When AI Actually Pays Off at Your Size (and When It's Theater)

Most AI advice written for your business was written for a company ten times your size, or for a startup whose product is AI. Neither is you. You run a 20-to-150-person operation, you have no data team, and every tool you buy has to pay for itself in months, not fiscal years. Fair enough. Here is where AI genuinely works at that size, from someone who ships this stuff for a living.

Where it actually pays off

Document processing. This is the workhorse. Invoices, purchase orders, delivery receipts, insurance forms, loan applications: anything that arrives as a PDF or a scan and currently gets typed into a system by a human. Modern document AI reads these and extracts the fields you care about with accuracy good enough to be useful, as long as a human reviews the exceptions. The reason it works: the task is narrow, the output is checkable, and the person who used to key it can spot a wrong number in two seconds. This is the first place I point almost every client.

Quote and proposal generation. You already have the pricing, the templates, and five hundred past quotes. "Draft a quote for this customer and this scope, matching our usual terms" is a task AI does well, because you can review the draft faster than you wrote the original. The output lands in a human's hands before anyone sees it, which is what makes it safe.

Support triage and routing. Sorting inbound email and chat by intent, pulling the customer's history into the thread, drafting the reply for a human to approve. Not a chatbot that talks to customers directly. A system that makes your two support people behave like four.

Internal search over messy files. Every company this size has a shared drive that is a crime scene. "What did we charge this customer last time?" should not require a archaeology expedition. AI search over your actual files, permissions intact, is quietly one of the highest-value uses because it costs almost nothing to set up.

Where it's theater at your size

Custom model training. You do not have enough data, and your problems are not novel enough. Off-the-shelf models through an API will beat anything you could train, at a fraction of the cost. Anyone pitching you a custom model for a 40-person business is selling you their time, not an outcome.

A chatbot because competitors have one. If a customer asks your website bot a real question and it fumbles, you have paid to annoy your best customers. A chatbot only earns its keep when it can actually resolve the top handful of intents end to end. If it can only deflect, put a phone number on the site instead.

Anything that requires a data team you don't have. If the proposal's architecture diagram has a role in it labeled "data engineer," it is a proposal for a bigger company.

The filter

When someone pitches you AI, run it through three questions.

1. Does it touch a workflow you already measure? If you cannot say what the task costs in hours today, you will never prove the AI saved anything. Automate measured work, not vibes.
2. Can you check its output cheaply? The good uses above all share one property: a human can verify the result in seconds. A wrong extracted invoice line is obvious. A subtly wrong strategic recommendation is not. Cheap checking is what makes AI safe at your size.
3. Does it fail safely? The worst AI deployments fail silently: the extraction quietly drifts, nobody notices for a quarter, and the error surfaces in your audit. The good ones fail loudly: low-confidence items get flagged for a human, and the exception queue is the first thing you look at.

If a use case clears all three, do it. If it fails the second or third, it is theater no matter how good the demo looked.

The unglamorous truth

At your size, AI is not a strategy. It is a tool for specific, boring, measurable tasks, and the companies getting value from it are not the ones with AI roadmaps. They are the ones that pointed it at three workflows, checked the outputs weekly, and quietly stopped paying humans to do data entry. The demo-driven ones are still shopping for a strategy.