A note to the small business that is scared of AI — from one that was. What we built, what it cost, what broke, and the order I'd do it in if I were starting today.
Reps get first refusal; agents catch the rest; a seventh agent reads every call.
Reads, looks up, drafts in the rep's voice, stops. The rep hits send.
A schedule, a lookup or two, a model to summarize, an email at the end.
Audit the catalog, decide the rules, one template per family, generate, check.
Two lists get tangled when people ask where to start: what to build, and what to set up. They're the same ladder read from two sides. Left is what you build at each rung; middle is the plumbing you connect to do it, in the order you actually need it. Climb from the bottom.
| What you build | What you plug in | Cost · risk | |
|---|---|---|---|
| 4 | Agents on the phone — supervisedSix specialists and a seventh who reads every call. Reps get sixty seconds of first refusal; agents catch the rest. | ElevenLabs for the voices and phone handling. Twilio for the numbers. Everything below is already wired, so the agents inherit the same lookups, the same rules, the same ERP hook. | ~$1,500/moCustomer-facing. Months, honestly. |
| 3 | An assistant in the inbox that never sendsReads what comes in, looks up the facts, drafts in the rep's voice, stops. | The Shopify app gets write scopes — draft orders, customers. And this is where the ERP hook goes in: a read-only webhook into NetSuite (or your system of record) for stock, lead times, order status, tracking. Built once; every later agent uses it. | Same toolsHalf-visible. A week to pilot. |
| 2 | Reports that show up on MondayA schedule, a lookup or two, a model to summarize, an email at the end. | n8n (where the workflows live). Resend or any SMTP for delivery. A model API key for the flows that need judgment. A Shopify custom app with read scopes so a flow can see orders and products. | ~$40/moNobody sees it. A weekend. |
| 1 | Product pages a machine can readAudit the catalog, decide the rules, one template per family, generate, check. | Your store export and the assistant from day zero. Nothing to buy. This is also where you write the rules list — discounts, origin, what each option means — that every rung above reads from. | $0Nobody sees it. An afternoon per family. |
| 0 | Pick the AI partner you'll build withClaude, or ChatGPT with code execution. Whichever one you work best with. | This isn't rocket science. You don't need to know how a webhook works; you need to be able to describe your problem, answer the questions that come back, and test what you're handed. The partner writes the code. You make the calls. | $20–$200/moYou probably already have it. |
Each connection shows up the first time a rung needs it, not before. The Shopify app arrives at rung 2 read-only because the sync checker needs to see orders; it gets write access at rung 3 because the inbox assistant creates draft orders. The ERP hook comes in at rung 3, not 4, so that when the agents arrive they find it already working.
A page, a report, a draft, or an agent can only say what you can prove. Every connection above exists to give the AI a fact it can look up instead of one it would guess. The order matters less than the direction. Start at the bottom. Don't skip.
I want to say that first, because it is the whole point of this document. I am not a developer. I am not a programmer. I run marketing and e-commerce at a 62-person millwork company in Arkansas, and a little over a year ago I could not have explained what a webhook did. Everything in the pages that follow was built by that person.
I started because of a simple problem. Our phones and our people are in Arkansas. Our customers are everywhere, and California — our biggest state — is two hours behind us. A designer with a question at 7pm her time, or on a Saturday, simply could not reach us. I believed that was quietly capping our growth. I still do.
The conventional fix — a bigger call center, an offshore service, an agency-built platform — was never on the table for a company our size. So I did the unconventional thing. I built it. Not alone: with an AI as the engineering partner, describing the problem, making the judgment calls, testing what came back until it worked. A year later we have seven AI agents answering the phones around the clock in two languages, a library of automations that brief the sales team every morning, three storefronts rebuilt from the ground up, and a supervisor agent that reads every call the others take.
This is not that story. That story is forty pages and it's called From a Thought to Reality; you can read it at the link on the cover. This is the shorter thing I wish someone had handed me at the start: what it actually costs, what order to do it in, and where the fear is misplaced.
People assume this is expensive. It isn't. It's the cheapest thing I've ever built, and I've built catalog mailers. Here is everything we pay for, in the order you'd add it.
| Layer | What it is | Roughly | When you need it |
|---|---|---|---|
| Product pages | Rebuilding your pages so machines can read them | $0 | Day one. Nothing to buy. |
| An AI assistant | Claude, or ChatGPT with code — the engineering partner | $20–$200 / mo | Day one. You probably have it. |
| Automation | n8n — where the workflows live | $20–$60 / mo | Rung 2. The first Monday email. |
| Model API | The "thinking" inside a workflow — pennies per run | $5–$50 / mo | Rung 2, when a flow needs judgment. |
| Transactional email | Resend or similar — how reports get delivered | $0–$20 / mo | Rung 2. |
| Voice platform | ElevenLabs — the agents' voices and phone handling | The bulk of it | Rung 4. Not before. |
| Phone numbers | Twilio lines routed into the agents | A few dollars each | Rung 4. |
The whole voice system — six agents, the supervisor, phone lines, models, automation — runs about $1,500 a month, which came to about $2.50 per conversation over our first twelve weeks. In the first five days, the model cost for the quote-building side was $86.94. Everything below Rung 4 is a rounding error next to that.
Each rung is safer than it sounds and more useful than it looks. Each one teaches you what you need for the next. And nothing on the first three touches a customer.
Every week someone asks ChatGPT or Perplexity where to buy what you sell. The assistant answers by reading web pages. If your page has a real table of sizes, a labeled price, and a few lines of invisible code that say "this is a Product," it gets cited. If it has a paragraph that says "premium quality craftsmanship," it gets skipped.
