There’s a large gap between how AI is discussed and how it’s actually being used by businesses that aren’t Google or OpenAI.
The press release version involves transformation, disruption, and revolution. The practical version is more interesting, and more achievable: businesses are using AI to do specific, bounded tasks faster and more accurately than before, and in doing so they’re freeing up time and generating better information to make decisions with.
Here’s what’s working.
Document processing
This is probably the most consistent source of measurable ROI we’ve seen. Invoices, purchase orders, contracts, application forms, compliance documents — any business that receives high volumes of structured or semi-structured documents in various formats has a processing problem that AI handles well.
The pattern is straightforward: instead of a person reading the document and manually keying information into a system, an AI model extracts the relevant fields and flags anything it’s uncertain about for human review. Processing time drops significantly. Error rates drop. Staff spend their time on exceptions, not routine extraction.
The key is that this doesn’t require a large language model doing complex reasoning. It requires a reliable, well-validated extraction pipeline with a sensible review workflow. Getting that right takes some care, but it’s well-understood territory.
Internal knowledge and information retrieval
A recurring frustration in most organisations is that institutional knowledge exists — in documents, emails, past proposals, policies, procedure manuals — but finding the right piece of information at the right moment is slow and unreliable.
Retrieval-augmented generation (RAG) systems let you build a query interface on top of your own documents. You ask a question in plain language and get a referenced answer drawn from your actual content. This is particularly useful in environments where staff regularly need to look up procedures, policies, or historical decisions.
The caveat is that the quality of the output is directly tied to the quality of the underlying documents. If your knowledge base is disorganised, outdated, or inconsistent, the AI will faithfully reflect that. Garbage in, garbage out still applies.
Data analysis and pattern recognition
Many businesses are sitting on data they collected but never properly analysed — transaction histories, operational logs, customer records, supplier data. AI can surface patterns in that data that would take a human analyst weeks to find manually.
This isn’t magic. It works best when there’s a clear business question (“which customers are most likely to churn in the next 90 days?” or “which product lines have margin compression we haven’t noticed?”) and clean enough data to work with. The output is usually a starting point for human decision-making, not a replacement for it.
Routine query handling
Customer-facing AI is more mature than its reputation suggests, but it’s also more limited than its proponents claim. The use cases that work reliably are narrow: answering common questions that have clear answers, routing queries to the right department, handling after-hours enquiries with useful information rather than just a “we’re closed” message.
What doesn’t work is deploying an AI model to handle complex, nuanced, or emotionally sensitive interactions without proper escalation pathways and human oversight. The technology isn’t there yet, and the reputational cost of a poor interaction is real.
What isn’t working yet
A few things worth being honest about:
Fully autonomous agents that independently make and execute decisions across complex systems are real in controlled lab environments and largely unreliable in production at most businesses. The gap between “impressive demo” and “reliably works in our messy real-world environment” is still significant.
AI that compensates for bad underlying data or processes is a fantasy. If your data is a mess or your processes are poorly defined, AI will amplify the problem, not solve it.
Off-the-shelf AI tools as a complete solution for complex, specialised business problems almost always need customisation, validation, and integration work to deliver real value. The tool is the starting point, not the finish line.
What to take from this
The businesses getting the most value from AI right now are the ones treating it like any other technology investment: starting with a specific problem, validating that the solution actually works before scaling it, and measuring the outcome honestly.
If you’re not sure where AI would create real value in your business, that’s a reasonable place to start — and it’s a question we can help answer.