AI solutions for small business come in four kinds, and each one fixes a different problem. A writing assistant fixes the blank page. An automation fixes the handoff that keeps getting dropped. A knowledge base fixes the answer that's buried somewhere in your own files. A voice assistant fixes the work that never gets written down because your hands were full. Name the problem first, then pick the kind.
TL;DR
Look at where the work stalls, not at which tool has the best demo. Ask four questions about the job that's costing you the most time:
Whichever question gets a loud yes points to the kind of solution. The tool lists that rank for this search sort AI by product name. Sorting by problem is faster, and it stops you buying a writing tool for a handoff problem.
The problem: first drafts. Customer replies, proposals, job descriptions, the monthly update you keep putting off.
A plain example: you paste in a customer's long, unhappy email and your notes on what happened, and get back a calm reply you can edit in two minutes.
What it won't do: know your prices, policies or history unless you give them to it, or send anything, update a record, or move data anywhere. A writing assistant drafts. Someone still has to act. To pick one, compare ChatGPT, Claude and Gemini on your own work.
The problem: handoffs. A web form becomes a CRM entry, becomes a quote, becomes a follow-up three days later, and a person copies it at every step.
A plain example: a new inquiry arrives, AI reads it and pulls out the name, job type and address, and the automation creates the record, drafts the quote and schedules the follow-up.
What it won't do: handle a step nobody can describe. Anthropic, which makes Claude, describes workflows as systems where models and tools "are orchestrated through predefined code paths" (Anthropic). If you can't write the path down, you can't automate it yet. Whether to use an off-the-shelf tool or a build is its own decision: weigh workflow tools against a custom build.
The problem: answers that exist but can't be found. The warranty terms in an old proposal, how you handled the last odd request, the steps only one person knows.
A plain example: a new hire asks the AI how you price a rush job, and it answers from your own past quotes and policy documents, with a link to where it found the answer.
The technique these are built on is called retrieval-augmented generation. AWS describes it as a way to extend AI models "to specific domains or an organization's internal knowledge base," with output that "can include citations or references to sources" (AWS). Google Cloud notes that providing facts this way "can mitigate 'gen AI hallucinations'" (Google Cloud). Mitigate, not prevent.
What it won't do: fix documents that are wrong or out of date. It answers from what you give it, so a stale price list produces a confident stale answer.
The problem: work that never gets recorded, because the person doing it has their hands full. Site notes, estimate details, the follow-up you promised in a driveway.
A plain example: you talk through the job in the truck, and the customer, scope and line items land as a structured estimate for you to check.
What it won't do: answer your customers' calls. That's a different product, an AI answering service, with its own tradeoffs. Here's whether an AI answering service is worth it.
Agents are the fifth kind people ask about. Anthropic describes them as systems where models "dynamically direct their own processes and tool usage," and recommends "finding the simplest solution possible, and only increasing complexity when needed" (Anthropic). Our view: the four simpler kinds cover the problems above. Agents fit open-ended work where you can't predict the steps. If you're weighing one, see agent platforms against owning your own build.
The four kinds sit at very different price points, which is another reason to match them to the problem first.
For a fuller picture, here's what AI costs a small business in its first year.
With the problem that costs the most hours each week, and only that one. If you're not sure which that is, see where AI for small business pays off first.
A writing assistant you can set up yourself this week. The other three take more setting up. AI Marketing Labs builds them as systems you own: Custom AI Automations for handoffs, a RAG System that answers from your own files with a link to where each answer came from, and a Voice Assistant that turns what you say on site into finished work.
Sometimes, if the problems sit in the same workflow. An automation can use a knowledge base to answer a question mid-handoff, for example. Build for one problem first and add the second once the first runs on its own.
Which kind of AI solution should a small business try first?A writing assistant is the easiest start, because it costs little and needs no setup beyond good instructions. Move to the other kinds when a specific problem keeps costing you hours every week.
Do you need clean data before using AI?For a writing assistant, no. For an automation or a knowledge base, the information it reads has to be current and in one place, or the output inherits the mess.