AWRA OpsHub Search
AI Trust Center

AI you can hand to your auditors.

AWRA OpsHub uses AI to make your operation faster — never to put it at risk. Every AI feature is grounded in facts, reviewed by a human, and private by default. This page tells you exactly which models we run, what data leaves your workspace, what never does, and where you have controls today.

Grounded, not guessing Human-in-the-loop No training on your data Granular controls
4Responsible-AI principles
Retrieval-groundedAnswer method
Per-workspaceIsolation
Per-moduleInsights opt-outs today
Our commitments at a glance
  • We ground answers in your factsThe assistant retrieves approved knowledge first, then writes — so it cites, not invents.
  • A person stays in controlAI drafts and suggests. Approvals, postings and payments always need a human.
  • Your data is not used to train modelsPrompts are used to answer your request and then discarded — not to improve any model.
  • Granular controls for insights modulesAdmins can disable AI insights per module today. A full workspace-wide assistant kill-switch is on our roadmap.
The four principles

Responsible AI, written down

These are not aspirations — they are the rules the product is built to. Everything on this page ladders up to one of them.

01 · GROUNDED

Grounded before generative

Our assistant retrieves approved, curated knowledge and answers from it — low-temperature, short, and traceable to a source. When it doesn't have grounds to answer, it says so and hands off rather than improvising.

02 · ACCOUNTABLE

A human is always accountable

AI accelerates the work — it never has the final word. Suggestions, drafts and summaries are proposed to a person who reviews, edits and approves. Money movement, approvals and compliance filings are never fully automated by a model.

03 · PRIVATE

Private and minimal by default

We send the least data needed to serve a request, isolated per workspace, over encrypted channels. Prompts are not retained to train models, and sensitive fields are kept out of what we send wherever we can.

04 · TRANSPARENT

Transparent and honest about scope

You can see where AI is used, which models we run, and who processes your data — right here, kept current. Where a control is on our roadmap rather than shipped, we say so instead of implying otherwise.

Full disclosure

Everywhere AI shows up in AWRA

No hidden AI. Here is every place a model is involved, what it does, and the data it touches.

Website & Help Assistant

The public assistant answers product and support questions. It searches our curated, approved knowledge base, then a model turns the top matches into a short, grounded answer — and hands off to a human when it isn't confident.

Sends: your question + public help snippets No customer records Grounded & deterministic
Live

Predictive Insights

Forecasts demand, flags slow-moving stock and surfaces trends across your operational data. Outputs are decision-support: they inform a buyer or planner, who decides what to act on.

Runs on: your operational metrics Isolated to your workspace Advisory only
Opt-in

Academy Assistant

Guides users through learning content by answering from the curated Academy knowledge set. Like the website assistant, it stays inside approved material rather than free-forming.

Sends: your question + course knowledge No customer records
Live

Assistive drafting & summaries

Where offered, AI drafts text (notes, descriptions, summaries) to save typing. Every draft lands in an editable field for a person to review and change before anything is saved or sent.

Sends: only the text you're drafting Human edits before save
Opt-in
No black box

The exact models we use

We use several providers for resilience so a single outage does not take the assistant down. All are used the same way: to phrase an answer grounded in retrieved knowledge.

Model Where it runs What we use it for Data handling
Meta Llama 3.3 70B Groq & OpenRouterHosted inference gateways Grounded answering for the website, help and Academy assistants. Prompt sent to phrase one answer; not retained by us to train models.See provider terms below.
Google Gemini 2.0 Flash Google AI Additional grounded answering route used for resilience. Prompt sent to phrase one answer; not retained by us to train models.See provider terms below.

Models change as the field moves. We may add, retire or reorder providers to improve quality, latency or cost. When we do, this table is updated — it is the current source of truth, not a one-time snapshot. Our commitments below hold regardless of which model serves a given request.

On our roadmap. A self-hosted inference option — where AI requests are answered entirely inside our own environment with no third-party gateway — is in active development. We will list it in the table above once it is production-ready.

Data boundaries

What leaves your workspace — and what never does

The single most important thing on this page. We send the minimum needed to answer a request, and draw a hard line around the rest.

What may be sent to a model

Only for the request you triggered, over an encrypted connection:

  • The question or text you typed into an AI feature
  • Relevant snippets from our curated, approved knowledge base
  • For drafting: only the specific field you're editing
  • Minimal instructions the model needs to stay on-topic

What stays inside AWRA

Never sent to a third-party model to serve the assistant:

  • Customer, vendor and contact records in bulk
  • Financial ledgers, invoices and payment credentials
  • Passwords, API keys and connector secrets
  • Another organisation's data — isolation is enforced
Under the hood

How a grounded answer is built

This is why our assistant doesn't hallucinate its way into trouble: it retrieves first, generates second, and stays inside what it found.

1

Retrieve

Your question is matched against our curated knowledge base. Only approved, relevant snippets are pulled.

2

Ground

Those snippets are handed to the model as the only material it may answer from — with instructions to stick to them.

3

Generate

The model writes a short, low-temperature answer. No source, no confident answer — it declines rather than invents.

4

Decline over invent

If the answer would be weak, the assistant declines and hands off to a human — it never fakes certainty.

In plain language

Our data-use commitments

Expand each one for the detail your security and legal teams will want.

