Most mid-sized businesses we talk to are in the same spot. Leadership knows AI matters, a few employees are already using chatbots on their own, and nobody is sure what the business should actually do first. The good news is that the first step is smaller and more concrete than most people expect.
Start with work, not with tools
The most common mistake is starting with a tool: "We should get an AI assistant" or "Let's try an AI chatbot on the website." Tools are easy to buy and hard to get value from if you don't know what problem they're solving.
Start instead with the work your team does every week. Where do hours disappear? Which tasks do people complain about? What gets done late because nobody has time? Those answers are where AI and automation earn their keep.
A useful exercise is to ask each department head to list the five tasks their team repeats most often, roughly how long each takes, and how painful it is when it goes wrong. You'll usually find the same patterns: copying information between systems, answering the same questions, chasing people for documents, and building reports by hand.
Find the right first workflow
Your first project should be chosen for learning as much as for impact. Look for a workflow that is:
- Frequent. It happens daily or weekly, so improvements add up and you get feedback quickly.
- Mostly rule-based. A new employee could learn it from a written checklist, even if some inputs are messy.
- Measurable. You can count how long it takes, how many there are, or how often mistakes happen.
- Low risk if something goes wrong. An error is annoying, not a legal or safety problem, or a person reviews the output before it matters.
- Owned. Someone on your team cares about it and has the authority to change how it's done.
Good first candidates often include sorting and routing incoming email, pulling data out of documents into a system, drafting routine replies, or generating a weekly report automatically. Poor first candidates include anything that makes decisions about customers' money, health or legal rights without review.
Write down how it works today
Before changing anything, document the current process in plain language: who does what, in which system, and how long it takes. Watch someone do it if you can. You'll often find steps nobody mentioned and workarounds that exist for good reasons.
Then record a baseline. It doesn't need to be precise. "About 40 invoices a week, roughly 6 minutes each, with a few errors a month that take an hour to fix" is enough to judge whether a change helped. Without a baseline, every project ends in opinions instead of evidence. Our guide to measuring AI ROI honestly goes deeper on this.
Pick the simplest thing that could work
Once you know the workflow, work up this ladder and stop at the first rung that solves the problem:
- Use what you already have better. Many CRMs, accounting systems and office suites include automation and AI features that nobody has turned on.
- Add an off-the-shelf tool. For common problems, a proven product is usually faster and cheaper than building.
- Connect systems with automation. Much of the value comes from moving data between tools without retyping it. AI helps where the inputs are unstructured, like emails and PDFs.
- Build something custom. Worth it when the workflow is specific to your business and valuable enough to justify ongoing ownership.
Notice that "AI" isn't a rung by itself. Sometimes the right answer is a simple rule-based automation with no AI at all. That's a win, not a failure. See build vs buy for how to make the call.
Put guardrails in before you launch
A little structure up front prevents most of the problems people worry about:
- A short AI use policy that says which tools are approved and what data can go into them. Our guide on writing an AI use policy includes an outline.
- A human checkpoint wherever an AI output reaches a customer or changes an important record, at least until you trust it.
- Business-grade accounts for any AI service that touches company data, with data terms you've actually read.
- A way to switch it off and fall back to the manual process if something goes wrong.
Run a small pilot with a finish line
Treat the first project as a pilot with a defined end date, usually a few weeks. Agree before you start what success looks like, for example "cuts time per invoice by half with no increase in errors," and who decides whether it passed.
Involve the people who do the work from the beginning. They know the edge cases, and they're the ones who will make it stick or quietly work around it. Collect real examples of where the system got things wrong; those examples are the fastest way to improve it.
At the end of the pilot, make one of three decisions: roll it out, adjust and run another round, or stop. Stopping is a legitimate outcome. You've learned something for a small cost.
Train people, then expand
When a pilot succeeds, the temptation is to immediately launch five more projects. Resist it long enough to train the team properly and write down how the new process works. Adoption, not technology, is where most AI projects stall. Our guide to training your team on AI covers what works.
Then go back to your list of workflows and pick the next one. Each project gets easier because your team, your data and your guardrails are in better shape than they were.
A 30-day starting plan
- Week 1: Ask each department for its most repetitive tasks. Pick one workflow using the criteria above and name an owner.
- Week 2: Document the current process, record a baseline, and write a one-page AI use policy.
- Week 3: Choose the simplest approach that could work and set it up with a human checkpoint.
- Week 4: Run it on real work, collect errors and feedback, and decide whether to roll out, adjust or stop.
If you'd like a second opinion on which workflow to start with, that's exactly what our AI readiness assessment is for, and a free call is a good place to start.