The Dimensionality Test: A Better Framework for Knowing Where AI Belongs in Your Business

Walk into any SMB-focused subreddit and the same question shows up every week: Do small businesses really need AI? Do brick-and-mortar shops need automation? Is this hype, or actually useful?

The question keeps recurring because most answers are bad. Vendors tell you yes. Skeptics tell you no. Both are arguing the wrong thing.

The real question isn’t whether your business needs AI. Businesses don’t have problems. Tasks have problems. And once you start looking at AI through the lens of tasks — not industries, not company size — the picture gets a lot clearer.

Here’s a framework I find useful.

Two axes, not one

Every task in your business can be plotted on two axes:

Repetition — how often does this task happen? Once a year, once a month, twenty times a day?

Dimensionality — how many variables meaningfully affect what the right output looks like?

This is not new ground. Economists have been studying which tasks suit automation since David Autor’s 2003 work on “routine vs non-routine” cognitive tasks, and in 2017 Erik Brynjolfsson and Tom Mitchell published a 23-criteria rubric for “Suitability for Machine Learning” that’s now been applied to over 18,000 work activities in the U.S. economy. What’s worth updating is the language and the implications — because LLMs have broken some of the predictions those earlier frameworks made.

Repetition is the axis everyone already talks about. “Automate repetitive tasks” has been the advice of every productivity blog since 2005. It’s true, but incomplete. If repetition were the only criterion, the right tool would still be Excel macros and Zapier — not AI.

The axis that actually distinguishes where AI shines is dimensionality.

What “dimensionality” actually means

A low-dimensional task is one you can write down as a flowchart. Take a customer’s name, email, and order number. Look up the order in the database. Send a tracking link. Done. There’s a finite number of branches and they all fit on a whiteboard.

A high-dimensional task is one where you can’t write a finite flowchart, because the right output depends on too many subtle variables interacting at once. Three quick tests will tell you which kind you’re looking at:

• The Instruction Test. Could a competent stranger do this correctly with a single-page training document?
• The Edge-Case Test. Does roughly every fifth instance break the rules of the previous four?
• The Substitution Test. Would two equally competent employees, given the same input, produce visibly different outputs?

If any of these tip toward no, one page isn’t enough or yes, edge cases are common or yes, employees would disagree — you’re in high-dimensional territory.

These tasks aren’t repeatable in the rote sense. Each instance is a little bit different from the last. But the pattern is repeatable. That distinction is the whole game.

What automation actually pays for

There’s an underrated economic insight beneath the dimensionality framing:

Traditional automation reduces volume cost. RPA, scripts, and macros let you handle more instances of the same task at lower marginal cost. They scale repetition.

AI reduces variance cost. LLMs let you handle the unpredictable variation between instances without writing a rule for each variation. They scale judgment.

If your task is expensive because there are a lot of instances of the same simple thing — that’s a volume problem. Hire a script, an autoclave, or an offshore team. If your task is expensive because each instance requires interpretation, context, or a small judgment call — that’s a variance problem. That’s the cost LLMs were built to attack.

The “high-rep, high-dim” quadrant is just the geometric statement of high volume and high variance, both at once. That’s where AI has a unique edge that no prior tool offered.

Why this distinction matters now

For decades, automation only worked on the bottom-left of this 2×2:

If a task was low-dimensional and high-repetition, you could automate it with rules. If it was high-dimensional, you needed a person — and the more a task depended on context, judgment, or nuance, the more it stayed locked behind human attention. The economics literature in the early 2000s explicitly predicted this: non-routine cognitive tasks were supposed to stay safely human, because rules-based computers couldn’t handle them.

LLMs broke that prediction. They are generalization engines over high-dimensional spaces — they don’t follow rules; they learn the shape of a subspace from examples and produce sensible outputs for inputs they’ve never seen before. That capability is what unlocks the top-right quadrant for the first time.

That’s why “AI for small business” feels both overhyped and undersold at the same time. The hype is wrong about which tasks. The skepticism is wrong about whether any tasks fit.

A working example: a dental practice

Take a dentist with a four-person staff. The Reddit question would ask: “Does this dentist need AI?” Wrong question. Let’s plot the tasks instead.

• Sterilizing instruments. High repetition, low dimensionality. Don’t use AI here. Use a checklist and an autoclave.
• Performing a root canal. Low repetition per patient, extremely high dimensionality. Don’t use AI here either. This is where human expertise lives.
• Submitting insurance claims. High repetition, low-to-medium dimensionality. Traditional automation handles this well.
• Answering the phone. High repetition, high dimensionality. Every caller is a slightly different problem with a slightly different intent. This is the AI candidate quadrant.
• Choosing whether to buy a new CBCT scanner. Low repetition, high dimensionality. Human judgment, possibly informed by AI tools, but not delegable to AI.

Notice that “the dentist’s business” doesn’t fit on the chart at all. Tasks fit on the chart. Some have a clear AI answer; most don’t.

Why “candidate” and not “sweet spot”

A first draft of this framework would stop here. It would conclude: high-rep + high-dim = AI sweet spot. Plug in a tool and go.

That conclusion is too strong. Living in the high-rep, high-dim quadrant makes a task an AI candidate — not an AI fit. Four filters determine whether a candidate actually flies. And critically: if any one of these filters fails catastrophically, the others can’t compensate. You don’t average your way around a veto.

1. Verification cost

When the AI gets it wrong, can someone catch it cheaply?

