There is a number in the latest US Census Bureau data that should bother anyone running a small company. Among businesses with 250 or more employees, 37% now use AI in some part of their operations. Among businesses with fewer than 20 employees, it is under 20%. The survey ran from December 2025 to May 2026, so this is not a forecast. It is the current state of play.
The gap is not about talent. It is about job descriptions. In a company of 500 people, somebody eventually gets handed AI as a project. In a company of 12, nobody does, because everyone is already busy doing the actual work. So the technology that would help the small company most is the one it never gets around to.
The good news is that the barrier moved. You no longer need an engineer to build your first useful AI integration. You need a chore, an afternoon, and a bit of discipline about what you measure. Here Softeta’s software and machine learning experts disclose, how to do it properly, and where beginners reliably fall over.
Start with a chore, not with a tool
The most common mistake is to open a tool and look for something to do with it. Reverse that and start with the work.
The best filter we have seen came not from a consultancy but from a small business owner writing on Reddit: rank your automation candidates by boredom, not by complexity. The tasks worth handing to a machine are the ones nobody wants, that happen constantly, that follow the same logic every time, and where a mistake is easy to spot. Invoice matching. Ticket routing. Pulling data out of PDFs. Chasing missing documents. Sorting inbound leads. Turning three spreadsheets into one report every Monday morning.
There is a second filter worth applying, from the same conversation: a boring task that still needs a human judgement call halfway through is harder than it looks and often costs more to maintain than it saves. Automate the steps that are dull and predictable. Leave the judgement where it is.
And apply one test before you build anything. Can you already say how often a human gets this wrong, and how long it takes? If you cannot, you will have no way to prove the automation helped. Spend the first week counting. It is the least glamorous step and the one that decides whether the project survives.
Climb the ladder in the right order
Beginners tend to jump straight to the top rung, building an autonomous agent that touches five systems. Then it breaks in a way they cannot diagnose. Go in order instead.
Rung one: do the job by hand, with a chatbot
Before you automate anything, run the task manually inside a normal AI assistant for a week. Paste in ten invoices. Ask for the fields you need. Watch where it gets confused. This is not a waste of time, it is your requirements document, and it costs nothing. Claude’s Projects and Skills are available on the free tier, which makes it the cheapest place to store a repeatable prompt and a reference document without paying anyone.
By the end of the week you will know two things you could not have known in advance: whether the model can actually do the task, and what the edge cases are. Nine times out of ten the edge cases are the whole project.
Rung two: connect two apps
Now make it happen without you. Zapier’s free tier gives you 100 tasks a month and two-step automations, which is enough to prove a point. Make gives you 1,000 credits a month and two active scenarios on its free plan, with a paid Core tier from $12 a month. If you have any technical appetite at all, n8n’s Community Edition is free forever when self-hosted and prices its cloud plans per workflow run rather than per step, from €20 a month, which changes the economics once your automations get long.
A word on the accounting, because it trips up everybody. Zapier bills per task, meaning each successful action, and does not charge for filters or formatting steps. Make bills per credit, and warns that its AI actions can consume more than one. n8n bills per execution, so a fifty step workflow counts once. These are three genuinely different pricing models, and the cheapest headline number is often not the cheapest bill.
Rung three: let it run, but keep a human at the gate
Only after the first two rungs work should the automation write anything into your system of record on its own. Even then, start with a confidence threshold: high-confidence cases go through, anything uncertain lands in a queue for a person. Two practitioners building document processing systems, working independently and in different industries, arrived at exactly the same architecture and described it almost identically. Layered extraction, confidence scores, a fallback, and human review for the edge cases. When two strangers converge on the same design, it is usually the right one.
What this actually costs
Here is the arithmetic nobody does before buying, worked through on the most common first project: pulling structured data out of 10,000 documents a month.
Assume a two-page invoice costs you roughly 1,500 tokens going in and 300 coming out. At published August 2026 API prices, running that through a mid-tier model costs between $12 and $60 a month depending on which one you pick. Google’s Gemini Flash-Lite sits at the bottom of that range, Claude Haiku and OpenAI’s mid-tier models around $30, a larger model closer to $60.
Now compare the purpose-built alternative. Google’s Document AI Invoice Parser and AWS Textract’s expense analysis both charge $10 per 1,000 pages, which on 20,000 pages is $200 a month. Roughly three to sixteen times more than the general-purpose model.
That does not automatically make the parsers a bad buy. What you get for the extra money is determinism: a guaranteed schema, confidence scores, bounding boxes telling you where on the page each value came from. If you are reconciling money, those things matter. But the honest headline is that the model is not the expensive part of your project, and neither is the software. Your automation subscription and the human review time will both dwarf them. Budget accordingly, and stop agonising over which model to pick.
