For a small or medium-sized business, the interesting question isn't "should we use AI?", it's "which specific step in our work is slow or error-prone because it involves language, classification or summarisation?" That's where modern AI models actually help.
Where AI genuinely saves time
- Understanding unstructured text, pulling key fields out of emails, PDFs, scanned documents or customer messages.
- Classification and routing, deciding whether an inbound email is a sales lead, a support ticket, or an invoice question, and sending it to the right place.
- Drafting, not deciding, first drafts of replies, summaries, internal notes, or product descriptions that a human reviews before it goes out.
- Structured extraction from PDFs, turning contracts, orders or delivery notes into rows in a system.
- Search inside your own knowledge, asking questions against internal documents, manuals, or past tickets.
Each of these is a narrow, measurable task where a model can replace minutes of manual reading or typing.
Where AI typically disappoints
- Anything that must be exactly right every time without human review (final invoices, legal statements, payment instructions).
- Tasks where the cost of a confident but wrong answer is high.
- Vague goals like "use AI to improve the business", without a specific workflow attached.
- Replacing entire jobs rather than removing specific steps.
A useful pattern: AI as a small component, not a product
The most reliable AI projects for SMBs don't look like "an AI system". They look like a normal internal tool with one step inside it that calls an AI model and then passes the result to a normal, deterministic process:
- Event happens (new email, new PDF, new form submission).
- Model extracts or classifies something.
- Structured data goes into your existing systems.
- A human reviews exceptions, not the routine cases.
This pattern keeps the AI inside a controlled environment: its inputs are known, its outputs are validated, and its failures land in a predictable place.
Build, buy, or combine?
For standard tasks (grammar help, translation, generic chat), existing SaaS tools are fine. Building custom AI automation becomes worthwhile when:
- the data is private or regulated, and you don't want it inside a generic tool,
- the process is specific enough that no SaaS matches it,
- the AI needs to sit between your existing systems (CRM, ERP, ticketing, email).
In those cases, a small custom integration around an AI model, connected to your data, is usually both cheaper and more useful than bolting together several external tools.
Common mistakes
- Starting with the model. Start with the task. The model is the easiest part of the project.
- Skipping evaluation. "It looks good in the demo" is not enough, you need a small set of real examples with expected outputs.
- Letting the AI make irreversible decisions. Keep the irreversible step on a human, at least until you have measured the error rate.
- Ignoring data privacy. Know exactly which data is sent where, and under which contract.
Getting started without a big budget
- Write down the five most time-consuming text-heavy tasks in the company.
- Pick one where errors are tolerable and the output can be reviewed.
- Prototype an AI-assisted version of that task on a small set of real examples.
- Compare accuracy, time and cost against the manual version.
- If it wins, integrate it into your existing workflow, not next to it.