Voice AI for a dentist works in part because callers self-correct (“no, I meant Tuesday”) and in part because someone at the practice glances at the day’s call log and notices anything odd. Mistakes surface quickly and are cheap to repair. AI-assisted medical diagnosis fails the same test in reverse: even if the model is 95% accurate, verifying which 5% is wrong costs as much as doing the diagnosis from scratch. The 5% you don’t catch is the lawsuit.

2. Failure severity

If the AI fails on a hard case, what is the worst plausible consequence?

Some errors are recoverable: a confused caller hangs up and calls back; a misrouted email gets re-sent. Some errors are catastrophic: a missed cancer indication, a botched contract clause, a wrong dose. AI deployed where the worst case is catastrophic is a liability regardless of average accuracy. The bar to ship isn’t “right most of the time” — it’s “wrong in a way you can survive.”

3. The tacit-knowledge ceiling

Michael Polanyi’s old observation, beloved of economists: we can know more than we can tell.

A lot of expert work is high-rep and high-dim and still resists automation, because the relevant expertise has never been written down. A senior trial lawyer’s instinct for which witness to call first. A head waiter’s read of which table to seat a difficult regular at. A master plumber’s ear for which gurgling sound means the wax ring is failing rather than the trap. None of that is in any training dataset. The AI will sound right while being wrong.

There’s a useful second-order observation here: the process of trying to automate a tacit task often forces an organization to articulate it for the first time. Even AI projects that fail outright tend to leave behind a written workflow, a checklist, or a shared vocabulary that didn’t exist before. That residual value is real — but you should treat it as a consolation prize, not the win.

4. Feedback loop density

How quickly does the system find out it was wrong?

Phone AI gets fast feedback: callers correct it in real time, the calendar shows whether bookings landed, the receptionist hears about complaints same-day. Contract-review AI gets slow feedback: an error compounds for months before anyone notices, and by then five other contracts have been signed on the same flawed pattern.

Fast feedback loops let you catch and fix early — and the system gets better in production. Slow feedback loops let small errors silently scale into big ones. For a task in the AI candidate quadrant, the feedback loop is often what determines whether the deployment improves over time or quietly degrades.

Subspace coverage: why vertical AI wins

A related point worth flagging: LLMs only generalize well over the subspaces they’ve actually been trained on or grounded in. A general-purpose model is fluent in customer service language, calendar scheduling, basic medical and legal vocabulary, and a thousand other common subspaces. It is not fluent in your proprietary internal jargon, your specific clinical pathway, or the niche language of an industry that doesn’t generate much public text. This is the gap that vertical AI products fill — and it’s the practical reason a dental-specific voice agent will outperform a general one on a dental front desk.

The phone-answering example survives all four filters and lives in a well-covered subspace: callers self-correct, the relevant expertise is articulable, mistakes are recoverable, feedback is fast, and language models have seen enormous volumes of dental-adjacent customer service language. That’s why it’s the SMB use case that consistently delivers ROI rather than just demoware.

A 2023 study at the ACM UMAP conference (Salimzadeh et al.) found that human-AI team performance actually decreases as raw task dimensionality rises — a finding that contradicts the naive theory but aligns perfectly with the four-filter view. Dimensionality alone isn’t a green light. Verification, severity, articulability, and feedback are what turn a candidate into a fit.

One more thing: workflow fit

Even a task that passes all four filters can fail in deployment if the AI doesn’t fit how work actually happens. If staff have to switch systems, double-check every output, override the AI more often than they accept it, or paper over its errors with manual reconciliation — the organization has added friction rather than removed it. The MIT NANDA GenAI Divide report (2025) found that about 95% of enterprise GenAI pilots fail, and that success correlates not with task suitability alone but with deep workflow integration. A well-fitted candidate plus poor integration produces shelfware.

This isn’t a fifth filter so much as a precondition: before you invest in any of the surviving candidates, ask whether the AI’s output will drop into the existing workflow without being re-handled. If the answer is “we’ll figure it out later,” figure it out first.

The exercise to actually do

Forget industry reports. Forget vendor pitches. Spend thirty minutes doing this:

1. List every recurring task someone in your business performs more than ten times a week.
2. Run the three dimensionality tests on each (instruction, edge-case, substitution). Any task that fails at least one is high-dimensional — a candidate.
3. For each candidate, score it on the four filters: verification cost, failure severity, articulability, and feedback density. If any single filter scores catastrophic, the task is out — regardless of how strong the others are.
4. Check workflow fit. Will the AI’s output land in the existing workflow without being manually re-handled? If not, fix that before investing.
5. The candidates that survive — high rep, high dim, four filters clean, workflow fit confirmed — are your real AI investments.

Most businesses, when they do this honestly, find one or two such tasks. Not twenty. Not zero. One or two. Those are the ones worth investing in.

The framing worth holding onto

The reason “should small businesses use AI?” generates so much circular debate is that it’s a category error. Businesses aren’t the right unit of analysis. Tasks are.

Repetition tells you whether automation is worth the setup cost. Dimensionality tells you whether AI specifically — variance-cost reduction — is the right kind of automation, versus a script, a checklist, or a hire. The four filters tell you whether an AI candidate actually deserves to ship. Workflow fit tells you whether it will survive contact with how your business actually runs.

The businesses that will get real value out of AI in the next five years aren’t the ones who adopt it broadly. They’re the ones who identify the small number of tasks in their operation that genuinely live in the high-repetition, high-dimensionality quadrant — survive all four filters cleanly — and then integrate the chosen one or two into how work actually gets done.

Everything else is theatre.

Leave this field empty if you're human: