Seven Questions Every Institution Should Ask an AI Language-Tool Vendor (Including Us)

If you evaluate educational technology for a living, you have sat through the demo where everything works, everything is "fully compliant," and everything your teachers asked about is "on the roadmap — shipping soon."

We are exploring an institutional version of our AI speaking tutor, so we are not a neutral party. But we believe the fastest way to earn trust in this category is to hand evaluators the interrogation script — and commit to answering it ourselves, in writing, including the uncomfortable parts. Here it is.

1. Where is student speech processed, and under whose terms?

Voice is the product in language learning, so this is the first question, not the tenth. Demand three specifics: which AI provider processes the audio, under which of that provider's API surfaces and terms (consumer and enterprise terms differ enormously — including on age restrictions and data use), and in which region the processing happens.

A vendor who answers "it's all GDPR compliant" without naming the provider and the terms hasn't answered.

Our answer today: speech is processed by Google's Gemini service; we name this plainly in our privacy policy. No new institutional cohort is authorized under the current provider path. A future cohort that includes minors requires provider authorization expressly permitting under-18 access, enforceable age and institutional safeguards, and the required residency and school-grade data-processing agreement. EU residency alone is not sufficient.

2. What is stored, for how long, and can we turn it off?

Real-time processing and storage are different things. Ask what persists after the conversation ends: recordings? transcripts? error logs? For how long, deletable by whom, and — the question that separates real data controls from brochure claims — can retention features be disabled for our organisation specifically?

Our answer today: because Gemini Live session resumption is enabled, Google documents that conversation state can be retained for up to 24 hours; provider abuse-monitoring records may be retained for up to 55 days. LinguaLive may separately save short, private correction replays. A daily cleanup targets clips older than 30 days on a best-effort basis, and account deletion immediately attempts an owner-scoped storage purge; backlog or storage failures can delay removal. The replay disable switch is global today; organisation-specific disable is a pre-pilot build-and-verification item. We do not claim a 30-day maximum, guaranteed complete erasure, or that no voice is stored anywhere. Institutional deployment remains blocked until the applicable provider authorization and data terms are confirmed.

3. Which of the features you just showed us exist today?

The single highest-yield question in edtech procurement. Ask the vendor to sort everything discussed into three columns: available today, committed for our deployment with a date, and roadmap. Then ask them to sign the middle column.

A vendor who resists the exercise is telling you which column their demo lives in.

Our answer today: the individual-learner product includes the live voice tutor, a learner-facing mistake ledger, and a typed-response level assessment whose result is saved for the learner. It does not currently include a rubric-graded speaking assessment, class or cohort provisioning, teacher reports, a teacher dashboard, or institutional exports. If a future pilot includes any of those, they belong in the dated, signed pre-pilot column and must be implemented and verified before participants are enrolled. EU-resident processing is also not shipped. We'd rather lose a deal on honesty than win it on futures.

4. What does it cost when students actually use it?

AI inference costs real money per minute, which means uncapped "unlimited" offers hide one of two things: an assumption your students won't use the product, or a future renegotiation. Ask: is there a hard usage cap, who enforces it, and is the contract price fixed against it? Then ask what happens the day a whole class hits the ceiling at once.

Our answer: every learner has a hard daily speaking cap enforced server-side. For any eligible future institutional pilot, our commercial design is a fixed price against agreed caps with no overage billing; the signed pilot terms would need to state what happens when participants hit those caps.

5. What evidence of learning will we receive — and could it survive an audit?

"Engagement" is not evidence. Minutes-used dashboards tell you the tool was open, not that anyone can speak better. Ask for the measurement design: baseline placement, progress measures tied to errors actually corrected, and a pre/post assessment graded against a rubric someone can inspect — ideally CEFR-mapped.

Our answer today: LinguaLive has a typed-response level assessment that saves a result for the learner, plus learner-facing correction and review data. That assessment is not spoken, is not a formal CEFR or certification judgement, and is not graded through an inspectable speaking rubric. We do not currently ship teacher-facing reports, cohort dashboards, term-end reports, or evidence exports. A pilot that depends on learning evidence must define, implement, and verify its baseline and endline rubric, teacher-visible reporting, and institutional export before participants are enrolled.

6. What happens to our teachers?

Ask the vendor to describe the teacher's role in their ideal deployment. Listen for whether teachers are users of the system or an obstacle to it. Tools positioned to replace conversation practice, grading, or the teaching relationship itself tend to fail in institutions for reasons that have nothing to do with technology — and they deserve to.

Our intended institutional role: the practice room between classes, with teachers retaining every curriculum and certification judgement. Teacher assignment workflows and teacher-visible reports are not shipped today, and the current product does not grade a learner for certification. If those workflows are required for a pilot, they are pre-pilot deliverables, not demo promises.

7. If you disappear, what do we keep?

Small vendors build most of the interesting things in this category — we're one of them, so we'll say it plainly: ask about the exit. Data export formats, deletion guarantees on termination, and what happens to the current school year if the company fails or is acquired.

Our answer today: no institutional export-and-termination package is live yet. An authorized pilot contract must define the supported export format, verified deletion workflow and residual legal-retention boundaries before deployment; CSV/PDF export must not be promised until that workflow is built and tested.

Print this, use it on everyone

These seven questions take twenty minutes of a vendor meeting and will tell you more than any demo. Use them on every vendor in your review — including us.

Going deeper: The speaking-hours gap: why university language centres can't staff their way to fluency · LinguaLive for schools, universities & academies