Someone just told your AI
she is not safe.
What happens next is on you. When a customer tells your assistant that a partner controls the money, tracks her phone or has just pulled up outside, Harmwise Guardrail makes the reply right, the referral real and the record provable. Two API calls. Nothing sits between you and your model.
Never blocks, edits or delays a reply. No customer content leaves your environment at the first level of adoption.
Customer assistant · 21:14 Sunday
It is already happening in your channel
Disclosure follows the interface. Every public facing AI hears it, whether or not anyone planned for it.
“My husband controls the account and I can't get money out.”
“The plan is in his name and he can see where I am.”
“I can't pay the bill. He takes my wages.”
Left to itself, a general purpose model does one of three things. It replies with something sympathetic and useless. It refuses and leaves her nowhere. Or it names a service that closed at five. A referral that dead ends is worse than none, because she has already spent the courage it took to ask.
Our organisation has delivered more than 1,000 digital safety audits since 2022 and works with 67 domestic and family violence agencies across Australia. We know what these conversations look like, and we built the framework that says what a safe reply is. Harmwise Guardrail puts that framework behind your product.
How it works
Three layers. Each useful alone.
Respond
Governed behaviour composed into your own prompt. The reply recognises the disclosure, believes it, stays inside what your product can do and says so, and never suspends the emergency path.
Refer
The right local service of the right kind, able to serve her in her language and open by its published hours, with what to expect when she calls and a 24/7 alternative in the same breath.
Prove
A sealed, hash chained record of what was detected, applied, said and referred, including the turns where the system could not answer and why.
Your model call stays yours
Harmwise Guardrail composes what goes into your prompt and grades what came out, after the reply has gone. It is a layer beside your model, not a gate in front of it. It is never between you and your model, and never between a person and her reply. If Harmwise Guardrail is down, your product works and the evidence queues.
// 1. Governed behaviour into your prompt
const p = await fetch("/v1/policy/block?jurisdiction=AU", auth)
system = [p.block, ...yourPrompt, p.session_block]
// 2. After the reply is sent, never before
await fetch("/v1/grade", { ...auth, body: { candidate: reply, turns } })Modules
One safety floor. Modules for the harms your channel actually hears.
Each module is a versioned, byte stable block of governed behaviour. You pin it by hash, cache it with your prompt and adopt new versions on your own schedule. Modules are sold separately or together, and the record shows which version applied to every graded reply.
| Module | Covers | Status | What it does |
|---|---|---|---|
| core | Safety floor | Live | The baseline behaviour every system gets: recognise, believe, stay in scope, keep the emergency path open. |
| tia | Trauma informed response | Live | The 18 scored principles of the published framework, extracted byte for byte from the safety module that runs in our own survivor facing products. |
| cald | Culturally and linguistically diverse context | In development | Language, community and settlement context for people who are not well served by an English first reply. |
| crisis | Suicide and self harm recognition | In development | Recognition and handover to a human service. The AI's role stops at recognising the risk and connecting the person; it never attempts the intervention. |
Harmwise Guardrail is a recognition and referral layer, not a counselling service. Where a disclosure needs a trained human, its job is to get the person to one.
Referral
The whole local service system, not one hotline.
A woman whose partner controls the money needs a financial counsellor more than a helpline. One who can't leave because of the dog needs animal welfare placement. Harmwise Guardrail knows which service, which kind, whether it is likely to be open, whether they will take her, and what she can expect when she calls.
- Hours as published, not asserted: every opening time carries the date it was last verified.
- Never a closed door alone: a service that is closed is always offered with a 24/7 alternative in the same reply.
- Ranked on fitness only: services our owners run are flagged, never preferred, and the self preference rate is published and independently audited.
Service kinds carried
Verified local coverage in Australia and New Zealand at launch; national services and emergency numbers elsewhere. Where coverage is missing the product says so in the reply rather than guessing.
The record
What did your system do on that night, and how do you know?
A regulator, an insurer, an enterprise customer and your own board will all ask the same question. Every graded reply writes a record, sealed into a per system hash chain. Alter or remove one after the fact and the chain breaks at a point anyone can check.
From the console you verify the chain from genesis, read every finding, and export a 30 day attestation that states what was applied and what was found. The attestation is yours to make. Harmwise Guardrail certifies nothing; it gives you the evidence to stand behind your own statement.
Zero retention by default
The record holds what was applied and found, and a hash of the exchange, not the words.
Honest about its gaps
Missed turns, degraded periods and our own instrument errors are recorded with cause and blame.
Tamper evident
Per system hash chain; heads published daily and co-anchored with you.
Yours to keep
Export the whole chain with a standalone verifier that works without us.
What you get to work with
A console for your team. An audit trail for everyone else.
Overview
Disclosures by kind, findings by severity and referral outcomes across every system you run, scoped to your organisation only.
Findings
Every graded reply with the principle it met or missed, the module version that applied and the exact span that triggered the finding.
Referrals
Which service kinds we can offer where your customers are, which were offered, and the gaps shown rather than filled with guesses.
Record and keys
Chain verification from genesis, attestation export, live and test keys, and the integration page with your current block hash.
Customers sign in through their own Microsoft tenant. Our staff see tenant performance system wide, run directory verification, gate module releases and keep their own audit log. Neither side can see the other's data, and no customer content is held by either.
The obligation already exists
Your sector code was written for staff. Deploying an AI channel did not suspend it.
Every code below already requires your organisation to recognise and respond to customers experiencing family violence or financial abuse. Harmwise Guardrail is designed to sit inside those obligations, not to add a new one. It gives you a documented, repeatable way to show that the AI channel meets the standard your people are already held to.
Banking
ABA Banking Code: vulnerable customers, financial abuse
Insurance
General Insurance Code: family violence
Energy
Retail Rules: family violence protections
Telecommunications
TCP Code and regulator guidance
Financial services (UK)
FCA Consumer Duty: vulnerable customers
Pricing
Priced on systems and modules, never on how carefully you watch.
Grading every reply costs you nothing extra. The price moves with the number of systems you connect and the modules you switch on, so there is never a reason to grade less. Figures are indicative monthly ranges; a pilot is priced on its scope.
Start today
Developer
Free
- ✓Core safety floor
- ✓Baseline on your transcripts
- ✓REST API, CLI and CI
- ✓Referral directory
Product teams
Growth
A$1,500 to 4,000/month
- ✓Trauma informed or CALD module
- ✓The sealed record and attestations
- ✓Referral fitness grading
- ✓Private regression suite
Regulated organisations
Enterprise
A$10,000 to 30,000/month
- ✓All modules
- ✓Self hosted grader, your own judge
- ✓SSO, residency and retention control
- ✓Platform connectors
1,000+
digital safety audits delivered by our practice since 2022
187
benchmark items across 23 classes in the published framework
18
trauma informed principles, each written to be scored
See what your channel is already hearing.
Baseline runs on your own transcripts, on your own machine. Nothing leaves. You get the count, the kinds and the worst examples in your product's own words. That report is the start of the conversation with us, and it costs you an afternoon.
Run a Baseline