We do not use your data to train AI models

Prompts sent to a model exist to answer your request in the moment. We do not collect, retain or use your prompts, documents or records to train, fine-tune or improve any model — ours or a provider's.

Where a request is served by a third-party gateway, we select and configure providers with the intent that inputs are not used for their model training. Provider terms are listed under Sub-processors.

We send the minimum, and keep it short-lived

Each AI request carries only what that request needs — your input plus relevant approved knowledge — not your database. Prompts are transmitted to generate a single answer and are not stored by us as a durable AI dataset.

Everything is isolated per workspace

AI features operate within your workspace boundary. One organisation's data is never blended into another's request, and knowledge retrieval is scoped so answers cannot cross workspaces.

Encrypted in transit, access-controlled at rest

Traffic to model providers travels over encrypted connections. Credentials and connector secrets are encrypted at rest and are never included in prompts.

AI never moves money or files compliance on its own

Models draft, summarise and suggest. Actions with real-world consequences — approvals, payments, tax filings, postings — require an authorised human. There is no path where a model alone commits an irreversible action.

Who processes what

AI sub-processors

The third parties that may process an AI request on our behalf. Review each provider's own data-use terms alongside our commitments.

Sub-processorPurposeData processedRegion
GroqLLM inference gateway Grounded answering (Llama 3.3 70B) Your question + approved knowledge snippets United States
OpenRouterLLM inference gateway Grounded answering (Llama 3.3 70B) Your question + approved knowledge snippets United States
Google Grounded answering (Gemini 2.0 Flash) Your question + approved knowledge snippets Google global infrastructure

Need this list formalised in a contract? Our Data Processing Agreement covers sub-processors, and we notify customers of material changes. A future self-hosted inference option will remove all third-party sub-processors for AI requests once it ships.

You're in control

Your controls & opt-outs

AI is a convenience, not a lock-in. Turn it down, turn it off, or take your data out — on your terms.

Turn AI insights off per module

Admins can disable AI insights per module — inventory, procurement, sales, accounting — from the insights settings. Modules keep working; only their AI insights go quiet.

Manage privacy choices →

Export & delete your data

Request a full data export or account deletion at any time. GDPR-style export is available from account settings; deletion follows our documented erasure flow.

Data & deletion →

Report an AI concern

Saw an answer that looks wrong, biased or unsafe? Tell us and we'll investigate and correct the underlying knowledge.

Responsible disclosure →

Review the legal basis

See how AI processing fits our broader privacy and data-processing commitments.

Privacy policy →

Tenant-wide assistant kill-switch ROADMAP

A single admin toggle to disable the website assistant, academy assistant, and drafting features workspace-wide is in progress. Today, granular controls exist for AI insights modules only.

Talk to us if this is a blocker →

Self-hosted inference ROADMAP

A deployment mode where AI requests are answered entirely inside our own environment — no third-party gateway — is in active development.

Register interest →
Accountability

How we govern AI over time

Trust isn't a launch-day promise. Here's how we keep these commitments true as models and threats evolve.

Human oversight

AI outputs feed decisions made by people. High-impact actions require explicit authorisation — there is no autonomous approval or payment path.

Quality & grounding checks

We tune retrieval and prompts so answers stay accurate and on-source, and we correct the knowledge base when an answer misses.

Change management

Adding or swapping a model is a deliberate change. This page — the model list and sub-processors — is updated to match.

Bias & fairness

Because answers are grounded in curated content rather than open-web free-form, we constrain the surface where bias can creep in.

Incident response

Reported AI issues route into our disclosure and response process. We investigate, remediate the source, and follow up.

Kept current

This is a living document. The version and date below reflect the last time our AI practices or model list changed.

Straight answers

Enterprise AI questions

Do you train models on our data?
No. We do not use your prompts, documents or records to train, fine-tune or improve any model. Prompts serve your immediate request and are not retained by us as a training dataset. See our data-use commitments.
Can we turn AI off entirely?
Today, admins can disable AI insights per module (inventory, procurement, sales, accounting) from the insights settings. A single workspace-wide kill-switch for the website assistant, academy assistant, and drafting features is on our roadmap — not yet shipped. If this is a procurement blocker, let us know.
Which exact models do you run?
Meta Llama 3.3 70B (via Groq and OpenRouter) and Google Gemini 2.0 Flash. The full, current list is in the models table. A self-hosted option is on our roadmap.
Does our financial or customer data get sent to a model?
Not in bulk. The assistant sends your question plus approved knowledge snippets — not your ledgers, records or secrets. See what we send for the hard boundary.
Can inference stay inside our own environment?
This is on our roadmap, not shipped. Today AI requests are served by one of the third-party providers listed in the models table above. If your procurement or regulator requires in-house inference, talk to us — we track this demand actively.
What happens if the AI gets something wrong?
Because answers are grounded and a human reviews high-impact outputs, a wrong answer is caught before it becomes a wrong action. You can also report it and we'll fix the underlying knowledge.
Last updated  10 July 2026  ·  Version 1.0

Bring AI to your operation — without the risk.

Grounded answers, human oversight, and controls you actually hold. That's the deal.

Help Center

Need a quick answer while you read?

Run inventory, procurement, assets, sales, and field work with approved AWRA guidance for setup, migration, integrations, security, pricing, and support.

Search all approved AWRA public help articles.

Open Help Center