Enterprise face recognition software: a buyer’s guide

There is no single best, and any page that names one without knowing your constraints is selling rather than advising. The choice is decided first by category — cloud API, device terminal, identity-verification service, on-premise platform or edge appliance — and that category is decided by whether biometric data may leave your premises, how many sites you have, and what latency a decision can tolerate. Only after the category is settled does comparing specific products become meaningful.

Start with the requirement rather than the feature list. Establish whether you need 1:1 verification or 1:N identification, the gallery size at three years, whether data may leave the site, the latency budget, and which systems must receive the result. Then ask every vendor the same six evidence questions: which algorithm identifier was independently evaluated and when; at what threshold and on what dataset any quoted figure was measured; the demographic breakdown behind it; which presentation attack instruments liveness was tested against and by whom; whether they will pilot on your cameras and report failures; and where templates live when the network drops.

They are different financial shapes rather than different amounts, and which is cheaper depends entirely on volume and time horizon. Cloud is operational expenditure that scales with transactions and never ends. On-premise is capital expenditure with a known maintenance cost. Model both against your actual transaction forecast over your actual amortisation period; a vendor asserting one is universally cheaper is describing what it happens to sell.

That the vendor submitted an algorithm to independent measurement on data it did not choose, on a specific date. That is meaningful — it is the strongest independent evidence available in this category — but it is evidence about an algorithm rather than a prediction about your deployment. It tells you nothing about camera placement, integration quality, support, or how the system behaves with your population. And there is no such thing as NIST certification: NIST publishes comparative reports and does not certify, approve or endorse vendors.

Identity verification services typically compare a selfie against an identity document during onboarding — a one-off, remote, consumer-facing flow with document authenticity checks and often a compliance workflow attached. Enterprise face recognition is usually continuous and operational: the same people, repeatedly, at doors, gates and cameras. They share underlying technology and almost nothing else, and a product built for one is usually a poor fit for the other.

The honest answer is that the question is incomplete. 1:N error rates rise with gallery size, so the limit is not a storage capacity but the point at which the error rate stops being acceptable for your use case — and that depends on your threshold, your capture quality and what an error costs you. A vendor quoting a maximum gallery size without asking about any of those has answered a different question.

It depends on whether your threat model includes someone presenting a photograph, a screen or a mask to the camera, which in turn depends mostly on whether the capture point is attended. An unattended gate benefits materially; a camera in direct view of a staffed desk benefits less. Where you do need it, specify it against named attack instruments rather than as a feature: ISO/IEC 30107-3 defines how such testing is reported, and a liveness claim that does not name what it was tested against covers nothing specific.

That varies too much to quote responsibly, but the structure should not vary at all: acceptance criteria agreed in writing before it starts, measurement on your own cameras with your own population, both error types recorded, the threshold documented, and the results reported whatever they show. A vendor unwilling to agree failure criteria in advance is proposing a demonstration rather than a pilot.