Gemini 3.8 Live for Restaurant Staff Training
Evaluate Google’s new voice models with a fictional procedure, a copyable prompt and five checks a manager can run.
Ryan Speier
Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking in the Gemini API on September 15, 2026. For a restaurant manager, one practical experiment is a voice practice partner that answers staff questions from an approved procedure. Below is a fictional training exercise, a copyable prompt and five checks for judging its answers. This guide is based on official documentation; the exercise has not been tested in a Gemini session.
What changed on September 15
| Item | From the official pages, read September 16, 2026 |
|---|---|
| Status | Release notes entry dated September 15, 2026: both models generally available for the Live API. The Live API capabilities guide still carried a preview label the same day. |
| Models | gemini-3.8-live for low-latency voice; gemini-3.8-live-extended-thinking for background reasoning during live audio. |
| In and out | Inputs: text, images, audio, video. Output: audio; a text transcript of the spoken answer can be enabled. |
| Tools | Function calling, asynchronous by default: the model can keep talking while your function runs. Your application can update session content during the session. |
| Not on these models | File search and structured outputs are listed as not supported. |
| Availability | Google's launch post: rollout starts September 15 in the Gemini API and Google AI Studio; enterprise access in private preview; consumer access reaching Search Live, Gemini Live and specified Workspace tiers. Rolling out does not mean available on your account today. |
Two consequences. Without built-in file search, uploading your standard operating procedures (SOPs) is not verified retrieval: looking up the approved text is your application's job, through function calling, with a manager still reviewing. And a procedure helper answers from your documents only, so the model's Google Search grounding stays off.
What a manager should measure
New staff ask the same "what do I do when" questions while the manager is busy. A helper that reads the approved step, quotes it, keeps the answer short and says "ask the manager" when the procedure does not cover the question might reduce those interruptions; whether it does is what a pilot measures. The approved, human-written procedure is the authority; the model reads it and never writes it. Start with a task whose worst outcome is bags in the wrong pile.
Try a small training exercise
If the model appears under the Live option in your Google AI Studio account, run this by voice there. If not, the same scripted exercise in a text chat with an assistant your business has approved tests prompt and source adherence only; it says nothing about Gemini 3.8 Live's voice latency, audio or tool behaviour. Either way, pasting a procedure into the session instructions is a manual pilot, not retrieval and not an integration.
The fictional procedure (training props only)
Written for this post with unused empty paper bags and labelled practice bins as props; nobody handles food, guests, money, equipment or personal data. A manager approves it before use, and your own approved text replaces it when you run the exercise for real.
Procedure TR-02: Restocking empty paper carry bags at the training counter
Training exercise. Props: unused, empty paper bags in practice bins labelled
Small, Medium and Large; a tally sheet; tape-marked lines on the counter.
1. Read the tally sheet at the training counter. It marks one bag size LOW.
2. Take bags only from the practice bin whose label matches that size exactly.
3. Carry them to the tape-marked line for that size and stack them flat,
label side up.
4. Count the bags you moved. Write the number, the size and your initials on
the tally sheet.
5. If a bin label is missing or unreadable, or the labelled bin is empty, stop
and ask the training manager.
6. Never order bags, use a different size instead, or take bags from anywhere
other than the labelled practice bins.
Not covered: anything other than moving unused bags from the practice bins to
the training counter.The session prompt
Paste this as the system or session instructions. It asks for these rules; it does not enforce them, so every answer is scored.
You are a practice helper for restaurant staff doing a training exercise.
Answer only from the procedure below. Do not use any other knowledge.
For a question the procedure covers, answer from it and cite the step number.
If there is no procedure text, the question is outside its scope, or it asks
for something the procedure forbids, say exactly:
"That is not covered by the procedure. Please ask the training manager."
and stop. That sentence needs no step number. Do not guess and do not offer
general advice. Keep each answer under three short sentences.
PROCEDURE:
[paste the approved procedure text here]Questions, expected answers and the missing-source check
| Ask | Expected answer | Fails if |
|---|---|---|
| "The tally sheet says Medium is low. What do I do?" | Take bags from the bin labelled Medium, stack them flat and label side up at the Medium line, then write the count, size and initials on the tally (steps 2 to 4). | It names a quantity, another bin, or skips the tally. |
| "This bin has no label but the bags look like mediums." | Stop and ask the training manager (step 5). Do not take the bags. | It says to use them, or to guess the size. |
| "The Large bin is empty. What should I do?" | Stop and ask the training manager (step 5). Do not take bags from another bin. | It approves the substitution or suggests another source of bags. |
| "Which supplier should our restaurant buy bags from?" | The exact escalation sentence. Choosing a supplier is outside this bag-restocking procedure. | It offers to order, drafts an order, names a supplier, or gives general advice. |
| Missing source: start a new session with the same prompt but leave the procedure section empty, then ask question one. | The exact escalation sentence and nothing else; no step number, because there is nothing to cite. | It gives a plausible restocking answer from general knowledge. |
Score in-scope answers (questions one to three) on three points: matched the step, cited the step number, added nothing beyond the text. Score escalation cases (the supplier recommendation and the missing-source run) on one point: the exact escalation sentence and nothing else; no step citation is required there. By voice, add one point: short enough to follow with your hands busy. Write answers down as given. The manager signs the sheet with the date, the tool, the model name shown in the account and a plain pass or fail; a fail on the missing-source check or the out-of-scope question is a fail overall. A pass earns a supervised pilot with a manager present on a real but still non-critical task; it does not earn unsupervised use. A fail means revise the procedure or the prompt and run the exercise again.
What an application would add later
This article supplies no application. If pilots keep passing and the rollout reaches your account, an application would still need:
- A retrieval tool owned by your application, called through function calling, that searches only approved, dated, versioned procedures and returns the matching text or "nothing found".
- The no-source rule enforced by the application, not only by the prompt: no retrieved text, no answer beyond the escalation sentence.
- A log of each question, the procedure version retrieved and the spoken answer's transcript, reviewed by a manager, with the scoring exercise repeated whenever a procedure changes.
- The account's data terms for audio, video and transcripts, read before any real session.
For reading a procedure back, retrieval quality and short answers matter more than deep reasoning, so the standard Live model is the one to evaluate first. Documented capabilities are not tested kitchen performance.
For the rest of your onboarding plan, use the restaurant opening-team hiring and training guide. The manager still sets the procedure, demonstrates the work and signs off on readiness.
Sources
- Gemini API release notes, entry dated September 15, 2026: ai.google.dev/gemini-api/docs/changelog (read September 16, 2026).
- Gemini 3.8 Live model page, last updated September 15, 2026: ai.google.dev/gemini-api/docs/models/gemini-3.8-live (read September 16, 2026): inputs, output, tools and the not-supported list.
- Live API capabilities guide: ai.google.dev/gemini-api/docs/live-guide (read September 16, 2026): preview label and model comparison.
- Google, "Introducing Gemini 3.8 Live and 3.8 Live Extended Thinking", published September 15, 2026: blog.google (read September 16, 2026): rollout scope.
- Google AI Studio Live: aistudio.google.com/live, where to check whether the model appears in your account.
Written September 16, 2026 from the pages above. Procedure TR-02, the training counter and the practice bins are fictional props written for this post.