Every technology wave produces the same pattern: early over-enthusiasm, a trough of disillusionment, and then a plateau of mature, practical adoption. AI is moving through this cycle faster than any previous wave — and we are already in the practical adoption phase for several categories of AI applications.
But that does not mean AI is the right answer to every problem. In our work with clients, the most valuable thing we often do is help them figure out where AI genuinely makes sense — and where it would add complexity without improving the outcome.
The Framework: Three Questions
Before committing to an AI approach, we ask three questions:
- Is the task ambiguous or does it require judgment? AI adds the most value where the rules are too complex or too numerous to enumerate manually. If you can write a deterministic function that handles all cases, do that instead.
- Is there enough data — or can you generate it? AI systems are only as good as the data they are trained or grounded in. If you do not have quality data and no clear path to getting it, AI is not a short-cut.
- What is the cost of being wrong? AI systems make mistakes. If errors are low-stakes, high throughput and imperfect accuracy is a fine trade-off. If errors are high-stakes, you need human oversight in the loop regardless of how confident the model seems.
Where AI Adds Clear, Measurable Value
- Document processing and extraction: Pulling structured data from unstructured text (contracts, invoices, medical notes) is one of the highest-ROI AI applications. The alternative is expensive manual data entry.
- Customer support triage and response drafting: AI can handle the long tail of repetitive, factual support queries and draft responses to complex ones for human review. The key is keeping humans in the loop for anything sensitive.
- Code generation and review: AI coding assistants genuinely accelerate developer productivity, particularly for boilerplate, tests, and documentation. They are most effective as a pair-programming tool, not an autonomous agent.
- Personalisation at scale: Recommendation systems and personalised content ranking are mature AI applications with well-understood evaluation frameworks.
- Anomaly detection: Flagging unusual patterns in logs, transactions, or sensor data is a natural fit for ML models, particularly where the definition of 'unusual' shifts over time.
Where AI Often Disappoints
- Replacing simple rule-based logic: If you can write an if/else, do. AI adds latency, cost, and unpredictability where determinism is possible and preferable.
- Tasks with very small datasets: Foundation model fine-tuning requires at minimum a few hundred high-quality examples. Few-shot prompting can work with much less, but there are limits.
- Decisions that require auditability: Regulators and legal teams often require that decisions can be explained step-by-step. Current LLMs are not reliably auditable in this way.
- Real-time applications with strict latency budgets: LLM inference latency is still measured in hundreds of milliseconds to seconds. If you need sub-10ms responses, a traditional algorithm will beat AI every time.
The Pragmatic Default
When in doubt, start without AI. Build the simplest version of the feature using traditional engineering. This gives you two things: a working baseline to compare against, and a clear specification of what the AI version needs to improve.
The teams that get AI right are not the ones that reach for it first — they are the ones that understand their problem well enough to know when AI is genuinely the best tool for the job.