AI tools are good at a short list of jobs and mediocre-to-bad at a long list, and the entire trick of getting value from them is knowing which list you're on. Here are the five jobs where AI reliably earns its keep in a small business, and how to pilot one without betting the quarter.
The five jobs
- Drafting. First versions of proposals, job descriptions, customer emails, policy documents. AI produces a solid 70 percent draft in seconds; a human takes it the rest of the way. The economics work because editing is faster than composing, for almost everyone.
- Summarizing. A 40-page contract into one page of terms. A month of support tickets into themes. The meeting recording into decisions and action items. This is arguably the most reliable job on the list.
- Extraction. Pulling structured data out of unstructured documents: invoice numbers and totals from PDFs, key dates from contracts, fields from scanned forms into spreadsheet rows. The full picture is in turning PDFs into data, including where humans stay in the loop.
- Triage. Sorting incoming volume: routing support requests by topic and urgency, flagging which of 200 inbox items need a human today, first-pass categorizing of expenses. AI doesn't make the final call; it orders the queue so humans spend attention where it matters.
- Lookup. Answering questions against your own documents: "what's our refund policy for X" answered from the actual policy, with the source shown. Powerful, with one dependency so common we wrote it up separately: it inherits your document chaos.
What's not on the list: final judgment calls, math you'll act on without checking, anything where a confident wrong answer costs real money unsupervised. AI is a fast junior colleague with unlimited energy and no shame about guessing. Assign work accordingly, and put the guardrails in a one-page policy.
Picking your first project
Score each candidate task on three things: it happens often (weekly-plus), it eats real time, and a wrong output is cheap to catch because a human reviews before anything ships. High on all three wins. For most businesses that's a drafting or summarizing task, which conveniently needs no engineering, just a tool choice.
The two-week pilot
One task, two or three volunteers who actually do that task today, one tool. Before starting, write down the baseline: how long the task takes now, per instance, honestly measured. Then two weeks of real use, tracking time including review time. At the end, the arithmetic in measuring AI ROI tells you whether to expand, adjust, or kill it. Small, measured, and killable beats the company-wide rollout that arrives with a press release and leaves with a shrug.
Want this handled instead of homeworked? That's the job.
Email us →