The 5 AI Projects Every SMB Should Run Before Hiring a Consultant

I talk to a lot of business owners who want to "do AI" but don't know where to start. My honest advice, before they spend anything: run these five experiments first. They cost nothing but time, they'll show you where AI actually fits your business, and they'll make you a much better client when you do hire someone.

Each experiment takes 30–90 minutes. All you need is a ChatGPT or Claude account (free tier is fine for most of these).

Experiment 1: Give your most repetitive writing task to Claude

Pick the single most repetitive writing task your team does. Proposal introductions, job postings, customer follow-up emails, social media posts, weekly reports — whatever your team does over and over with slight variations.

Write one example yourself, then ask Claude or ChatGPT to produce five more variations in the same voice. Then ask it to produce ten.

What you're measuring:How much editing does the output need? Could a junior team member use this to produce a first draft that a senior person quickly polishes? If the answer is yes, you've just found a real automation opportunity.

What this reveals: Whether your writing has enough consistent structure for AI to replicate. If AI output is all over the place, your actual process is also inconsistent — and that needs to be fixed before automation, not after.

Experiment 2: Build a rough chatbot with a custom GPT (no code)

Go to ChatGPT, create a custom GPT, and give it a system prompt that describes your business and how you handle customer questions. Upload your FAQ page or a few pages of your website. Then ask it the ten most common questions your team gets.

This takes about an hour and no technical skill.

What you're measuring: Does the AI answer accurately and in your voice? Which questions does it get wrong or refuse? Are the gaps because your documentation is thin, or because the questions require real judgment?

What this reveals:Your documentation quality. Most businesses discover their FAQ content is full of implicit knowledge — answers that assume the reader knows things they don't. AI surfaces these gaps immediately. You'll also discover whether customer questions are actually answerable by information or require relationship and judgment. The former is automatable; the latter probably isn't.

Experiment 3: Have AI sit in on a week of your inbox

Export or copy a week's worth of customer emails (anonymized) and paste them into Claude. Ask it to: categorize them by type, identify the top 5 most common requests, draft template responses for the top 3, and flag which ones required unique judgment that a template couldn't handle.

What you're measuring: What % of inbound messages are genuinely repetitive vs. unique? How good are the draft responses?

What this reveals:The ROI of an email automation project. If 60%+ of your inbox is repetitive and the draft responses are 80% usable with light editing, you have a strong case for AI-assisted email handling. If everything requires unique judgment, the opportunity is smaller — but you've still saved time figuring that out before paying someone to build it.

Experiment 4: Write a prompt for your most painful internal process

Pick the internal process your team hates most — the one that's slow, error-prone, or falls through the cracks regularly. Write out how it currently works in plain language.

Then paste that into Claude and ask: "What parts of this process could AI automate or significantly speed up? What parts can't be automated and why?"

Then actually discuss it with the AI. Push back on suggestions that seem off. Ask follow-up questions.

What you're measuring:The quality of the analysis. Can the AI identify genuine opportunities you haven't considered?

What this reveals:Whether your painful process is painful because of information problems (AI can help) or people/culture problems (AI cannot help, and may make worse). This is a crucial distinction. A lot of "operational inefficiency" is actually a management or incentive problem wearing the costume of a process problem.

Experiment 5: Track one week of missed calls or unanswered messages

For one week, manually log every call that went to voicemail, every chat message that took more than an hour to respond to, and every after-hours inquiry that sat until morning. Record the volume and, if possible, whether those contacts followed up or went silent.

Then do the math: multiply the number of dead leads by your average deal value and your close rate.

What you're measuring: The cost of unavailability. Is it big enough to justify a Voice AI or automated follow-up system?

What this reveals: Either a real financial problem (in which case Voice AI or automated follow-up has obvious ROI) or that your lead volume is too low for automation to matter yet (in which case, focus on lead generation before automation).

What to do with the results

After running all five experiments, you should have a clear picture of:

Where AI can genuinely help (and where the hype doesn't apply to your business)
Where your documentation and processes are too thin to automate yet
Which single opportunity has the clearest ROI to tackle first
Whether your team has the appetite and capacity to adopt new tools

If you bring this analysis to a discovery call with any competent AI consultant, you'll get a dramatically better conversation — and a much more accurate quote. If the consultant doesn't ask about any of this, that's useful data too.