AI for SMEs: where to actually start

The honest state of AI for most SMEs

There's a significant gap between what gets discussed at industry conferences and what's practical for a business with 50 employees, a modest IT budget and no in-house data team. Enterprise AI – the kind that makes headlines – typically runs on years of clean, structured data, significant infrastructure investment and teams of specialists to build and maintain it. Most SMEs have none of those things yet, and that's fine.

The tools that deliver the most immediate value for smaller businesses aren't custom-built AI systems – they're the AI features already embedded in software you might already be paying for. Microsoft Copilot, HubSpot's AI tools, Notion AI, and general-purpose assistants like ChatGPT and Claude are all powered by what are called LLMs – Large Language Models. An LLM is a type of AI trained on vast amounts of text to predict and generate language. You give it a prompt, it gives you a response. The technology is genuinely impressive, but it has one critical limitation you need to understand before using it for anything that matters: hallucination.

Hallucination is when an AI confidently states something that isn't true. It doesn't flag uncertainty – it just generates plausible-sounding text, which may or may not be accurate. This isn't a bug that will get patched; it's a fundamental characteristic of how these systems work. It means that any AI output that will be acted upon – sent to a client, used in a contract, published as fact – needs a human to verify it. With that caveat clearly understood, there's still a great deal of useful work these tools can do.

Where AI adds value without a big project

The following five applications are low-risk, low-cost entry points that don't require a development team, a data warehouse or a change management programme. They're places where AI tools are already good enough, and where the downside of a mistake is manageable.

Drafting. AI is genuinely useful for producing first drafts – proposals, emails, job descriptions, tender responses, policy documents. The output won't be publication-ready, and it shouldn't be. The value is in getting from a blank page to a structured draft in minutes rather than hours. Your team still reviews, edits and approves. The AI does the scaffolding; the human does the judgement.

Summarisation. Long documents, meeting transcripts, research reports, supplier contracts – AI handles summarisation well. Tools like Copilot for Microsoft 365 can process a one-hour meeting transcript and produce a list of action points. Claude and ChatGPT can digest a lengthy PDF and answer specific questions about its contents. The time saving here is real and immediate, with relatively low risk because a human is still making decisions based on the summary.

Customer support drafts. Rather than having support staff write responses to common enquiries from scratch, AI can draft a response that a human then reviews and sends. This keeps a person in the loop – which is important, given the hallucination risk – while meaningfully reducing the time each interaction takes. It works particularly well for repetitive enquiry types where tone and accuracy can be checked quickly against a known answer.

Data extraction from documents. Pulling structured data from unstructured sources – reading an invoice and extracting supplier name, date and line items; parsing an email and logging the key details into a spreadsheet – is something AI handles well and that previously required either manual effort or expensive custom software. This is one of the more immediately practical applications for businesses that process a high volume of similar documents.

Internal search across your own documents. This is where a technique called RAG – Retrieval Augmented Generation – becomes relevant. RAG gives an AI access to your specific documents or data before it answers a question, rather than relying solely on what it was trained on. In practice, this means you can ask a question like "what does our supplier agreement with X say about payment terms?" and get an accurate answer drawn from the actual document, rather than a hallucinated guess. It significantly reduces the risk of inaccurate outputs when working with proprietary content. Building a RAG system over your own document library is a meaningful project, but it's one of the more compelling custom AI investments for a business with a large internal knowledge base.

What requires more investment to get right

Some AI applications get a lot of attention and are genuinely worth pursuing – but they're not starting points. They require data quality, system integration and ongoing maintenance that most businesses aren't ready for on day one.

Customer-facing chatbots. The appeal is obvious: a chatbot that handles customer enquiries around the clock, reduces support load and improves response times. The reality is that deploying a chatbot that faces customers without careful guardrails is a reputational risk. Hallucination doesn't just mean an incorrect answer – it can mean an incorrect answer about your prices, your policies or your legal obligations, delivered with full confidence. Getting this right requires a defined scope of what the bot can and can't answer, regular testing and monitoring, and clear escalation paths to a human. It's achievable, but it's a project, not a toggle.

