DEV Community

T.M. Gunderson
T.M. Gunderson

Posted on

Your AI Advantage Isn't the Model — It's the Wiring

The first question most small-business owners ask about AI is "which model should we be using?" It's becoming the wrong first question. When the frontier models sit at rough parity, the model stops being the differentiator — every vendor is renting you a very similar brain. What separates a business that gets real work out of AI from a business with an impressive AI org chart is the wiring around the model: what you connect it to, what you refuse to let it judge, and whether you can grab the wheel when it stalls.

The builder who said it out loud

On a recent episode of the TWIML AI Podcast — "The Brutal Reality of Building Agentic Products Today" — Sam Charrington interviewed Dan Jeffries, creator of Nautilo, an open-source platform for personal AI agents. Much of the conversation was Jeffries complaining about his own industry, and the complaints double neatly as a buyer's checklist.

His sharpest one: a lot of the multi-agent "org charts" circulating online are, in his words, cosplay. Agents with role titles talking to other agents, reporting lines, coordination layers — the setup is the content, and the actual output never appears. He extended the critique to demo culture generally: a large share of the "this is insane" agent videos are mostly setup, and the part where the thing does your work never gets filmed.

His tool-design philosophy he calls the "dumb hammer": a hammer doesn't think, and neither should your database or your file mover. If the AI has to be clever to operate a tool, the tool is badly designed. Cleverness belongs at the judgment layer; certainty belongs in the tool. When someone showed off an agent tidying a desktop, Jeffries' response was that a plain shell script is enough for that job.

None of that is model talk. It's all architecture — and it translates directly.

The wiring rules for a small business

1. One agent that knows your business beats an org chart of agents that don't. If you've assembled a research agent, an analyst agent, a writer agent, and a manager agent to coordinate them, ask what any of them shipped last month. If the honest answer is "a very good diagram," you've built the cosplay: structure that performs for an audience of none. Complexity is a cost, and output is the only thing that pays for it. Most small businesses get more from one well-instructed agent with access to their real documents than from a federation of generic ones.

2. Anything a script can do, a script should do. Moving data between systems, renaming files, reformatting the invoice, the every-Monday report — that's deterministic work, and deterministic work is a dumb-tool job. If your AI agent is doing it, you're paying judgment prices for work that needs no judgment, and you're inheriting variance for free. Jeffries compared delegating to agents to a slot machine: every pull lands on a different employee — one who did it wrong, one who says it's done when it isn't, one who almost got it. That variance is tolerable for genuinely fuzzy work — drafting the awkward customer reply, categorizing the odd invoice — and pure downside for work that was never fuzzy to begin with.

3. Keep a takeover path. The moment a tool gets you 90% of the way there, you should be able to finish it yourself in two minutes: open the draft and edit it, fix the file, correct the last row. The alternative is re-prompting — asking the agent to fix its own output, again and again, with a new result each time — and that's the hidden cost that never appears in demos. Even the person running the podcast pays it: Charrington described his own year as roughly 10% doing the actual work and 90% tuning the agents that were supposed to do it for him. When you evaluate any AI tool, look for the takeover path before the feature list. If the only way to correct the tool is to keep talking to it, that's a design flaw you'll pay for every week. It's also a real design choice, not a wish: Jeffries built his own product so that when the agent produces something, you can simply take over and edit it directly instead of prompting and praying.

4. Assume the operator won't debug. On the episode, Jeffries put agent adoption at roughly the 2% mark of the diffusion curve — still bleeding-edge — and told a story about his own best friend, a technically capable person, asking how to even start using agents. For people like Jeffries that question answers itself: talk to it, and when it fails, figure out what went wrong and try again. But most people, he argued, want a button — and they quit at the first error. He compared this era to personal computers in the late eighties and early nineties: you expected them to crash, and knowing how to put them back together was the skill. We are not yet in the iPhone era of AI. So design accordingly: whatever you hand to staff should fail visibly, restart simply, and start from a template. The difference between a tool your team keeps using and one that gets abandoned is whether recovery from the first error is obvious.

The honest boundaries

Three caveats before you treat any of this as law. First, it's one builder's design philosophy, argued from months of building an agent platform — not a measured result, and even sympathetic coverage of the episode notes his own architecture is unproven against the alternatives he criticizes. Treat it as a strong prior. Second, parity is a snapshot, not a permanent condition: models leapfrog, and when one genuinely pulls ahead on your specific work, that's worth knowing — run a quick check with inputs you already know the answers to before believing any leaderboard. Third, the free, out-of-the-box agents from the big platforms make all this sound optional. It isn't. The free tool that almost does what you need is exactly the case where your own wiring — the one integration, the one template, the one takeover path — decides whether it's useful or shelfware.

Tonight

  1. List every AI tool and agent your business touches. Flag any agent whose whole job a script, template, or plain automation already does.
  2. Pick one deterministic task currently done by an agent and move it to a dumb tool.
  3. Check the takeover path on your most-used AI tool: when it's 90% right, can you finish it in two minutes without re-prompting?
  4. If you have a multi-agent setup, ask it for last month's output. If there isn't any, cut to one agent and give it something real to do.

The model you rent will keep changing, and you control none of it. The wiring — what's connected, what's deterministic, where you can grab the wheel — is the part you own, and the only part that compounds.


Want the wiring pre-built? The Boring Automation Pack from SMB Scale Up includes plug-and-play automation blueprints plus the standing-instructions worksheet — the deterministic half of the system, so your AI only has to handle the judgment.

Top comments (0)