Skip to content

For risk and compliance teams

Evidence you can check, for every channel your customers write to

Last updated 17 November 2026

Your developers add Harmwise to your chat, whether an AI, a scripted bot or a person is answering. This page is for the person who signs off: what the record shows, how to check it without trusting us, and what Harmwise does not do. Harmwise certifies nothing. It gives you the evidence to back what you say.

What the record shows

  • For every checked reply: what was detected, each problem found with its rule and how serious it is, the service the person was pointed to, and which version of the safety instructions shaped it.
  • For every piece of guidance: whether it was guidance only, or a draft held because a wrong answer would have been dangerous, and why, with where each reason came from (the reply's own words, something you declared, the person's own words, or a guess).
  • What could not be checked, and why: a language the rules cannot read, instructions that were not used, or our own failure. A check that fails on our side is recorded as an instrument error and never counted as a pass.
  • Who wrote the reply, for a project set as human live chat: a person is checked as a person, and never marked down for not saying they are an AI.
  • No conversation text by default: a fingerprint of the exchange, never the words. If you choose to keep the words from a reply behind a problem, they are encrypted with a key for your organisation for that month, and destroyed with it.

How to check it yourself

  • Each system's records form a hash chain. If a record is changed or removed later, the chain breaks at that point. Your team can verify it from the start in the console, or through the API (/v1/records/verify).
  • Every check returns a receipt signed with a key we publish at /.well-known/harmwise-keys.json.
  • Each day we publish a signed summary of every chain at /heads, and you can get proof that yours is in it. Once a copy exists outside our control, no earlier record can be changed without it showing.
  • Records older than a year move to archive files that check offline, without our services.

The evidence pack

  • A 30 day attestation, signed, generated from the records (from the Pro plan). It is your organisation's own statement, backed by the record.
  • The findings list with each problem's workflow (assigned, resolved, disputed), exportable from the console.
  • The audit trail of every change to keys, the team, settings and findings, and every step our staff take to reach your content, exportable by anyone on the team.
  • A compliance summary for any period through the API, computed from the record.

Mapped to obligations you already have

Harmwise adds no new obligations, and it never says that you comply. It gives you evidence of how each conversation was handled, to sit beside the codes you already follow, which apply whoever is answering:

  • Banking Code of Practice: extra care for customers experiencing vulnerability, including financial abuse and family violence.
  • General Insurance Code of Practice 2020, Part 9: a published family violence policy.
  • National Energy Retail Rules: family violence protections.
  • Telecommunications (Domestic, Family and Sexual Violence Consumer Protections) Industry Standard 2025.
  • UK FCA Consumer Duty and FG21/1: fair treatment of vulnerable customers.

Rules coming for chat services

Chat services, and AI in particular, are getting rules of their own. Each asks for a protocol and proof that it ran, which is what the record gives you. None of this is legal advice; whether a rule covers your service is a question for your own advisers.

  • New York: since 5 November 2025, AI companion operators must detect signs of suicidal ideation or self-harm and direct users to crisis service providers, and tell users they are talking to AI (New York State).
  • California: since 1 January 2026, companion chatbot operators must keep a crisis protocol that refers users to crisis services, publish it, and report annually from 1 July 2027 (Morrison Foerster).
  • European Union: from 2 August 2026, Article 50 of the AI Act requires people to be told they are interacting with an AI system (Jones Walker).
  • Australia: the Online Safety Amendment (Digital Duty of Care) Bill 2026 is before a Senate committee, with a 12-month transition after passage (Pinsent Masons).
  • United Kingdom: in February 2026 the government said it would bring standalone AI chatbots into the Online Safety Act's illegal content duties (Lewis Silkin).

Where Harmwise fits best

Any organisation with a public-facing chat can use Harmwise. It matters most where these are true:

  • A customer chat is live, or launching within months: an AI assistant, a scripted bot or live agents.
  • An industry code, regulator or law already asks for extra care for customers in vulnerable circumstances.
  • Risk or compliance asks for evidence, not assurances.
  • Conversations about money, accounts, addresses, devices or contact details, where financial and technology abuse come up.
  • More than one chat channel, such as a bot that hands over to live agents: one integration covers both, with one standard and one record.
  • Customers in Australia, New Zealand or the UK, where the referral directory and the code mapping are strongest today.

What organisations do instead

  • Build it in house (prompt rules, bot flows, agent scripts): cheap, but hard to get right without casework, rebuilt for every channel, and never checked against what was actually said.
  • Trust the model or platform defaults: free, but no control over the reply and nothing to show a board or regulator.
  • A keyword trigger that ends the chat and shows a number: it cuts the person off at the moment they asked for help, and the number may be closed or wrong for them.
  • A general AI guardrail: it protects the organisation from misuse, for AI only. It does not know what someone who is not safe needs to hear. Harmwise works beside it.
  • Sampling transcripts by hand: slow, partial and after the fact.
  • A one-off assurance review: comfort at one point in time, and nothing about next month's conversations.

Data handling

How we protect the record and your data, who can see what, and where it is held, are set out in the security overview, the privacy notice, the data processing terms and the list of sub-processors. Our staff see metadata only; reaching any of your content needs a stated reason, a second person's approval and a four hour limit, and every view is written to your own audit log.

What it does not do

  • It certifies nothing, and a finding is not legal advice.
  • It is not a counselling service. When someone needs a trained person, its job is to get them to one.
  • Its fixed rules read English. Other languages are read by the AI judge, on Team and Enterprise, where it is turned on.
  • It holds a reply only when your code asks it to check a draft, and only where a wrong answer would be dangerous. It never ends a conversation, and if it cannot be reached your reply goes out as normal.

Start with your own conversations

A Baseline checks a sample of your past conversations and shows how often people raised harm and how they were answered, which turns “we already handle this” from a belief into a number. Request a Baseline, or see how it works in depth.