Operations — Building the Workflow

Republished with permission from the FrUn SMB AI Guide. Canonical chapter: Superhuman Docs → · All contributors

Mitch Schwartz · AI Operations Designer | Ops Machine · LinkedIn

1. Mindset — What an SMB Should Believe Before Spending a Dollar on AI

I know what you heard this week: that artificial intelligence is changing absolutely everything, that the world is sprinting and you are standing still, that your competitors are already through the door you haven’t found yet. Luckily for you almost all of it is foam.

Watch the ocean. On the surface it is all violence: wind, chop, spray flung in your face, one wave shouting over the last. But drop down a few feet, just a few, and it is quiet. It is slow. It is more or less exactly as it has always been. The storm is real, but the storm is not the sea. And your business, the actual living thing, is not won up in the spray. It is won down in the deep, where the water moves the way water has always moved.

What’s down there in the deep? Two very old questions:

• How do you run a good process?
• How do you lead your people and your work well?

These two are the whole game, and they were the whole game long before anyone taught a machine to talk. This is where it is won or lost. So when the surface gets loud, the move is not to thrash harder. The move is to ground yourself.

Here then, is the one belief worth more than any tool you could buy: AI changed how you build a system, but it changed nothing about what makes a system useful. The instruments are new. The music is the same music it always was.

And here is the unshiny thing: this new instrument leans on those two old skills harder than anything before it.

Consider: every tool you ever bought did as it was told, and every person you ever hired made judgments, and so needed managing. AI is the strange new creature that is both at once: a process that also decides, that takes a fuzzy, half-said thing and acts on it. So running it well is nothing exotic. It is those two skills played together, process design and management, pointed at the same work. That overlap is the only genuinely new note, and you already own both halves of it.

So you have every problem you had a year ago, the same eternal tasks of guiding your people to:

• work together,
• write things down and actually look at them,
• store those things so that they can find them,
• hold their focus and context steady,
• do the work the same way twice even when the inputs are messy.

Oh, and the modern agony of too many systems to live in, which this new instrument is great at solving.

Point it at chaos and you get chaos at scale. Point it at a clean process and, well, you can apparently run things in four hours a week from a hammock somewhere, hosting multiple podcasts.

Either way, nothing fundamental has moved in twenty or 100 years, if you keep yourself grounded.

Now, that MIT statistic: you’re sick of it, I know. 95% of corporate AI projects return nothing measurable.

When it landed, every gum-flapper on the internet was banging it like a drum, mocking the AI-hypesters, and broadcasting that they understood the whole situation just as poorly.

Read the report and it does not say AI fails. It says the failures had almost nothing to do with the machine and almost everything to do with people strapping it onto broken, contradictory workflows and hoping.

The five percent who won did the unglamorous thing first: they got small, chose one real problem, and tended the operations before they reached for the magic.

And that is the good news. It’s not a secret kept by giants. You, with your twelve people, can do it better and faster than a corporation buried under eleven committees and a chief executive waving AI at teams who never asked for it.

And one more piece of good news: maybe your processes are loose? AI is about to make those foundation cracks impossible to keep ignoring, and that is a gift.

Whatever cracks exist in that foundation have been taxing you quietly every week for years; you just learned to push through it.

Now the tax is coming due, loudly. So you’ll fix it, at last.
Not from virtue, but because the numbers will stop making sense any other way.

2. The SMB Tool Stack — Start Cheap, Stay Simple

You need far less than the internet wants to sell you. A useful AI setup has four parts: a brain (the AI model), skills (instructions), context, and a connection to a working system.

The goal to aim at: one AI interface for everything. Not a different AI bolted into every app you own, each with its own quirks and its own learning curve. One capable brain, wired into your systems, that you and your team actually get good at.

That’s why I also skip most products’ built-in AIs. The assistant baked into each product is usually underwhelming, and even when it’s fine, it’s one more interface to learn. Connecting your central AI into those systems instead costs less, works better, is easier to train and to train your team on, and is simply better practice.

Where I’d start:

• An AI workspace (~$20/user/month). Your company brain, and your first hire. Treat it as a workspace, not a chat box. For most operations work (drafting SOPs, summarizing, answering “how do we do X,” working with emails and transcripts, CRMs, task management tools…) this one subscription covers about 80% of what an SMB needs.

So far, Claude is the best non-developer tool I’ve found for being that single interface, and it connects cleanly into a Microsoft or Google stack. Start here.

