Every small business owner has now been told they need AI agents. Very few have been told which ones, or why most of the projects quietly die after the demo. We build these for clients, and the pattern separating the ones that stick from the ones that get switched off is consistent enough to write down.
The three ways these projects fail
1. The task was never repetitive enough. An agent earns its keep on volume. If a task happens twice a month, a human doing it in ten minutes will always beat four weeks of engineering. The threshold we look for is roughly twenty repetitions a week — below that, the automation rarely pays back inside a year.
2. The information does not exist in a usable form. "Have the agent answer customer questions from our knowledge base" fails when the knowledge base is a shared drive of 2019 Word documents, three contradictory price lists and the sales manager's memory. The agent is not the bottleneck; the missing source of truth is.
3. Nobody defined what "wrong" looks like. Teams ship an agent with no way to detect a bad output, so the first embarrassing mistake becomes an argument about whether the whole thing was a mistake. You need a review path from day one — even if that path is just "the agent drafts, a human sends."
What reliably works
The pattern behind every automation that survives contact with a real business: a narrow, boring, high-frequency task where a wrong answer is obvious and cheap to catch.
Inbound lead qualification
An enquiry lands — a form, an email, a WhatsApp message. The agent reads it, extracts budget, timeline, location and intent, scores it against your criteria, writes a two-line summary, and either books a call or sends a polite decline. The owner sees a clean queue instead of an inbox.
Works because: high volume, structured output, and a human sees every result anyway. Typical payoff: hours a week, plus faster response times, which is what actually wins the deal.
Support triage and draft replies
Not a bot that talks to customers unsupervised — an agent that reads each incoming ticket, tags it, pulls the relevant policy or order record, and drafts a reply for a human to approve. Your support person goes from writing 40 replies to approving 32 and rewriting 8.
Works because: the human stays in the loop, so an occasional bad draft costs seconds, not a customer.
Document extraction
Invoices, purchase orders, delivery notes, timesheets, ID documents. Anything where a human currently retypes numbers from a PDF into a spreadsheet or accounting system. Modern models handle messy, non-standard layouts that defeated the previous generation of OCR tooling, and you can verify every field against the source in one glance.
Works because: the input is bounded, the output is structured, and correctness is checkable.
The internal "where are we on X?" agent
A model with read access to your project tracker, calendar and shared drive, answering questions like "what did we promise the Kapoor job last week?" and "which invoices are 30 days overdue?". Cheap to build, immediately useful, and low risk — it reads, it does not write.
Content and follow-up drafting
Quotes, follow-up emails, listing descriptions, job posts. Anything you write from a template today, with variables you already have in a system. The agent does not need to be creative; it needs to be fast and consistent.
What usually wastes money
- Fully autonomous customer-facing chat. The failure mode is public and the upside is a deflection rate you could have got from a decent FAQ page. Do it eventually, do not do it first.
- "Give the AI access to everything and see what it does." This is not a project, it is a security incident with a roadmap.
- Agents that replace a decision you have not written down. If you cannot explain your quoting rules to a new hire, you cannot explain them to a model.
- Anything that needs to be right 100% of the time with no human check. Payments, legal commitments, medical anything. The tech is not the problem; your risk tolerance should be.
How to scope one properly
Before writing a line of code, we make the client answer four questions. If any answer is fuzzy, the project is not ready.
- What is the trigger? An email arrives, a form is submitted, it is 9am on Monday. If you cannot name the trigger, there is no workflow.
- What does the agent need to know? List every system it must read from. Every one of those is an integration, and integrations are where the time goes.
- What does "done" look like? A row in a sheet, a draft in an inbox, a message in Slack. One concrete artifact.
- Who checks it, and how fast can they tell it is wrong? If the answer is "nobody" or "we'd find out next quarter", redesign until it isn't.
Build it in the right order
The sequence that keeps these projects alive:
- Do it manually, once, with a checklist. If a human cannot follow the checklist to a correct answer, the model has no chance.
- Automate with a human approving every output. Ship this. It is useful on day one and it generates your evaluation data for free.
- Measure. What fraction of drafts get sent unedited? Where does it fail? Which cases does it get wrong the same way twice?
- Loosen the leash only where the numbers earn it. Auto-send the 80% of cases the agent nails; keep routing the rest to a human.
Notice that this order gives you a working, valuable system at step two — before you have made any bets you cannot walk back.
What it actually costs
The surprise for most owners is that the model is the cheap part. At small-business volumes — a few hundred documents or enquiries a week — inference typically lands in the tens of dollars a month. See our note on retrieval vs fine-tuning for why you rarely need the expensive option.
The real cost is engineering: connecting to your CRM, handling the API that returns malformed data on Tuesdays, deciding what happens when the model is unsure, and building the review interface that makes a human fast rather than annoyed. Budget accordingly, and be suspicious of anyone quoting you for the prompt.
Have a task that eats your week?
We build focused AI agents that plug into the systems you already use — scoped in days, live in weeks, with a human in the loop from the start.
Tell us about it →