The problem: every deck says AI-powered
Open any enterprise alumni platform pitch in 2025 and the AI section reads identically across vendors: intelligent search, personalised careers, smart matching, "AI-native." The claims have converged because the marketing has converged — several vendors have moved quickly to replicate, at least at the announcement level, features first shipped by leading platforms. The products have not converged. The gap between the cheapest claim and the most expensive delivery is large enough that a buyer needs a map, and the map is a ladder: four tiers of "AI-powered," each with a distinct capability, a distinct cost to build, and a distinct test that separates shipping from announcing.
Tier 1 — The keyword box wearing an AI costume
The bottom tier is a faceted search — filter by graduation year, department, location — repositioned with AI language in the release notes. Some of these products predate the current wave entirely and have changed nothing but vocabulary.
The test: type an awkward natural question into the live product during evaluation, not a scripted filter. "Show me partners in Singapore who left in the last five years and now work in fintech." A tier-1 product cannot parse the sentence; it will surface a search box that wants you to pick filters. Vendors route around this test by demoing their own planted queries — which is why the test belongs in a pilot on your data (see below), not in a demo.
Tier 2 — Copilot content tools
The second tier is genuine large-language-model usage aimed at the program team rather than the alumni: newsletter drafting, event descriptions, outreach personalisation, survey summarisation. It is real value, and several honest vendors sit here deliberately and say so. The confusion starts when vendors price tier 2 as if it were tier 3, or when "AI-powered engagement" in a deck turns out to mean an email writer.
The test: ask which persona the feature serves — the program team or the alumni. "AI-powered" that only your communications manager touches is an operations tool, priced like one.
Tier 3 — Natural-language query over a verified graph
Tier three is the current dividing line in the enterprise category: alumni asking the platform questions in plain English and getting answers computed over the employer-verified alumni graph — tenure, roles, career history, current employer, engagement. This is genuinely hard, and the hardness is not the language model; it is the substrate. Query-oriented search over people data requires a deduplicated, effectively-dated, employer-verified record underneath, plus evaluation discipline: a versioned battery of test queries with known-good answers, run on every model and index change, the way a database vendor runs benchmarks.
Market leaders are shipping this; most vendors are demoing it; many are announcing it. The tells are specific. A vendor that ships tier three will happily answer: What does your eval set contain, when did it last block a release, and what is your top-k relevance on awkward queries? A vendor that is demoing it will describe accuracy in adjectives. And a tier-three feature over a poorly linked base degrades to confident nonsense — which is why tier placement and data readiness are the same evaluation.
The test: a bounded pilot on your own data, with your own awkward questions, scored by your own reviewers. A vendor confident in delivery accepts this; a vendor confident only in the demo describes pilots as unnecessary.
Tier 4 — Continuous alumni-to-opportunity matching
The top tier of the current generation: alumni-to-opportunity matching across roles, advisory seats, mentoring, and referral moments, running continuously rather than on request — the system noticing that a specific alum is now the right person for a specific opening, and surfacing it. Tier four consumes everything below it: verified graph, career history depth, current-role freshness, and enough active engagement to generate signal. It is genuinely rare, and its scarcity is architectural: you cannot bolt continuous matching onto a base where two-thirds of profiles have not refreshed in three years, whatever the roadmap slide claims.
The test: ask the vendor to state, unprompted, the data preconditions the matching feature needs and the conditions under which it refuses to match. Credible vendors answer this reflexively — verified history, freshness windows, dedupe tolerance — and will tell you your base is not ready. The vendor who says matching works on any data has sold you a rules engine with a chat interface.
Why announcements outrun delivery — and will keep doing so
Replication economics explain the claim inflation. Announcing an AI feature is cheap: a prototype, a press release, a slide. Shipping it is expensive: data modelling, evaluation infrastructure, and years of unglamorous quality work that customers never see in a launch post. In a consolidating category — and industry research describes this market settling around two to three dominant platform providers — the announcements also serve a survival function: mid-list vendors must appear on the same feature list as the leaders to stay on shortlists. Buyers should expect the announcement-to-delivery gap to widen, not narrow, precisely because the incentive to inflate is strongest in the vendors least able to deliver.
Establishing tier placement: the five-question script
– Which AI features are generally available to all customers today, which are in beta, and which are roadmap? Status phrasing matters; "available" should mean a paying customer uses it in production, and the vendor should name one under NDA.
– Show us the eval. For tier-3 claims: the eval set's shape, its last regression, how quality is measured. No eval discipline means they cannot tell a regression from an improvement — and neither will you after go-live.
– What data preconditions do these features need, and how do you show us ours meets them? The honest answer includes a readiness assessment on your data, not a generic assurance.
– Will you pilot on our data with success criteria we define? The artefact is the pilot itself. Decline is a result, and a cheap one.
– What happens contractually if a tier-3 or tier-4 feature slips? Service credits, price step-downs, or an out clause. Roadmap language survives signature only if someone wrote it down.
The buying position
Score vendors on two axes: highest tier credibly shipped — pilot-verified, not announced — and highest tier your program could operate in the next twenty-four months given your data readiness and ownership. Buy the platform whose shipped tier covers your operable tier plus one. That buys headroom without paying a premium for tiers that will sit idle while your base catches up. And treat every replicated announcement from a vendor that has never shipped a tier above one as what it is: a marketing event with a press release, priced into your licence either way. The tier ladder is not marketing cynicism — it is the honest structure of a market where the difference between tiers is years of engineering, and the difference between a tier and an announcement is nothing at all.