For the assistant layer, current list prices are simple enough to memorise. ChatGPT Business runs $25 per user a month billed monthly, or $20 billed annually, with a two-seat minimum. Microsoft 365 Copilot is $30 per user a month on an annual commitment. Claude Team is $20 a seat a month annually, and Claude’s free tier already includes Projects and Skills. Google bundles Gemini into Workspace Business Standard, which it repriced to $14 per user a month back in 2025. Check the live page before you sign anything, because these move.
Four traps that are live right now
This is the part a generic guide will not tell you, because it goes out of date fast.
Microsoft is retiring AI Builder credits. If you read anywhere that Power Automate Premium comes with 5,000 free AI Builder credits, that is true today and will not be after 1 November 2026, when seeded credits are removed. New customers cannot buy the add-on at all and must use Copilot Credits at $0.01 each. Any Power Platform plan you build on that assumption needs redoing.
OpenAI’s no-code Agent Builder is being deprecated, with a shutdown date of 30 November 2026. It is a lovely drag-and-drop canvas and it is exactly the kind of thing a beginner would build their first workflow on. Do not.
In Copilot Studio, authoring burns credits, not just running the agent. So does previewing and evaluating it. People discover this after a week of enthusiastic tinkering.
Vendor prices are not always what the page says. Microsoft’s own Power Automate pricing page carries a disclaimer that the numbers are “for marketing purposes only and may not be reflective of actual list price.” Azure’s Document Intelligence pricing page currently renders every rate as a placeholder. Get a quote.
Your data, and the law
The four major providers all say the same thing about training on your data, and they all say it with the same escape hatch. OpenAI: “We do not train our models on your data by default.” Anthropic: “By default, we will not use your inputs or outputs from our commercial products to train our models.” Microsoft, on 365 Copilot: prompts, responses and data accessed through Microsoft Graph “aren’t used to train foundation LLMs.” Google, on Workspace: your content “is not human reviewed or otherwise used for Generative AI model training outside your domain without permission.”
Read the qualifiers by default, without permission and outside your domain. On a paid business tier the default is safe. The risk sits in three places: consumer and free tiers, which are governed differently; an admin or an employee switching on an opt-in; and feedback or bug reports, which every one of those policies carves out. So the rule for your team is short. Business account for business data, and think before you click “submit feedback” on a document you would not email to a stranger.
If you operate in the EU, three things are already in force as of August 2026. The prohibited practices and the AI literacy obligation have applied since early 2025, and the literacy one lands on ordinary businesses using AI, not just on vendors. Since 2 August 2026, interactive AI systems must tell people they are talking to a machine, and AI-generated content has labelling requirements. If you put a chatbot on your website, it has to say what it is.
The heavier compliance machinery for high-risk uses was pushed back to 2 December 2027 under the AI Omnibus agreed in June 2026. Worth knowing which uses those are, because they are exactly the ones a small company reaches for first: recruitment, employee management, and credit scoring. You have a runway there, not an exemption.
When to stop and call someone
We at Softeta build ML and AI systems for a living, so treat this with appropriate suspicion. But the honest boundary is not hard to draw.
Stay in no-code territory when the task is a single decision, the volume is modest, the data lives in tools that already have connectors, and a person can review the output. Get help when the automation starts writing to a system where a wrong entry costs real money, when you need an audit trail because of a regulator, when the accuracy plateaus below what the job needs, or when you have four automations quietly depending on each other and nobody can say what breaks if one fails.
That last one is the usual moment. It arrives about six months in, and it is a good problem, because it means the first ones worked.
The first 30 days
If you want a plan you can start on Monday, this is it.
- Week one. Pick one boring, repetitive, high-volume task. Count it. How many per week, how long each, how often does a human get it wrong.
- Week two. Do that task by hand inside an AI assistant, twenty times. Write down every case where it failed. Save the prompt.
- Week three. Wire the two apps together on a free tier so it runs without you, with the output going to a review queue rather than into your real system.
- Week four. Compare against the numbers from week one. If it is better, widen the scope. If it is not, you have spent a month and roughly nothing, and you know something specific about why.
That is the whole method. It is not glamorous and it does not need a data scientist. What it needs is somebody willing to count first and build second.
If you would rather have a second opinion before you start, bring us the process and we (experienced AI and software engineers from the Softeta) will tell you honestly whether AI is the right tool for it. Sometimes the answer is that you need a rules engine, a cleaner spreadsheet, or two systems that finally talk to each other. We will say so.