Where to Start with AI in Your Business: A Guide

The question "where to start with AI in your business" is asked badly more often than it's asked well. I've heard it phrased as "what AI tool should we buy", "should we hire an AI person", "what's the killer AI use case", and a dozen other variants. None of those is the right opening move.

The right opening move is to pick one workflow that's slow, repetitive, and doesn't touch a customer in real time, and ship an agent for it in eight weeks. That's the answer to the question. The rest of this post is why, what to look for, and what to skip.

The frame: don't start with AI, start with a workflow

Most SMB owners I talk to want to start by deciding which AI tools to buy. That's working backward. The tool choice is downstream of the workflow choice. If you pick the workflow first, the tool choice is usually obvious or unimportant. If you pick the tool first, you'll spend two months trying to wedge it into a workflow it doesn't fit, then quit.

So the opening question is not "what AI should we buy". It's "where in our business does someone spend a lot of time doing something repetitive, where the rules are knowable, where a mistake doesn't end us, and where the team would actually use the help".

Three out of four of those constraints have to be met before AI is even a sensible answer. I've written about the full roadmap exercise elsewhere. The version that fits in this post is shorter.

What "ready to start" actually looks like

The conversation gets clearer when an owner can answer these three questions about a candidate workflow.

How many hours per week does someone in the business currently spend on this? If the answer is "I'm not sure, maybe a few", you're not ready. The hours number is your North Star metric. You can't measure success without it.

Who in the business will own operating the agent after it ships? If the answer is "we'll figure it out", you're not ready. AI agents need a human owner for the first quarter, every quarter. Five to ten hours a week of attention, then one to two hours a week ongoing. If no one has that capacity, you have a hiring problem before you have an AI problem.

What happens if the agent gets it wrong 5% of the time? If the answer is "we lose a customer" or "we get sued", pick a different workflow. Internal triage and document extraction have low-stakes failure modes. Customer-facing real-time chat does not.

If you can answer all three for a specific workflow, you're ready to start. If you can't, the time spent answering them is more valuable than the time spent building anything.

I cover a more structured readiness assessment if you want the longer version of this gate. The short gate above gets most owners 90% of the way there.

The shape of a good first project

The first AI project that works in an SMB has a specific shape. I've seen this shape enough times that I'd bet on it.

It reads from one place. Inbound emails, inbound leads, inbound documents. One input stream. Not three.

It writes to one place. A queue, a category, a CRM field, a draft folder. One output. Not a fan-out across systems.

It runs asynchronously, not in real time. The team can review outputs at the start of their day. There's no SLA tied to milliseconds of latency.

It produces an output a human will review before action. Not autonomous. Not "AI sends the email and we'll see how it goes". A human looks at the agent's classification, summary, or draft and clicks confirm.

It has a measurable baseline. Today the team handles 200 of these items a week, spending 14 hours, with a 4-hour average response time. The agent's job is to take the 14 hours to 6 hours within 60 days.

This shape sounds boring. That's the feature. The boring first project ships, produces real hours-saved numbers, and earns the team's trust before you move to the more ambitious second project. Most SMBs that start with an exciting first project end up retreating from AI entirely after the first project goes sideways.

Concrete first-project candidates

To make this real, the patterns I recommend most often as a first project, in rough priority order.

Lead triage. Inbound leads get classified by intent and quality, enriched with public data, routed to the right salesperson with a pre-drafted reply. Saves SDR time, improves response speed, no customer-facing AI until the human approves the draft.

Inbound document extraction. Vendor invoices, customer contracts, expense receipts, shipping forms. Extract structured data into your accounting or ERP system. Human reviews exceptions. Saves AP and ops time.

Support ticket classification. Tickets get tagged by category and severity, drafted reply suggested for common cases, human reviews. Cuts triage time, improves first-response speed, leaves the actual reply in human hands until the team trusts the agent.

Internal knowledge lookup. Team members ask questions of internal docs (policy, product specs, past customer notes) through a chat interface. Saves the experts' time answering the same questions repeatedly.

Notice what's not on the list. No customer-facing chatbot. No "AI that handles all of customer service". No content marketing generation. These are all real projects that work in production, but they are not first projects.

What to skip while you're starting

A short list of things SMB owners try first that I'd skip.

Don't start with a customer-facing chatbot. The risk profile is wrong for build one. Save it for build three or four when you've learned how your agents fail.

Don't start with content generation as the headline project. The quality bar is high, the brand risk is real, and the team's job becomes editing AI slop, which most writers find more painful than writing from scratch.

Don't start by hiring an "AI consultant" who wants a six-month engagement before shipping anything. The good ones can show you a working agent in eight weeks. The ones who can't are selling you slide decks.

Don't start by signing a multi-year contract with a platform vendor. You'll learn more about what your business needs from one 8-week build than from any vendor pitch. Get the build done first, then evaluate platforms.

Don't start by trying to be product agnostic about model choice on the first build. Pick one model that's appropriate for the task and ship. You can swap later. Most SMBs that try to be tool-agnostic on day one spin in evaluation paralysis.

The 60 to 90 day version

For an owner who wants the calendar, the version I sketch most often.

Weeks 1 to 2: pick the workflow. Score the candidate list. Confirm the operating owner has capacity.

Weeks 3 to 4: scope the build. Define inputs and outputs. Pick build vs buy (I cover the build vs buy decision separately). Get access to source systems.

Weeks 5 to 8: build, integrate, test in a sandbox.

Weeks 9 to 10: pilot against real production traffic with human review on every output. Measure error rate against the baseline.

Weeks 11 to 12: tune. Hand off. Document. Pick the next candidate.

If you're working with a partner, the build phase compresses to weeks 5 to 8 and the pilot starts week 7. If you're building yourself with no prior AI experience, give yourself the full 12 weeks and don't rush.

A word on starting too small

Starting boring doesn't mean starting tiny. An agent that saves 30 minutes a week is not worth building. The team won't care, the math won't work, and you'll lose momentum after the first ship.

Target a first project that saves at least 8 to 12 hours a week of real work, with a clear baseline you can measure against. Below that, you're doing a demo, not building a capability. The demo is fine if your goal is to learn, but be honest about it. Don't burn $20,000 on a demo and call it a first agent.

The right starting move is unglamorous, measurably useful, owned by a specific human, and shipped in under three months. Anything more ambitious is build three. Anything less ambitious is a learning exercise dressed up as a project. The middle is where SMBs win with AI.