Every technology company in the world is now describing itself as an AI company. Every piece of software has an AI feature. Every conference talk leads with AI. In this environment, it’s genuinely difficult to separate the tools that create real business value from the ones that are primarily marketing dressed in impressive language.
This article is an attempt to be honest about both sides: where AI creates genuine, measurable business value today, and where the expectations have run ahead of what the technology can reliably deliver.
Where AI Is Delivering Real Value Right Now
Document processing and information extraction. Extracting structured information from unstructured documents — invoices, contracts, application forms, reports — is genuinely well-suited to current AI capabilities. This has traditionally required either manual data entry or brittle rule-based systems. Modern language models handle varied document formats far more robustly than their predecessors, and the economics of automation at scale are compelling.
First-draft generation. AI assistants have substantially reduced the time required to produce first drafts of written content: job descriptions, internal documentation, proposal templates, FAQs, email responses. The output typically requires human review and editing — it isn’t a replacement for skilled writing — but it significantly reduces the blank-page friction for routine tasks.
Code assistance. For software development teams, AI code assistants have demonstrably improved developer productivity on a range of tasks: code completion, explaining unfamiliar codebases, generating boilerplate, suggesting test cases. The productivity gains are real and measurable in many organisations.
Customer-facing conversational interfaces. For well-defined, high-volume enquiry types — order status, account information, common support questions — AI-powered chat interfaces can handle a significant proportion of queries without human involvement. The key word is “well-defined”: the better the system is constrained to a specific domain with clear data sources, the better the results.
Search and retrieval over internal knowledge. Businesses accumulate large amounts of internal documentation, policies, process guides, and historical information that is genuinely difficult to search effectively. AI-powered retrieval systems (Retrieval-Augmented Generation, or RAG) allow staff to ask questions in natural language and receive answers grounded in the organisation’s own documentation.
Where the Hype Outpaces the Reality
Autonomous agents for complex decision-making. The vision of AI that independently manages complex workflows — making decisions, calling external systems, and completing multi-step tasks without human oversight — is more limited in practice than the demos suggest. Current systems perform well on narrow, well-defined tasks in controlled conditions. They fail unpredictably on tasks that involve ambiguity, novel situations, or the kind of contextual judgement that experienced humans apply almost unconsciously.
Replacing skilled knowledge workers. AI augments skilled workers; in most business contexts today, it does not replace them. A lawyer who uses AI to draft contract clauses is faster and more productive. The AI alone, without the lawyer’s judgement, would produce work that is plausible but unreliable in ways that matter — and in a high-stakes domain, unreliable is unacceptable.
“We just need to add AI to it.” Many projects start from the assumption that adding AI will improve something, then search for a way to do it. The better starting point is: what specific problem is not currently being solved well, and is AI the right solution for it? Sometimes it is. Often the better answer is a simpler integration, a process change, or a better conventional tool.
Accuracy on factual claims. Large language models have a well-documented tendency to produce plausible-sounding but incorrect information — “hallucination” in the technical literature. For tasks where factual accuracy matters, outputs need verification. Deploying AI in contexts where incorrect outputs could cause real harm (medical advice, legal interpretation, financial decisions) requires careful design and robust human oversight.
A Practical Framework
When evaluating an AI opportunity, we suggest asking:
-
What is the specific task? Not “AI for customer service” but “handling X type of enquiry, where the answer is always retrievable from Y data source.”
-
What does failure look like? A wrong answer in an internal productivity tool is a minor annoyance. A wrong answer in a customer-facing system that affects money, health, or legal standing is serious. Your tolerance for failure should shape the design.
-
What is the baseline? If the alternative is a manual process taking ten hours a week, an AI system that gets it right 85% of the time is a significant improvement. If the alternative is an existing automated system with 99% accuracy, an AI system at 95% is a step backwards.
-
Who reviews the output? The most reliable AI deployments today are systems where AI handles the volume and humans handle the edge cases, errors, and exceptions. Designing the human oversight into the workflow from the start produces better outcomes than treating AI as fully autonomous.
Starting Well
The businesses getting genuine value from AI right now are mostly not doing anything dramatically ambitious. They’re automating a specific document-heavy process. They’re giving their developers an AI code assistant. They’re building a knowledge retrieval system for a domain where staff spend significant time looking up information.
Small, specific, measurable. That’s where the real value is.
Thinking about where AI fits in your business? Get in touch — we can help you identify the specific opportunities worth investing in and build them properly.