AXIOM FORGE / FIELD NOTES

AI Automation · Illustrated edition

Gemini can route the return. Who approves the refund?

Google’s new Decisions API scores predefined choices. For a returns desk, the useful question is which queue gets the message—not whether to release money.

A cobalt mug with a chip in its rim sits in an opened carton.
Illustrative scenario: a second damaged mug. The buyer wants someone to stop another shipment.

The short answer

Use the API when the answer you need is a label from a list you control. Let it suggest where a message goes, then let your order system establish what actually happened. Before enabling automatic routing, compare its mistakes with those of your current process, including the cases neither should decide.

A message can sound exactly like a refund request and still contain none of the evidence needed to approve one.

In this article

01 / 04

One message, two desks

A customer photographs the chipped rim of a cobalt mug.
One photo can show the damage. The customer’s message adds a second instruction: do not send another replacement.

“The replacement arrived cracked too. Please don’t send a third one.” Imagine this message arriving with a photo. The damage desk can inspect the packaging; the returns desk can stop another shipment. Pick just one label and the first agent may do exactly what the customer asked them not to do.

On November 11, 2026, Google introduced the Gemini Decisions API and its Gemini Decision 1 model. Instead of asking for a written answer, an application can submit defined questions and receive scores. That makes it interesting for the moment before an agent opens the ticket. This is a documentary analysis of Google’s release and documentation, not a hands-on benchmark. The returns desk is illustrative.

02 / 04

What changed is the answer you buy

A conventional classifier can already route tickets. A general language model can already return a label in structured JSON. Neither idea is new. The distinction here is that the service scores fixed choices, yes/no questions or ordered rubric levels directly, without generating a text reply. A team that needs an explanation for the customer still needs another step.

For our broken replacement, ask separately whether the message describes damage and whether the sender asks to stop a shipment. Google documents that questions in a request are independent: the second does not see the first answer. Combine the results in your own application. If both apply, create one case with two tasks and one owner for the customer reply, not two competing conversations.

Google announced access through the Gemini API in Google AI Studio. The documented endpoint is v1beta; the announcement described Google Cloud availability as forthcoming. Verify access for the intended project and region before making it a dependency. No production latency, price advantage or service-level commitment is assumed here.

03 / 04

Give the sorter an ‘other’ tray, not a cash drawer

A fixed-choice question still allocates probability among the options you supply. If every option names a returns queue, an unrelated invoice enquiry can look like a returns case. Add an explicit out-of-scope route. For uncertainty between two valid routes, use a staffed triage queue; that is a different problem from an enquiry that does not belong here at all.

A support specialist compares the mug photograph on screen with the order record.
Routing changes who reads the request. The reviewer still has to check the order before authorising a refund.
QuestionWho should answer?What happens next?
What is the sender asking for?Classifier proposes a route.Agent can move the ticket.
Has the replacement already shipped?Read the fulfillment record.Stop picking only if the workflow still permits it.
Is this order eligible for a refund?Policy applied to the order, with an authorised reviewer for exceptions.Record a separate approval before a payment action.

04 / 04

Buy the comparison before buying the rollout

Ask the supplier to replay a permitted, de-identified set of earlier enquiries without changing a single live ticket. Have support leads mark acceptable destinations before they see model output. Include short messages, translated messages, missing photos and second returns. Where two agents disagree, record both acceptable routes or mark the item for review; don’t turn an ambiguous policy into a model error.

Run the current rules and the proposed classifier over the same cases. For each queue record how many were routed there correctly, which requests were missed and how many a person still had to read. Look at the high-confidence mistakes individually: sending a billing question to dispatch wastes time; missing a request to stop an imminent shipment can be harder to undo. A single accuracy percentage hides the difference.

Release only the routes whose errors the team can tolerate and reverse. If almost everything lands in manual review, compare that staffing cost with simply keeping the old rules. If the API is unavailable, tickets should still enter the ordinary inbox. Our recommendation: make reassignments observable and reversible before granting even that limited authority. Launch documentation is evidence of an interface, not proof that it improves your queue.

Sources and editorial scope

This is Axiom Forge’s procurement and product-delivery guidance, not a guaranteed project plan. The example is illustrative. Sources checked on 12 Nov 2026.

From uncertainty to a clear scope

Know what you are commissioning.

Send the workflow, audience and constraints you already know. We’ll review the context privately and discuss a suitable next step. Please leave out passwords and confidential customer data.

Write to us

Email us — we reply in English, Russian or Armenian.

Our time now:

Analytics preferences

We use Google Analytics and Google Ads to understand which pages help visitors and to measure enquiries. Form contents and personal details are never sent to them.