Build the First 80% of Your AI AgentsYourself. Hire an Engineer for the Last 20%.

A Shopify founder turned down our AI proposal this week after building the first 80% of her agents in-house, and she was right to do that part herself. Basic AI setup is now DIY for any business owner, which means you should only pay an engineer for the last 20% where agents touch live customer data, money, and production keys.

Why basic AI setup is going the way of vibe coding

My YouTube feed is full of videos claiming AI consulting is dead. Six months ago, setting up an agent meant fighting developer frameworks like OpenClaw and Hermes just to get a basic workflow running. Today anyone can open Meta's Muse or Grok Bot and have a working agent in minutes.

I saw that shift firsthand this week. We pitched a founder who runs a growing wellness brand on Shopify. Before our first call, she had already built a company brain from her brand docs, connected Gorgias for customer support tickets, used Parker for ad creative and analytics, and wired Claude into Klaviyo.

When we sent a proposal for a full multi-agent system across marketing, customer support, and operations, she declined the marketing build because she was almost at the finish line herself. She only left the door open for operations and multi-pack inventory math, where 1-pack and 3-pack SKUs make demand forecasting and purchase orders tricky.

Looking back, her instinct was spot on. When an AI tool is only drafting ads, emails, or support replies for a human to review, you do not need to pay an engineer. Off-the-shelf tools let any founder build that first 80% in-house.

Where the last 20% breaks DIY agents in production

The danger starts when founders treat the last 20% the same way they treat a vibe-coded landing page. That last 20% is the line where agents stop drafting text and start touching live customer data, bank feeds, inventory counts, and production keys.

When an agent has write access to your live systems, small gaps turn into expensive failures. With recent package compromises and this week's Railway security incident forcing teams to rotate their stored database and login keys, a fragile setup can do real damage to a business.

I saw that exact line this week while running a production-readiness and security audit on an app another founder built with a Replit agent. Out of 1,421 commits in the repository, 1,197 were written by the Replit agent. Building the prototype himself saved him months, and on the surface the app looked clean and worked great in a single-user demo.

Under the hood, he was smart to call me before launching to real users. During the audit, I found three categories of production blockers that the coding agent missed:

• Two users overwriting each other's data: Because of how whole-list and offline sync was wired, if two people on the same account used the app on two phones at once, the second phone's sync overwrote and deleted the first person's additions. Even worse, if the app opened while offline, it could upload blank data 1.5 seconds later and wipe out the account's real records.
• Exposed master keys and screen-only login checks: Earlier in the project, the coding agent had pasted a master database key into a config file where it sat across 46 branches. On top of that, two-factor login was checked only on the phone screen, not on the server, so a password alone could read transactions through the API.
• Unfiltered prompt injection and math drift: Across 64 realistic household budget scenarios I tested against the app's calculation code, 38 came back wrong, often telling a family they could safely spend $1,500 or $1,880 more than they actually had. In offline tests, a bill name written like an instruction also reached the AI assistant unfiltered.

None of those issues required throwing away his app. We retired the old database key, locked down the server checks, and are fixing those exact gaps right now so he can launch safely.

Five questions to ask before hiring an AI builder

In a world where anyone who uses ChatGPT calls themselves an AI expert, you do not hire someone to write prompts. You hire them for the last 20%. Ask these five questions before hiring an AI builder for your business:

1. What happens if two people or two agents update the same record at once? A demo builder only tests one user on one screen. A production engineer explains how they prevent a second sync from overwriting and deleting the first person's data.
2. Where do my secret keys live, and what breaks if a host like Railway gets breached? Listen for server-only secret storage, least-privilege database roles, login checks enforced on the server, and a clear plan to rotate keys in minutes if a hosting provider has a security incident.
3. Which actions can the agent take alone, and which pause for human approval? Drafting an email or ad is safe. Sending money, issuing refunds, or modifying live customer records should start in draft-only mode with a human approval gate.
4. How do you stop an incoming email or file from hijacking the agent's instructions? Outside emails, web forms, and bank merchant names come from untrusted sources. A real builder treats outside text strictly as data, never as system instructions.
5. When do you use regular code instead of AI? Language models guess. If a candidate uses an AI model to calculate account balances, inventory counts, or user permissions instead of deterministic regular code, walk away.

Three takeaways for business owners

• Build the first 80% yourself when AI is only drafting content, ads, or support replies for a human to review.
• Bring in an engineer before an agent gets write access to money, inventory math, or live customer data.
• Vet AI builders on production safety, not demo speed, using the five questions above.

Every founder can build working AI agents today. The real advantage is knowing where DIY ends and engineering begins for your business.

Where do you draw the line between what you build with AI and what you hand to an engineer?