When to use AI
Use AI when the input space is too large or too varied for hand-written rules to keep pace. Examples: parsing free-form documents, classifying support tickets across hundreds of categories, recommending products from millions of items, forecasting demand under shifting conditions, detecting fraud patterns that change weekly, recognising objects in images.
In each of those cases, a deterministic rule engine becomes a maintenance nightmare. Rules need to be rewritten constantly, edge cases multiply, and the team owning the rules grows faster than the value delivered. AI is not a magic answer — but it is the right tool when the problem itself does not have stable rules.
When traditional software wins
Traditional software wins when the logic is fully known and stable. Examples: tax calculations, payment routing, regulatory compliance checks, inventory accounting, scheduling under fixed constraints. In these cases, AI introduces noise where deterministic code introduces certainty. Buyers are sometimes pushed toward AI by hype, but for stable rule-based problems, the right answer is "write the rules".
A useful test: if you can write down the rules in a single afternoon and they will not change for two years, do not use AI.
Cost: where the money actually goes
Buyers consistently underestimate the cost of AI in two places: data and operation. A custom AI system rarely fails because the model is not accurate enough; it fails because the data labelling pipeline was underfunded, or because nobody owns the model in production after launch.
A typical AI project spends roughly 30% of its budget on data, 20% on model and integration, 20% on evaluation and guardrails, and 30% on MLOps and operation. Vendors that quote 80% of the budget on "model development" are mispricing the work — usually because they will not be the ones operating it.
Risk profile
Traditional software has well-known risk profiles: bugs, downtime, security. AI adds new risks: drift (the world changes and the model gets worse), data leakage (training data containing PII shows up in outputs), prompt injection (untrusted input changes model behaviour), and hallucination (confident but wrong outputs). A serious AI development partner builds defences against these into the system from day one.
For buyers, the practical implication is that AI projects need ongoing operation, not just initial build. Treating AI like one-time software delivery is the most common reason AI projects fail in their second year.
Most real systems are hybrids
In practice, the right answer is rarely "all AI" or "all rules". Most production systems combine the two: rules for the deterministic parts (compliance, pricing, scheduling), AI for the open-ended parts (text understanding, classification, recommendation). Zeven's engagements consistently end in hybrid architectures because hybrid is what works.