Skip to content
Commerce360

How our AI works

AI you can let take the order

The assistant answers from your own data, works as a small team of agents with a checker on every step, and needs a buyer’s confirmation or a person-issued spending limit before money moves. Here is how, in plain language — open “Technical detail” under any section for the specifics.

  1. 01

    A buyer asks

    In chat, WhatsApp, email — or Teams and phone, on request.

  2. 02

    Finds answers in your data

    Your catalog, prices, and help content — not the open internet.

  3. 03

    Planner decides

    Works out what to do and which specialist should do it.

  4. 04

    Checker reviews

    Looks over the plan and the reply before anything happens.

  5. 05

    Action

    An answer, a quote, a cart, or an order the buyer has confirmed.

Backup model. If the main model fails, a second one takes over automatically.

A buyer asks; the assistant finds answers in your data; a planner decides; a checker reviews; then it acts. A backup model takes over if the main model fails.

Grounded answers

Answers come from your data

The assistant looks things up in your own help content and catalog before it answers, so replies reflect your products, prices, and policies.

  • It searches by exact words and by meaning, so “nitrile gloves, large” finds the right item even when the wording differs.
  • It rewrites a vague question into a better search before looking.
  • Weak matches are dropped instead of being used as an answer.
Technical detail
  • Hybrid retrieval: keyword search combined with vector (semantic) search.
  • Embeddings from Google Vertex AI.
  • Query rewriting before retrieval.
  • A similarity floor: results below a minimum relevance score are discarded.

Who sees what

Staff-only knowledge stays staff-only

Every piece of knowledge has an audience. The assistant only uses what the person asking is allowed to see.

  • Four audiences, from most open to most restricted: anyone, signed-in users, buyers, and your staff.
  • If the assistant can’t confirm who is asking, it treats them as the most restricted case and shows less, not more.
Technical detail
  • Knowledge-base audience tiers: anonymous < signed-in < buyer < staff-only.
  • Fail-closed: an unverified or unknown audience resolves to the lowest tier.

Planner, checker, specialists

A team of agents, each with one job

Instead of one model doing everything, a planner decides what to do, specialist agents do it, and a checker reviews the result.

  • Specialists for shopping, cart and checkout, order follow-up (status, tracking, and cancelling unshipped orders), and handing off to a person in support.
  • A router sends each request to the right specialist, combining fixed rules with the AI model.
Technical detail
  • Planner + critic (checker) loop over specialist agents.
  • Specialist agents: shopping, cart/checkout, order lifecycle, support hand-off.
  • Hybrid router: deterministic rules plus model-based classification.

Guardrails

Safe to let it act

The assistant can place orders and take payments, so it only uses the tools you allow, and no money moves without a buyer’s confirmation or a spending limit a person issues.

  • Instructions hidden in a message or document can’t take over the assistant. In our compatibility suite, 0 of 39 injection attempts led to a write.
  • A checker reviews replies before they go out.
  • Sensitive personal details — such as ID numbers, phone numbers, and card numbers — are masked before the model sees a message, and again on the way out.
  • Each assistant may only use the tools you allow it. Payments need the buyer’s own confirmation, and an outside AI assistant can only spend under a mandate a person issues.
  • Your customer data is only used when a request carries a verified customer identity; requests without one get no customer data. The amount of AI work per conversation and the size of each reply are capped.
Technical detail
  • Prompt-injection guard; output checker.
  • PHI/PII redaction (pattern-based) on the way in and on the way out; the audit log keeps only the redacted text.
  • Token budgets per turn and per session.
  • Per-assistant tool permissions.
  • Tenant identity comes only from a signed credential, never from a client-supplied field; a turn without one is answered without tenant or catalog data.
  • Response size limits on model replies.

Reliability

A backup model takes over if one fails

The assistant runs on a Commerce360-managed model, served by Google Cloud — no additional AI vendor. If the main model has a problem, a backup takes over automatically.

  • Gemini is the main model; Gemma is the backup.
  • If a model call still fails, the assistant ends that turn with a short message instead of an error or a half-finished action.
  • The model is served from Google’s global endpoint, so we do not commit to a processing region for the AI model today.
Technical detail
  • Primary: Gemini on Vertex AI. Fallback: Gemma.
  • Automatic failover with a circuit breaker.
  • Graceful turn termination on model failure.
  • No regional processing commitment for chat.

Tested every release

Checked before it ships

Changes to the assistant are tested automatically before they go out, so behavior that worked yesterday keeps working.

  • Known questions must still reach the right specialist.
  • Sample chats are scored against expected answers.
  • Full conversations are replayed across channels.
Technical detail
  • Router golden tests.
  • Chat evaluation cases.
  • Cross-channel conversation evals, run in CI.

On the roadmap

Planned · not available today

Bring your own model

Run the assistant on a model in your own cloud, or with your own provider key.

Dedicated AI tier

A model tier dedicated to a single customer.

Where data is stored and which providers process it: Security & Trust and Subprocessors.

Watch it take an order

Book a demo and we’ll run the assistant against your own catalog, prices, and policies.

Book a demo