The AI Software Buying Guide: Telling Real Capability From a Marketing Label

'AI-powered' is a label, not a spec

Every SaaS pricing page now has an AI section, which means the word itself has stopped telling you anything useful. Some of those features are a genuine model doing real work on your specific data. Some are a rules engine that existed before the word 'AI' got added to the marketing copy, with a chatbot skin layered on top. The label doesn't distinguish between them — three concrete questions during a sales conversation usually do.

Question 1: what specific task does it do, and on whose model?

A vendor that can say precisely what the feature does — 'it drafts a first-pass reply to support tickets using GPT-4' — and which underlying model or approach it's built on is telling you something falsifiable and checkable. A vendor that answers with marketing language ('it learns your business and gets smarter over time') without naming a specific task or model is often describing a feature that doesn't do much more than basic automation with a new label.

Question 2: where does your data go, and can you see it work on your own data?

A vendor confident in a real capability will let you test it live against your own support tickets, your own documents, your own records — because that's the best way to prove it works. A vendor that resists this, or only offers pre-canned demo examples, is often protecting a feature that performs worse than advertised outside a curated demo script.

Question 3: is the price premium proportionate to the capability?

A genuine AI feature solving a real, narrow problem — auto-categorizing inbound tickets, drafting a first pass at a routine document, flagging anomalies in a dataset a human would otherwise scan manually — is reasonably priced at a 20-50% premium over the non-AI tier. A 3x price jump attached to a vague 'AI-powered' banner, with no specific new capability to point to, is a sign the premium is priced against the hype rather than the feature.

Features where the premium is usually worth paying — and where it usually isn't

Narrow, well-scoped tasks — auto-categorization, semantic search across a large document set, a reasonable first-draft generator you're expected to edit — tend to deliver real, measurable time savings because the task itself is narrow enough for a model to do reliably. Broader claims — full meeting-notes automation with no editing needed, autonomous task prioritization, revenue forecasting from thin historical data — tend to underdeliver, not because AI can't help with any of them eventually, but because the current version of the claim is usually broader than what the underlying model can reliably do today.

The bottom line

The label doesn't tell you anything a vendor's answers to three direct questions won't tell you better: what specific task, on what model, verified on your own data. Ask for a live demo, expect a proportionate premium, and treat vague marketing language as the disqualifying signal it usually is.