Why the hallucination guarantee is structural, not a design choice
A rule-based system operates by matching an incoming question against a defined set of patterns or decision-tree branches and returning a pre-authored, human-reviewed response tied to that match — it never synthesizes new text from a language model’s statistical prediction. Because every possible output is a known, reviewed string written in advance, there is no mechanism by which the system can produce a fluent-but-false novel claim; the hallucination failure mode described in the companion piece on AI hallucination simply does not exist in this architecture, by construction rather than by tuning.
What "doesn’t hallucinate" does not mean
This guarantee is narrower than it might sound: a rule-based system can still return an outdated or incorrect pre-authored answer if the underlying content itself was never updated (a stale policy, an old price) — that is a content-maintenance failure, not a hallucination in the generative sense, but it is still a real accuracy risk that requires ongoing human curation. "Doesn’t hallucinate" describes the absence of one specific failure mode, not a general accuracy guarantee independent of how well the underlying content is maintained.
The real cost: coverage and flexibility
A rule-based system can only recognize and respond to questions it has been explicitly built to match. A customer who phrases a common question in an unusual way, combines two questions in one message, or asks something genuinely novel that the decision tree wasn’t designed around will not get a useful automated answer — the system either fails to match anything or falls back to a generic "let me connect you with an agent" response. A generative system, by contrast, can often produce a reasonable response to a novel phrasing it was never explicitly trained on, precisely because it generates rather than matches — which is the same mechanism that creates hallucination risk.
The genuine trade-off, stated plainly
This is a real trade-off, not a case where one approach is strictly better: rule-based systems buy reliability and predictability (every possible answer is known and reviewable in advance) at the cost of coverage (novel or oddly-phrased questions fall through to a human); generative systems buy broader coverage and more natural handling of varied phrasing at the cost of a structural hallucination risk that has to be actively managed. Neither claim — "rule-based is always safer" or "generative is always more capable" — holds up as a blanket statement; the right choice depends on the specific use case, the tolerance for a wrong-but-confident answer in that context, and how much operational effort can go into content maintenance versus AI-output review.
Where this leaves Voz360’s Answer Engine specifically
Voz360’s Answer Engine is rule-based — decision-tree and FAQ-matching bots that resolve repeatable customer questions across chat, voice IVR, and messaging — which means it carries the coverage limitation described above: it handles questions it was built to recognize and defers to a human for genuinely novel ones, rather than attempting to generate a plausible-sounding answer to something outside its defined scope. That is a deliberate trade-off, not a claim that rule-based automation is universally superior; Voz360’s roadmap generative Assist capability is intended to eventually extend coverage into more novel phrasing, with the hallucination-risk management that requires designed in from the start rather than added after the fact.
Can the vendor tell you — in one sentence — which of their AI capabilities are rule-based, which are generative, and which are still roadmap?