What people mean by an “AI agent”
A chatbot answers questions. An agent takes steps: it reads something, decides what to do, and then does it in your other tools. “Read this invoice, file it under the right client, and create a task for whoever approves payments” is an agent job. “What does this invoice say?” is a chatbot job.
The difference matters because the value is in the steps. Most small teams don’t lose time reading. They lose it on everything that happens after: filing, forwarding, chasing approvals and remembering deadlines.
Jobs agents handle well today
- Sorting incoming documents. Recognizing an invoice, a contract, a change order or a certificate of insurance, and pulling out the fields that matter: amounts, dates, names, reference numbers.
- Routing and filing. Putting each document in the right folder and sending it to the right person for review.
- Turning text into tasks. Reading a contract or meeting notes and creating follow-ups with due dates, like “send the revised quote by Friday.”
- First drafts. Status updates, client emails, reports and summaries from your own files, ready for a person to check.
- Answering questions about your own documents. “What’s still open on the Level 3 job?” answered from your RFIs and emails, with sources.
Where they still need a person
- Money and commitments. Let an agent prepare a payment or a signature request. Keep a human click on sending it.
- Anything it can’t check. If the agent can’t see the source, it can sound confident and be wrong. Good tools show where each answer came from.
- Messy, one-off judgment calls. A dispute with a long-time client is not an automation problem.
How to start without a big project
- Pick one paper flow that repeats every week. Invoices, change orders, onboarding forms or inbound leads are good candidates.
- Write the rule in a sentence. “Invoices over $5,000 go to Maria for approval; everything else is filed under Accounts Payable.” If you can’t say it in a sentence, it isn’t ready to automate.
- Run it with approval on. For the first two weeks, the agent proposes and a person confirms. You’ll quickly see where it’s reliable.
- Measure time saved, not features. Count how many documents went through untouched.
What to look for in a tool
Look for plain-language rules you can read back, a history of every action the agent took, approval steps for anything that matters, and answers that cite your documents. Avoid tools that need a consultant to set up your first workflow. If the first useful automation takes more than an afternoon, it probably won’t stick in a small team.
ProjectOnUs was built around this idea: your documents arrive, AI reads them, and the workflow you described in one sentence runs, with approvals and a full history.