How we keep AI honest
The safety story of Hi5Agents isn't policy — it's mechanisms. Every claim below points at code shipped in the product. If you find one that isn't true, tell us and we'll fix the product or fix the page.
Grounded output — no invented facts
Where our agents summarize numbers or documents, the output is checked against the source before you see it.
- How it actually worksEvery generated metric summary is re-parsed and diffed against the source query. On mismatch, we retry once, then render the table with a 'summary unavailable' note.
- How it actually worksThe generator runs an invention checker that flags any numeric or safety claim without a page citation, and a warning-preservation check that never softens or drops manufacturer warnings.
- How it actually worksEach candidate score includes an evidence quote extracted verbatim from the resume; if none is found, the criterion is scored as 'insufficient evidence'.
Humans decide on money and reputation
Anything that spends money, changes CRM records, replies to customers, or writes to your store passes through your approval unless you explicitly change the policy.
- How it actually worksEach of these actions writes to the approvals queue with the exact payload; nothing executes until an approver confirms. Policies are per-workspace and per-action.
- How it actually worksReversible write-backs record the pre-image; one click restores it. Non-reversible actions are labeled as such before you approve.
Provenance & labels — AI content is always marked
Anything generated or altered by AI carries a label that follows it out of our product.
- How it actually worksProvenance is stored per file and burned into the render spec as an on-frame corner badge; it can't be toggled off in the export.
- How it actually worksThe generator refuses to emit a claim without a citation index that resolves to a real source page.
Your keys, your data, your accounts
BYOK/BYOA is the default. We don't resell your data or lock you in.
- How it actually worksKeys are stored per-workspace, encrypted at rest, and used directly to call the provider. Your bill from the provider is your bill.
- How it actually worksEvery table scoped to a workspace has an RLS policy that checks membership; a member of workspace A cannot read data from workspace B, ever.
- How it actually worksStandard retention: operational logs 90 days, generated media 12 months unless deleted sooner. Contact support to purge on request.
Published, honest limits
Every agent's public manual has a Limitations section written by a human and reviewed before publish.
- How it actually worksLimitations are a required, non-empty section of every agent doc. Drafts are AI-assisted from a curated facts registry, but a human reviews before publish.
- How it actually worksBoth run a data-sufficiency precheck; below threshold, they emit a 'insufficient data' report naming the missing inputs.
Safety escalation — when we hand off
Some domains require a professional. Our agents recognize that and stop.
- How it actually worksA safety-domain classifier gates guide generation; matched topics produce a professional-required card instead of steps.
- How it actually worksPer-workspace forbidden-topic lists filter outgoing speech; a sampled fraction of live calls are audited post-hoc against the KB.
Email trust@hi5agents.com. We'll fix the product or fix the page. Every claim is backed by shipped code — the audit trail is the git diff.