive analysis is based on real companies (important detail, that), whether the A+ content tells a story that will resonate with buyers in that specific category.

This is where the "replace junior staff with AI" thesis falls apart in practice. Companies that cut entry-level roles assuming AI handles those tasks have often discovered they've eliminated the human capacity to catch errors. The senior people who remain can make strategic decisions, but they're too far from the operational detail to notice when an AI-generated listing contains a fabricated claim, or when a data analysis has quietly dropped 20% of the input rows.

The result is a compounding problem. AI produces more work. Fewer people review it. The errors that slip through are harder to detect because they sound right. And the business accumulates a growing layer of decisions built on foundations nobody has verified. Gartner research has consistently shown that fully automated ecommerce operations experience critical failures at roughly six times the rate of human-in-the-loop systems. And that ratio has held as models improved, for a straightforward reason: better models produce more convincing errors, not fewer.

The Slop Problem (Or: When Everyone Sounds Like the Same Robot)

There's a second-order problem here that's less about reliability and more about differentiation — but it matters just as much commercially.

Every seller in your category has access to the same models. Those models are trained on the same data. When everyone feeds them similar prompts with similar inputs, the output converges. Listings start to sound alike. Ad copy follows the same structures. Brand stores adopt the same visual patterns. No individual output is bad — it's competent, functional, adequate. But adequate, multiplied across an entire category, becomes invisible.

The industry is increasingly calling this "slop" — not a single identifiable flaw, but a pervasive sameness that both audiences and algorithms are learning to recognise and penalise. AI discovery systems, including the retrieval mechanisms behind Google's AI Overviews and ChatGPT's shopping features, are already filtering for signals of originality and engagement. Content that reads like default model output — because it is default model output — is structurally disadvantaged in exactly the environments where product discovery increasingly happens.

Here's the thing: in an environment where AI-generated content is abundant and converging, the scarce resource is the judgment that makes output specific, distinctive, and fit for purpose. That judgment is human. And it is not getting cheaper.

The Maintenance Tax Nobody Mentions

The vendors selling AI automation tools have a curious habit of not discussing the ongoing cost of making them actually work. This matters because it materially changes the ROI calculation.

Agents and automated workflows are not set-and-forget systems. They need continuous maintenance: prompt refinement as model behaviour shifts between versions, monitoring for output quality drift, updating context as your business conditions change, debugging failures from model updates, and managing the token costs that accumulate as workflows scale. (Think of it less like installing software and more like adopting a very talented but temperamental pet that needs constant attention and occasionally eats your data.)

Organisations that have gone deepest on AI adoption report that maintaining agent infrastructure requires dedicated engineering resources — and that agents assigned to individual team members tend to go stale and get abandoned unless someone actively invests in keeping them effective. For a large enterprise with dedicated AI engineering teams, this is manageable. For a mid-market ecommerce brand or a growing Amazon seller, the honest cost comparison isn't "AI tool subscription versus employee salary." It's "AI tool subscription plus the time your team spends reviewing output, correcting errors, maintaining workflows, and managing the system" versus "the employee who did this work before."

That comparison often still favours AI adoption, particularly for high-volume, repetitive tasks. But the margin is narrower than the sales pitch suggests.

The Sandwich Model (Yes, We're Calling It That)

None of this is an argument against using AI. The tools are too powerful and improving too quickly to ignore. The argument is for using them with an accurate understanding of what they can and can't do.

The most effective deployment model isn't maximum automation. It's what we might as well call the sandwich model: a human sets the frame — defines the task, provides the context, specifies what good looks like — the AI does the production work within that frame, and a human reviews the output and makes the consequential decisions.

This isn't a transitional phase on the way to full automation. For the foreseeable future, this is the architecture that works.

The businesses getting burned are the ones skipping the bread on either end of that sandwich. They're handing vague briefs to AI tools and publishing the output without review. They're replacing the people who understood what good looked like and discovering, too late, that nobody remaining can tell the difference between a competent draft and an expensive hallucination.

The Bottom Line

The state of AI in mid-2026 is this: it is a powerful amplifier of human capability and a poor substitute for it. The tools can make a competent operator significantly more productive. They cannot make the operator optional.

The vendors selling full automation know this, even if their marketing doesn't say it. The benchmarks they cite measure model performance inside carefully designed evaluation frames — prompts loaded with context, constraints, and human judgment that shaped the task before the model ever saw it. Strip away that framing, and performance degrades. The gap between what AI does when expertly directed and what it does when left alone is the gap between the pitch deck and the production environment.

Use AI aggressively for the tasks it handles well. Maintain human oversight where errors have consequences. And treat anyone promising a future where your team is optional with the same scepticism you'd apply to any vendor whose incentives don't align with yours.

The humans aren't a transitional cost. They're the layer that makes the automation worth running.

P.S. If you've already fired your listing specialist and handed the job to Claude, this is not the time to panic. It is the time to quietly re-read some of those listings before your return rate tells the story for you.

P.P.S. Yes, I am aware of the irony of an AI newsletter telling you not to blindly trust AI. We contain multitudes. We also contain quality control processes, which is sort of the whole point.

• TikTok launched TikTok GO for in-feed travel booking and Branded Buzz to match advertisers with creators at scale, pushing harder than ever to become a full commerce platform.

• Google appears to have changed how it handles self-serving "best of" listicles in AI Overviews. Your list may now help competitors get recommended while your own brand gets left out.

• Anthropic launched a small business tier for Claude. One toggle installs everything, connects to tools like QuickBooks, HubSpot and PayPal, and runs end-to-end workflows from day one.

• Higgsfield launched Supercomputer, a multi-agent creative platform for running full production pipelines. Script, cast, shoot, edit and publish, with memory and slash command skills built in.

• Anthropic overtook OpenAI in business adoption for the first time, hitting 34.4% vs 32.3%, per Ramp data. Anthropic quadrupled adoption over the last year while OpenAI grew just 0.3%.

• Google pushed its UCP Buy button into standard search results for the first time, letting shoppers complete purchases without leaving the page. Currently live for Wayfair, with Etsy and Target expected next.

• Tavus launched Image-to-Replica, letting anyone build a fully conversational AI human from a single photo. Historical figures, brand mascots and AI-generated personas are all now fair game.

📝 Superlist - Create to-do lists, capture thoughts or detailed notes, assign tasks to teammates, and more

🗣️ Just words - Optimize your product's copy for user growth.

🙋🏻‍♀️ Kraftful Surveys GPT - Create product surveys for insightful customer feedback

⏺️ Strut AI: Quickly capture projects, notes, drafts, and more in collaborative workspaces powered by AI.

 🌐 Dorik AI - Generate beautiful websites from a single prompt.

😍It was great, all I needed

🙂Good, not great

🙄It sucked