I recommend orienting towards Claude’s (or if you must, Copilot’s) Cowork sections right away. Shift your mental model from chat to workspace. This is the future. Google is also starting to build this with GDrive’s new project workspaces.

The connection is the piece most people miss, and it’s simpler than it sounds. It isn’t a heavy automation platform like Make or Zapier. It’s plugging your one AI directly into the tools you already work in: Notion, ClickUp, or Wrike, plus Google Drive, Gmail, and Google Calendar. Once it can read your docs, your tasks, and your schedule, it stops being a chat box and starts acting as an assistant and a strategist at the same time.

Zero budget?
Your first job is to gain vision of how $20/mo/seat gives you a 20x ROI. If it’s not a no-brainer then you’re probably right: there’s work to do to get there first. Start the journey.

Related caveat: This write-up is about the execution layer: picking the right job and getting AI to do it well. It assumes you’ve laid the foundation first. Before you wire AI into real systems and real data, get your data privacy, security, access control, and a simple team-use policy sorted. None of it is exotic. Any good ‘AI person’ should be able to lead you through this. Skipping it is how good automation turns into expensive mistakes.

If that groundwork isn’t in place yet, that’s job zero, and walking it with someone who’s done it before is high leverage. Don’t drag it out; time is leverage. You’re spending it either way.

3. The First 30 Days — Where to Start

Your first 30 days isn’t an “AI project.” It’s onboarding one new worker, and you’ve onboarded people before. Don’t try to transform the company. Hire one capable worker, give it one job, and manage it closely.

One expectation: your first attempt will teach you more than it delivers. Most owners overestimate month one and underestimate month twelve.

First, pick the job. List the tasks that eat your team’s time, then score each one. A good first job scores yes on all five:

• Frequent: happens daily or weekly
• Repeatable: same steps most times
• Text or data based: email, documents, spreadsheets, transcripts
• You can describe how it’s done: out loud, start to finish, without hand-waving
• Measurable: you can name what it costs you today, in time or errors, so you’ll know whether AI actually beats it

Strong first jobs in most SMBs: sorting and triaging the shared inbox; turning call or meeting recordings into notes and action items; drafting first-pass proposals or follow-up emails from a template; matching payments or invoices to records; cleaning and formatting data; answering the repeated “how do we do X” questions.

Skip for now: anything customer-facing that ships without review; anything that needs judgment you can’t put into words; anything irreversible or high-stakes; and anything touching sensitive data you haven’t cleared for AI yet.

Weeks 1–2 — Write the job description. (Design the work.)

• Pick the single highest-scoring job. One, not three.
• Write the steps the way you’d hand them to a new hire
  • What triggers the process?
  • What is the output of the process?
  • What are the skills required?
• List the common edge cases and what to do with each
• Note where the information lives: which inbox, folder, board, or sheet
• Decide how you’ll measure quality and cost if you aren’t already, and take a “before” reading: time per run, error rate, current cost. That baseline is what you’ll judge the AI against.
• If you get stuck writing it down, try dictating it to a team member or to AI.

Get the process on paper. Workflow diagrams are a great tool. Text works.

Weeks 3–4 — Bring the worker on, and supervise. (Manage the hire.)

• Give the AI the written job description and 1-3 quality examples to work from
• Run it beside your person on live work, and compare the two outputs
• When it’s wrong, fix the instructions, not the model, then rerun
• Tell it to flag anything it’s unsure of and to immediately ask, never to guess
• Once it’s reliable, connect it to where the work lives, read-only first
• Keep a human approving every output before it counts

What “good” looks like at day 30:

• One job runs AI-first-pass, human-approved
• You trust it enough to keep using it next month
• It hands back real hours, net of the time you spend checking it, measured against the baseline you started with
• You’re still the one making the calls

One job that genuinely works beats ten half-built pilots. That’s the whole difference between MIT’s winning 5% and the 95%.

4. A Real SMB Use Case

A roughly 10-person real-estate valuations firm I worked with had a slow, thankless job buried in their back office. Payments arrived by email, hundreds of them, and someone had to read each one, find the matching account in their Monday.com receivables board, then manually record it. Hours of careful, mind-numbing work, and a wrong match created a mess downstream.

This is the kind of back-office grind MIT found delivers the real returns.

