Ryan Lewis | AI Automation

I break down practical AI automations for online businesses.
Founder @BackendBrilliance (AI that books jobs)
New: The Practical AI Automation Blueprint ↓

[Ryan Lewis | AI Automation @Mr_RyanLewis]

Most AI automations don't fail because of the AI.

They fail because nobody designed the control.

Here's the 6-stage operating model I run every workflow through before it goes live: 🧵

[Ryan Lewis | AI Automation @Mr_RyanLewis]

@MyraCodes_ the part people skip is what happens between attention and money — most businesses lose the sale in the follow-up gap. that's usually the first thing worth automating, not the marketing.

[Ryan Lewis | AI Automation @Mr_RyanLewis]

The second-biggest leak after missed calls: the quote that goes out and dies in the inbox.

Here's the fix I built: estimate sent, no reply in 48 hours, they get a text. Not a nudge to call the office — a specific next step: "The quote for the maple removal is ready whenever you are. Want me to hold a spot this week?"

Most businesses follow up once or never. Two follow-ups is where the bookings start.

[Ryan Lewis | AI Automation @Mr_RyanLewis]

You're already asking the right question in the replies. It was never about waiting on the slow model — it's whether the call needed it at all. Cheap model plus deterministic checks covers most production traffic; opus is for the few calls where judgment actually moves the needle. What's the one task you're still not handing to sonnet?

[Ryan Lewis | AI Automation @Mr_RyanLewis]

The trust question gets smaller when you split it. Fetch, filter, route are deterministic — verifiable, no trust needed. The LLM is the only part that needs watching, so you shrink its surface area. My rule: agents run the pipeline, humans get a ping only on judgment calls. What are yours touching — data or decisions?

[Ryan Lewis | AI Automation @Mr_RyanLewis]

The pattern is everywhere, not just Wayland. People blame the model or the tool; the failure is almost always in the workflow scaffolding around it.

The MCP server is the interesting part. A screen recorder is fine, but giving agents a real interface into the workflow beats making them scrape pixels any day.

[Ryan Lewis | AI Automation @Mr_RyanLewis]

@JamesMalsawm Good picks. Of the two, customer support is the stickier business. Lead gen lives or dies on lead quality, which is mostly out of the builder's control. Support has a clean metric: tickets resolved without a human touching them. Which one are you building for right now?

[Ryan Lewis | AI Automation @Mr_RyanLewis]

Honest question for anyone running AI in their business: what's the one task you'd pay $500 a month to never think about again? Not "marketing" — name the actual thing. I'll tell you whether it's automatable.

[Ryan Lewis | AI Automation @Mr_RyanLewis]

@WukongNumber1 Exactly. The naming-the-job part is the whole trick. A generic "sorry we missed you" reads as spam and gets ignored. A text that names the actual service gets answered because it proves someone's actually paying attention.

[Ryan Lewis | AI Automation @Mr_RyanLewis]

Missed calls are the most expensive leak in a local business.

Here's the fix I built: someone calls, no one picks up, they get a text within a minute. Not 'sorry we missed you' — it references the actual job: 'Were you calling about the tree removal in Weirton? I can get you an estimate this week.'

Generic follow-up gets ignored. Specific follow-up gets booked.

Phone webhook → LLM → SMS. Four steps, a few dollars a month to run.

[Ryan Lewis | AI Automation @Mr_RyanLewis]

Genuine question for business owners: how many calls do you think you missed this week that you'll never know about?

Not the voicemails. The ones who never left one.

[Ryan Lewis | AI Automation @Mr_RyanLewis]

The arbitrage doesn't disappear, it moves. When everyone has the same tools, the edge shifts to execution speed and distribution. Same models, same APIs — but one team ships and iterates daily while another ships quarterly.

I've seen the same automation produce wildly different numbers based purely on who runs it. Access is equal now; the bottleneck is implementation.

[Ryan Lewis | AI Automation @Mr_RyanLewis]

The second bullet is the one that matters for real deployments. An agent that loses its identity across reconnects re-announces itself and re-asks for context every time — that's the difference between a bot you can hand to a client and one you have to babysit.

Identity continuity plus a compacted memory log is what turns a demo into something that survives production. The speed tiers are nice; surviving a cold start without amnesia is the actual feature.

[Ryan Lewis | AI Automation @Mr_RyanLewis]

The sorting rule I use: SOUL.md is who the agent is, USER.md is who it's talking to, MEMORY.md is what has happened. Three different jobs.

The failure mode I see in practice isn't misfiling, it's that MEMORY.md has no compaction rule. It grows into a junk drawer of stale context the agent reads but can't trust. A weekly summarize-and-prune pass matters more than which file things land in.

[Ryan Lewis | AI Automation @Mr_RyanLewis]

@tonysimons_ The automation people ask for isn't really automation — it's reliability. Blueprints get you to demo. The real work is run #200 working as well as run #1, and no blueprint covers that: retries, idempotency, escalation when the lookup returns garbage at 2am.

[Ryan Lewis | AI Automation @Mr_RyanLewis]

This is a solid loop — you basically built gap detection into your workflow. One upgrade worth trying: make those rules ranked and explicit. When two course lessons conflict, the agent needs to know which one wins. I keep mine as numbered ranked rules in a single system prompt; without that, an agent with conflicting instructions defaults to the most restrictive one and starts refusing.

[Ryan Lewis | AI Automation @Mr_RyanLewis]

Fun comparison, but for anyone running real workloads: day-zero vs day-10 matters less than whether your workflow survives the update. I keep a fixed set of my actual client flows and rerun them against every new model — same prompts, same scoring, twenty minutes. That's what tells you if the update is safe to ship.

[Ryan Lewis | AI Automation @Mr_RyanLewis]

@tonysimons_ Control is the whole game. But the flip side: the more freedom the agent has, the more explicit the guardrails need to be. In my client booking flows the most capable agents run under the tightest ranked rules — complete control without them quietly turns into complete liability.

[Ryan Lewis | AI Automation @Mr_RyanLewis]

@CaptainJeromeFr Good. One thing to watch for: provider model updates quietly change how strictly mid-prompt ranked rules get honored. Re-run your hardest test case after every update — that's when this setup breaks silently.

[Ryan Lewis | AI Automation @Mr_RyanLewis]

Exactly. And the control is where most of the real work hides — in my booking flows the model is maybe 20% of the build. The rest is routing: which replies the cheap model can send, which ones escalate to a human, what happens when the CRM lookup returns garbage. Nobody screenshots that part, but it's what keeps the thing running in production.

[Ryan Lewis | AI Automation @Mr_RyanLewis]

Most AI automations don't fail because of the AI.

They fail because nobody designed the control.

Here's the 6-stage operating model I run every workflow through before it goes live: 🧵

[Ryan Lewis | AI Automation @Mr_RyanLewis]

Most AI workflows burn money on the wrong model. The cheap one does the grunt work — extracting, summarizing, drafting. The expensive model only steps in for the final decision. One routing change took my API cost per run from ~$0.40 to ~$0.06, same output quality.