Process automation with AI agents. AI agents – sometimes called agentic AI – are AI systems that can take actions rather than just generate text. They can browse the web, write and run code, call other software via APIs (the connections between applications) and chain together multi-step tasks autonomously. This is powerful and the technology is advancing quickly. But agents operating in production environments need careful design: clear limits on what they're allowed to do, error handling when things go wrong and human oversight at defined checkpoints. Treating agentic AI as a set-and-forget solution is how businesses create expensive errors.

Predictive analytics. Using AI to forecast demand, predict churn or identify at-risk accounts is valuable, but it depends entirely on having clean, complete historical data to train on. If your CRM data is patchy, your sales records are spread across spreadsheets and your data hasn't been collected consistently, the model will reflect those gaps. Predictive analytics is often where the conversation about AI ends up leading – but it's rarely where it should start.

The data readiness question

The most common reason AI projects stall isn't the AI itself – it's the data. AI is only as good as what you feed it. A model trained on incomplete CRM records will produce incomplete predictions. A RAG system built on disorganised, outdated documents will surface outdated information. An extraction tool applied to inconsistently formatted invoices will struggle with the inconsistencies.

Before committing to a custom AI project, it's worth being honest about the state of your data. Are your key systems capturing data consistently? Is it stored in one place, or scattered across spreadsheets, email inboxes and shared drives? Do you have a data warehouse – a central, structured repository of your business data – or is reporting done by pulling exports from individual systems?

For many SMEs, the highest-leverage work isn't the AI model – it's getting the data infrastructure right first. That's less exciting to talk about, but it's what determines whether an AI project delivers value or just adds complexity. Embeddings – the numerical representations of text that capture meaning and power semantic search and document retrieval – only work well when the underlying documents are well-organised and up to date. Garbage in, garbage out applies here as much as it ever has.

Build vs buy: when to use AI platforms vs build custom

The build vs buy decision is where a lot of SME AI projects go wrong. Businesses either buy an off-the-shelf tool when they need something custom, or they spend months building something that already exists as a $50-per-user product.

The default should be to buy. If your problem is generic – drafting content, summarising documents, searching internal files, managing customer communications – there's almost certainly a tool that already does it. Microsoft Copilot for Microsoft 365, Notion AI, HubSpot's AI features, and tools like Otter.ai for meeting transcription are all mature products that work today without any development work on your part. Start there.

Build custom when you have a specific, high-value use case that no off-the-shelf product solves. The clearest cases are: a RAG system over your own proprietary content that gives genuine competitive advantage (your pricing models, your technical specifications, your institutional knowledge); a custom AI agent that automates a specific workflow unique to your business; or fine-tuning – further training a pre-built AI model on your own data so it specialises in your domain – when you need the AI to understand your specific language, products or processes in a way a general-purpose model doesn't.

Fine-tuning is often misunderstood as the first step. In most cases it isn't – a well-prompted general model will outperform a poorly fine-tuned one, and fine-tuning introduces significant ongoing maintenance overhead. Reach for it only when you've exhausted what's achievable with prompt engineering and RAG.

How to start without wasting budget

The most reliable way to waste money on AI is to start with a technology and then look for a problem it solves. The most reliable way to get value is the reverse.

Identify one process that meets three criteria: it's repetitive, it involves text or structured data, and it has a clear quality metric you can measure. Drafting responses to a specific type of customer enquiry. Extracting data from a consistent document type. Summarising a weekly report. These are bounded, testable, low-risk starting points.

Run a pilot. Use existing tools where possible. Measure the outcome – time saved, error rate, user satisfaction – against a baseline. If the result is positive, you have the evidence to invest further. If it isn't, you've learned something useful for a small outlay rather than a large one.

The businesses that get the most from AI at this stage aren't the ones that launch the most ambitious projects – they're the ones that pick a real problem, prove value quickly and build from there. That's as true for a 25-person professional services firm as it is for a 200-person manufacturer.