Where Should a Small Business Start with AI?
Pick a process, not a product. It has to repeat, run on predictable inputs, and produce something you can check. Four weeks is enough to know whether it works.
Most small businesses should start with a single internal process: one that repeats, runs on predictable inputs, and currently occupies someone whose time is worth more than the task. Starting with a specific workflow rather than a tool category cuts the failure rate, and it usually produces something measurable within 30 days.
The wrong question is "which tool should we use?"
Starting with tool selection is the most reliable way to spend four months and produce nothing useful.
The pattern repeats itself. A leadership team reads something about AI, books a demo, buys licences, and runs a kickoff meeting. Six months later the tool is open in two tabs and ignored everywhere else. McKinsey's State of AI report from 2025 put numbers on the distance between the small companies and the large ones [1].
- of companies under $100M in revenue have reached the scaling phase with AI29%
- of companies above $5B in revenue47%
The gap is not about access to tools. Both groups can buy the same software. The difference is whether anyone redesigned the work. Buying software is easy. Changing how the work actually gets done is the project.
Start with a problem. The tool comes last.
Which processes actually qualify?
A process is ready for AI when it meets three conditions: it repeats, it has clear and consistent inputs, and its output can be verified.
Repetitive means it happens more than once a week and follows roughly the same steps each time. Clear inputs means someone could hand it off on day one with a written description. Verifiable output means you can tell, within a reasonable time, whether the result is right or wrong.
The processes that fail this test almost always fail on the third point. If you cannot verify the output, you cannot trust the automation. And if you cannot trust it, someone ends up checking every result by hand, which means the time saved was zero.
The first question in every AI assessment we run is the same one: what takes the most time with the least variation? That intersection is where AI earns its cost.
What the first 30 days look like
A realistic first cycle runs four weeks. Two weeks mapping and testing. Two weeks running in a sandbox before anything touches production.
- Pick the process, not the toolFind the work that takes the most time with the least variation. It has to repeat, run on clear inputs, and produce an output someone can verify.
- Sit with the person who does itOne or two shadowing sessions. Watch the process as it actually runs and ask about every decision point, including the workarounds nobody wrote down.
- Write down what a good result looks likeA tiered framework: what a strong result looks like, what a weak one looks like, and what falls outside scope. That framework is what the AI gets checked against.
- Run it in a sandbox against real inputsTwo weeks in parallel with the real work, reading live data and writing nothing back. The person who owns the process reviews every output and flags the errors.
- Only then let it touch productionWhen the review period closes without surprises. The edge cases the shadowing missed turn up here, and there are always some.
One of our first implementations was for Tuco Marine Group, a workboat builder that needed a better grip on how it found and qualified leads. Better sales calls were never the problem. The work was finding the right companies, checking them against specific criteria, and getting clean records into the CRM. That sounds simple until you sit with the person doing it and watch how many judgment calls happen in a process nobody ever wrote down.
So we ran a shadowing session. One or two working sessions where we watched the process as it actually ran, asked about every decision point, and mapped what made something a qualified lead rather than one to skip. From that we built a tiered framework: what a strong lead looks like, what a weak one looks like, and what falls outside scope entirely. That framework became the verification layer the AI worked against.
The sandbox deserves its own emphasis. The goal is not to ship slowly. The goal is to ship fast and modular, and to run the system in parallel against real inputs without writing to the actual data source. The person who owns the process reviews the outputs, flags the errors, and confirms the tier calls are right. Only when that review period closes does the system touch production. It catches the edge cases the shadowing session missed. There are always some.
The mistake that kills most first attempts
Most first attempts do not fail on the technology. They fail because AI was put on top of a process that was already broken.
If a process has undocumented exceptions, informal workarounds, and rules that live only in one person's head, AI will not fix that. It will fail on the exceptions and produce output that needs constant correction. The process gets abandoned, or worse, it runs in parallel with manual checking, which leaves you with two processes instead of one.
Document the process as it actually runs before you automate any part of it. Map the workarounds and understand why they exist. Some are there for a reason that matters when something goes wrong. Automating over them means the reason disappears, and the failure arrives later without an explanation.
Fix the process first. Then automate.
How much does it cost to start using AI in a small business?
It depends on scope. AI features inside software you already pay for, like Microsoft 365 Copilot, run around $30 per user per month. A structured implementation with outside help typically starts around €1,500 and rises with the complexity of the process being redesigned. Our own prices are on the pricing page. The first project almost always costs less than the time currently spent on the process it replaces.
Do I need a dedicated IT team to implement AI in my company?
No. Most first implementations at small companies involve no custom development. The tools are configured, not coded. What you need is someone with enough time and authority to map the process being changed and to run the transition period, typically four to six weeks of supervised deployment. The bottleneck is almost never technical.
How long before we see results from AI implementation?
A well scoped first implementation usually shows measurable results in four to six weeks. That assumes the process is documented going in. If the documentation work has to happen first, add two to four weeks. Projects with nothing to show after three months almost always started without a clear process definition, not with the wrong tool.
What is the difference between using ChatGPT and implementing AI in my business?
Using ChatGPT is ad hoc. Implementation is systematic. With ChatGPT, a person uses the tool when they remember to. With implementation, the AI sits inside the workflow, so it runs every time the process runs, with consistent inputs, monitored outputs, and a defined path for anything that falls outside normal range. One is a helpful tool. The other changes how the work gets done.
Which department should implement AI first?
Whichever one has the most time consuming process with the most predictable inputs. In practice that is often finance, operations, or customer support, because those functions run on structured data and repetitive tasks. Avoid anything that needs nuanced judgment or where an error is expensive. Legal review and performance evaluation are poor first projects.
- [1]McKinsey: The State of AI (2025)mckinsey.com