We rebuilt 7,754 product pages across three storefronts this way. Not by writing 7,754 pages — by writing one template per product family and letting each product's own data fill it in. The AI read our catalog export, audited it, asked me the questions it couldn't answer from data, built the templates, generated the pages, and produced the import files. Along the way it found four trim pages describing the wrong finish, nineteen products priced at $0, and 704 pages quoting the wrong trade discount. Those had been live for months.
Why start here: it can't break anything, it makes you look at your own data honestly, and it teaches you the rhythm — audit, decide the rules, template, generate, check — that every later rung uses. The full how-to, with the prompts, is in the starter kit.
This is the rung I'd tell everyone to start with if they'd only do one. An n8n workflow with a schedule, a couple of lookups, maybe a model in the middle to summarize, and an email at the end. No app to open. No dashboard. It just arrives.
Ours: every Monday at 9am, a flow researches four competitors and lands one briefing in the whole team's inbox. Three mornings a week, another one sends an industry trend review — association events, design trends, where our products are showing up. Every weekday, each rep gets a personal summary of where they stand. And the simplest one, which may be the highest-leverage automation we run: a daily check for orders that never synced to our ERP, which used to mean orders that quietly didn't ship.
The lesson in this rung: knowing which job needs intelligence and which just needs reliability. The competitor review uses a model, because it has to read and summarize. The sync checker uses no AI at all — two queries and an email — because reliability is the whole point. Most of the craft is telling those two apart.
The next step points the same tools at a rep's email. The assistant reads what comes in, does the lookups — order status, stock, pricing — and leaves a finished draft in the rep's own Drafts folder, written in their voice, with a note on top saying what it checked and what to verify. Then it stops. The rep reads, edits, and hits send. Or doesn't.
That rule — drafts only, never sends — is what makes Rung 3 safe. The assistant cannot email a customer. It preps the shot; the rep takes it. On the first day of live testing, asked to quote a product that didn't exist, it refused to invent a price and flagged it for a human. That was the moment I trusted it.
We proved it on one inbox first, measured what share of the mail it could handle, then mirrored it to the rest of the team. One engine, five voices. Ten minutes of setup per rep.
This is also the rung where the ERP hook goes in — a read-only line into the system of record for stock, lead times, order status and tracking. We built ours in an afternoon, without buying a seat, and it's the piece every agent above uses. The kit walks through it.
This is the rung people picture when they hear "AI," and it's the one to do last. Ours is a team of six specialists — reception, quotes, trade, overflow, service, design — plus a seventh who reads every call the other six take and emails management a review every Monday. They build real quotes, create real Shopify draft orders with trade discounts applied, route real leads by territory, in English and Spanish, at any hour.
Two rules made it work. Collaborate, don't replace: during business hours the human reps get sixty seconds of first refusal on every call; the agents catch what the reps can't, and after hours they carry the phones alone. Supervise by design: the seventh agent reads full transcripts, not samples, and flags what a manager would flag. In its first week it cleared an agent of a second-hand complaint — the record showed the call was handled correctly. Supervision defends the agents as often as it corrects them.
Twelve weeks in, before we'd promoted the lines to a single customer: 1,710 conversations, a 90.8% success rate, a 4.8 customer rating. One call in eleven needs work. We know which ones, because someone reads every transcript.
I won't pretend it was smooth. The pattern behind every failure was the same: the data was ambiguous and the AI resolved the ambiguity by itself. The fix each time was a human answer, a check, or both.
| What happened | What we do now |
|---|---|
| A voice agent heard "Ryse Construct" as "Arise Construction" on a $24K quote | Agents spell company names back phonetically before submitting anything. |
| Inch-marks in product titles broke the data going into Shopify | Rules in the prompt and a sanitizer in the code. Defend at every layer. |
| An agent said "done" before the tool had actually succeeded | Confirm only after success. Optimistic confirmation erodes trust the moment something fails. |
| 704 product pages quoted a discount that had changed months earlier | Numbers that change live in one place. Pages get regenerated, not edited. |
| A resin option was described as "a cast version of the same profile" because nobody told the AI what it was | Answer the questions before generating. Every answer changes thousands of pages. |
| A corner block linked to the wrong crown — the SKU suffix was a different size | Match on the exact SKU. Spot-check ten random pages before generating all of them. |
| One site's schema told Google the products lived on a different domain | A machine-readable claim is still a claim. Check the URLs. |
If I'm honest, the reason I made it through is that I never told myself how hard it would be. Being naive was a blessing. Had I known at the start everything it would take, I might never have attempted it.
What surprised me most was my own patience — and that I refused to accept failure. The constant breaking, over and over, the same thing failing in a new way each time. And then it works, and a customer calls at 8pm on a Friday for a moulding quote and gets one, and Monday morning there's a clean draft order in Shopify and a routed lead in the right inbox. The routine is covered. The relationships are still ours.
Six months from now, newer tech may well replace what I've built. That's fine. What I keep is what I learned: how to actually use AI to solve real-world problems. That doesn't get replaced.
I believe we are at the beginning of something extraordinary, and I believe it's no longer gated by company size, headcount, or budget. It's gated by the decision to start. A four-person department at a family millwork company put a governed, revenue-producing AI team into production. If we can, you can.
Here is what I think the goal is for every small company, ours included: to be accessible 24 hours a day, 365 days a year. Not because a machine should replace the people who know your customers — because the customer with a question at 7pm should get an answer, and the one who calls on a Saturday should not have to wait until Monday. The companies that get there first will win a loyalty that's very hard to take back. That's the whole race. And the first step in it is small.
Start with the Monday email. Everything else follows from it.