We built them an AI worker for it. The interesting part wasn’t the automation. We onboarded it like an employee instead of deploying it like a tool. A good new hire doesn’t guess in silence when they’re unsure; they flag it and ask. This one works the same way:

• When it’s confident about a match, it makes updates and sets the status as “to approve.” A human reviews it before anything counts.
• When two records look alike, or an amount doesn’t add up, it doesn’t force a guess. It marks the item “needs a human” and leaves a comment to explain why.
• Every item flows through a simple review board: to approve, approved, reviewed-incorrect, needs help. Nothing is hidden. The work is tracked, the way you’d track any team member’s.
• A person on their team reviews and approves.

It was plugged into the same working systems as the team because it’s an automated coworker.

The tools were ordinary: their existing project board, a Make.com connector, and a standard AI model. The build was modest and runs for about $20 a month.

A person handles what AI is bad at: judgment on the edge cases. The AI handles what people are bad at: staying alert across hundreds of repetitive items, cost-effectively. A multi-hour slog became a few minutes of review.

None of this required a foundation-model breakthrough. Just writing down what a good match looks like, and managing the AI like a junior teammate who’s fast, tireless, and needs clear instructions.

5. Guardrails — What an SMB Should NOT Do

Every guardrail here is the same idea wearing different hats: these rules aren’t new, and they aren’t really about AI. They’re the two skills from the start of this chapter, good process and good management, drawn as lines you don’t cross. You already hold these lines with your people. AI just raises the cost of crossing them.

The big one: don’t make the tool your starting point. “We should use AI” is not a plan; it’s the sound the 95% made on the way in. A solution shopping for a problem is expensive suffering. Start in the depths, the problem and the process, and let the tool be the last choice you make.

The rest are those two skills as limits, plus the groundwork they both stand on:

• Don’t automate a process you wouldn’t have been proud to run by hand. (Process.) AI amplifies whatever it touches. Point it at a clean process and it scales the clarity; point it at a mess and it scales the mess, faster. The sloppiness was always costing you. Fix it first.
• Give it access without giving up control. (Management.) Example: we set up Emergency Management Logistics Canada, an emergency-response platform run by a non-technical founder, so AI could edit its website directly, saving hours of copy-and-paste and manual work compared to just using ChatGPT, with two critical rules: it can read but it can never delete anything, and nothing goes live until a human explicitly says so. Full access, zero loss of control.
• Make it escalate, not guess. (Management.) A good hire says “I’m not sure” and asks; a bad one is confidently wrong and quiet about it. Build the first behavior in, the way the real estate firm above did, and give a person the job of catching what gets flagged.
• Don’t skip the foundation. (Foundation.) Everything above is the execution layer, and it assumes the groundwork is already in place. Before you wire AI into real systems and real data, get your data privacy, security, access control, and a simple team-use policy sorted, the job zero from Section 2. The most common trap is the data itself: don’t paste client financials, health records, or anything confidential into a consumer tool without knowing where it goes. Business data uses business-tier accounts; when in doubt, keep it out.

None of it is exotic, and none of it is really an AI skill. It’s the discipline of running a good shop, written down as a few hard limits. The 5% who win aren’t the ones with the cleverest tool; they’re the ones who held these lines while everyone else went shopping for magic.

6. Lessons Learned + One Thing You’d Tell Every SMB Owner

A few lessons from doing this work, each able to stand alone:

1. The 95% didn’t fail at AI. They failed at operations and blamed the AI. They bolted a brain onto a broken process and called the result a technology problem. You skip that whole movie by doing the unglamorous part first.
2. If you can’t write it down, AI can’t do it, and your last new hire couldn’t either. The unclear process was always the bottleneck. AI just stops letting you look away from it. Clarity is the work.
3. Your team already knows what to automate: it’s the work they hate. Ask them what they’d happily never do again and you’ll get your shortlist. It’s different at every company, and it’s almost always the repetitive grind, never the work they take pride in. Automate what they resent and they’ll help you build it; aim at what they enjoy and you’ll fight them the whole way.
4. A 12-person shop can beat the enterprise at this. You can pick one real problem and ship it this month. The corporation needs eleven committees and a CEO mandating AI at teams that aren’t ready. Small and grounded wins the race that hype loses.
5. Measure it. Measure the success rate, measure the original task’s cost and quality, measure the new outcomes. That’s how you know if you’re gaining ground or chasing your AI tail.

And the one thing, if you remember nothing else:

AI hasn’t changed what makes a business good. It has only made the price of broken process much louder.

The work in front of you is the work you already know: run a clean operation, manage your people and your tools well, and stay the one who makes the decisions. Do that, point a capable tool at it, and you’ll land in the 5%. Not by chasing the storm but by anchoring in the depths.

Part of the FrUn SMB AI Guide. Read on canonical guide →