Let's talk
All posts
AI & Automation

AI Automation for Real Businesses: What Actually Works

2025-02-18 · 8 min read read

There’s a version of the AI automation conversation that’s exhausting. It involves a lot of talk about agents, superintelligence, and 10x productivity gains that somehow never materialize into anything you can point to in a spreadsheet.

I work with growing businesses in the $1M–$10M range, and I want to give you the more useful version of this conversation: what’s actually working, what’s hype, and how to think about where to start.

The honest framing

AI automation isn’t magic, and it isn’t a silver bullet. It’s a set of tools — some mature, some half-baked — that can meaningfully accelerate specific types of work. The key word is specific.

The businesses that are getting real value from AI automation right now have done one thing correctly: they’ve identified a concrete, bounded process that was previously manual, and they’ve automated that specific thing. They didn’t try to “implement AI” — they automated a process.

That’s the mental model shift. Not “how do we use AI?” but “which specific repetitive process is costing us the most time or money right now?”

Where it’s working

1. Content workflows

If your business produces regular content — blog posts, social copy, email campaigns, product descriptions, proposal templates — AI is legitimately useful here. Not to replace a writer, but to change the shape of the work.

The workflow that’s working: brief → AI generates a structured first draft → human edits for voice, accuracy, and nuance → publish. The AI handles the blank page problem and the structural skeleton. The human handles everything that requires judgment.

For a business producing 8–12 pieces of content per month, this typically cuts production time by 40–60%. That’s real.

2. Data extraction and classification

Any process that involves a human reading something and extracting or categorizing information is a strong automation candidate. Examples:

These were technically possible before LLMs, but they required specialized ML models that were expensive to build and maintain. Now they’re accessible with a prompt, an API key, and a tool like n8n or Make to wire it together.

3. Outbound research and personalization

For sales teams running outbound, the research-per-prospect step is a notorious time sink. An automation workflow that takes a prospect name and company, pulls relevant recent information (news, job postings, LinkedIn summary), and produces a paragraph of personalized context is achievable today. Not perfect — it still needs human review — but it changes the ratio of time spent on research versus time spent on actual outreach.

4. Internal knowledge retrieval

If you have documents, SOPs, or a knowledge base that your team references regularly — and that knowledge is scattered across Google Drive, Notion, and email threads — there’s real value in a retrieval-augmented system that lets people ask questions in plain English and get sourced answers. This is more infrastructure work than the previous examples, but for organizations with genuine knowledge management problems, it solves something expensive.

Where it mostly isn’t working

Complex, multi-step decision-making

“Let the AI handle customer escalations” sounds good until you’ve seen what happens when it handles a customer escalation badly. For anything where a wrong call has real consequences — client communication, hiring decisions, contract negotiation — AI is a research and drafting assistant, not a decision-maker. Treat it accordingly.

Processes you haven’t documented

This is the most common mistake. A business tries to automate a process that lives entirely in one person’s head. The automation fails — not because of an AI limitation, but because the process wasn’t defined clearly enough to encode. If you can’t write a clear step-by-step for how a human would do it, you’re not ready to automate it.

High-stakes, low-volume work

The ROI calculus for automation depends on frequency. Automating a report that gets run twice a year is a poor use of effort. Automating a check that runs 200 times per day is obviously worth it. Most decisions are somewhere in between — but the frequency question should drive the priority list.

How to start

Step 1: Make a list of the 10 most repetitive processes in your business. The ones that have a predictable input, a predictable output, and that someone on your team does over and over. These are your candidates.

Step 2: Score them on two axes: time cost (how many hours per week does this consume, across everyone who touches it?) and failure risk (what’s the worst case if the automation makes a mistake?). High time cost + low failure risk = automate first.

Step 3: Pick one. Not three. One. Build it, run it for 60 days, measure the time savings. Once it’s running reliably, pick the next one.

Step 4: Use tools that match your technical sophistication. n8n and Make are where most businesses should start — they’re visual workflow builders with pre-built integrations for the tools you already use. Zapier is fine for simple one-step automations. Custom code (Python scripts, AWS Lambda) is appropriate when you’ve outgrown the visual tools for a specific workflow.

The stack that’s worth learning

If you’re a founder or ops leader who wants to understand what you’re buying when you hire someone to build this, here’s the landscape:


The honest summary: AI automation is a real capability, not a buzzword. But the businesses getting value from it are treating it like any other operational improvement — specific, scoped, measured. Not “implementing AI,” but solving a problem.

If you want to map out where automation makes sense for your specific business, that’s exactly what a discovery call is for.

If this resonated, let's talk.

30 minutes.
No pitch.

An honest conversation about what you're building and whether we can help.