How we use AI
AI drafts. A person decides.
Trisova uses AI to save people time: reading documents, drafting plans and answers, and spotting what needs attention. It never acts on its own. Here is every AI feature, what it sees and who stays in control.
The rules every AI feature follows
- AI drafts and suggests; a person decides. No AI output changes a record on its own.
- AI only sees your organisation's data. When a person asks it something, it sees only what that person is allowed to see.
- Suppliers never see AI drafts — they are an internal staff review queue.
- Every AI output is traceable: what it said, which model, and who acted on it.
- Your organisation can switch AI off at any time, or switch off individual features; nothing is sent to the AI provider by a feature that is off.
How you can tell
AI drafts that become part of your records carry a label like this, before a person approves them and after: AI-drafted When you ask the AI a question, the answer says it comes from the AI. Emails and documents the AI drafts are sent only after a person has read, changed if needed and approved them, so they go out under that person's name.
Which AI, and what happens to your data
- The AI is Anthropic (Claude), used through its commercial API. Under those terms, what we send isn't used to train its models.
- Before text is sent, Trisova automatically removes people's names, email addresses and phone numbers it recognises, wherever a feature doesn't need them. This reduces what the AI sees but can't catch every name. Documents and photos you ask the AI to read (a contract, a permit, a photo of a scene) are sent as they are.
- The AI only sees your organisation's data. When a person asks it something, it sees only what that person is allowed to see; the overnight risk suggestions look across your organisation's records.
- For features that create or change a record, what the AI produced, the model, and who approved or changed it are kept with the record and in the audit log.
- An owner or admin can switch AI off for the whole organisation, or feature by feature, under Settings → AI governance. A switched-off feature sends nothing.
What the AI is not used for
- Making decisions about people: hiring, performance, pay or discipline.
- Deciding anything on its own that has a legal or similar effect on a person or a supplier. Awards, approvals, ratings and findings are always a person's decision.
- Emotion recognition, biometric identification or scoring people's behaviour.
The EU AI Act: where Trisova stands
- Trisova provides AI features built on a general-purpose AI model (Claude). Your organisation uses them, which the Act calls a deployer.
- None of Trisova's AI features is prohibited, and none is high-risk under Annex III: they don't decide who is hired, promoted or disciplined, don't monitor or score workers, don't assess people's creditworthiness or access to services, and don't act on their own. They draft text and suggest next steps about suppliers, sites, incidents and documents, and a person decides. Supplier risk ratings are about companies, not people, and a person sets them.
- Transparency (Article 50): the assistant says its answers come from AI, and AI drafts carry the AI label in Trisova before and after a person approves them. Text a person reviews and takes responsibility for before it's sent or published doesn't need a label of its own.
- AI literacy (Article 4) is a duty on every organisation that uses AI. This page, the feature list below and the downloadable register help your staff understand what each feature does and where a person stays in charge.
- We review this whenever we add an AI feature, and when EU guidance or the Act's dates change.
Download the AI register (CSV) for your own AI inventory: every feature, what it sees, who stays in control and what's recorded.
Every AI feature (32)
Daily risk suggestions
What it does: Drafts a plain-English explanation of what it means for the business, what happens if nothing is done, and a recommended action when a new monitoring signal, H&S incident or nearby global event affects one of your suppliers.
What it sees: That supplier's signals, open findings, H&S incidents and active contracts — your organisation's data only.
Who stays in control: Nothing happens until a staff member with edit rights approves it (it becomes a finding) or dismisses it.
What's recorded: Who reviewed it, when, and which AI model drafted it.
Ask the agent
What it does: Answers questions a staff member types about supplier risk.
What it sees: Only what the person asking is already allowed to see — module access rules apply to the AI exactly as they do on screen.
Who stays in control: Read-only. It answers; it never changes data.
What's recorded: Every question and answer, who asked, and the model — in a log that can't be edited or deleted. Each answer lists the suppliers it was based on.
New-supplier risk check
What it does: Gives a short brief on a prospective supplier before onboarding.
What it sees: Public registry checks on the supplier's name plus how similar suppliers in your account have performed.
Who stays in control: Advisory only — it never blocks or approves onboarding.
What's recorded: Logged with the same question/answer audit trail as Ask the agent.
Safety Data Sheet reader
What it does: Reads an uploaded SDS and extracts hazard class, H-statements, exposure limits, PPE and controls for the chemicals register.
What it sees: The SDS PDF only: a supplier's product document, not personal data.
Who stays in control: The reading is shown for review; nothing enters the register until a person applies it, and only into empty fields.
What's recorded: Which model read it and when, stored with the substance.
Incident investigation assistant
What it does: Drafts a 5-whys chain, root cause and corrective actions from what was recorded about an incident.
What it sees: The incident's type, severity, description, location and immediate actions, with names and contact details removed first.
Who stays in control: The investigator applies the draft (only into empty fields) and raises each suggested action themselves.
What's recorded: Which model drafted it and when, stored with the incident; actions it suggested are marked as AI-suggested.
Observation → hazard
What it does: Turns an at-risk safety observation into a hazard statement, a provisional likelihood/consequence rating and a suggested control.
What it sees: The observation's category and note, with names removed first.
Who stays in control: A person creates the hazard from the suggestion; it then goes through normal assessment.
What's recorded: Which model suggested it and when; the hazard links back to the observation.
Checklist drafting
What it does: Drafts an audit or inspection checklist from a short description, for audit templates.
What it sees: Only the description typed in, with names removed first.
Who stays in control: The draft is shown in an editable box; nothing is saved until a person creates the template.
What's recorded: Saved templates are marked as AI-drafted and reviewed by a person.
Environmental permit reader
What it does: Reads an uploaded environmental permit and pulls out its details and each condition, to pre-fill the permit form.
What it sees: The permit document as uploaded (a regulator-issued document).
Who stays in control: Fills a form a person checks and saves; compliance actions are only raised if they tick the box.
What's recorded: Saved permits are marked as AI-read and checked by a person.
Energy & fuel invoice reader
What it does: Reads a fuel, gas or electricity invoice and extracts consumption and billing period for Scope 1 & 2 reporting.
What it sees: The invoice as uploaded (supplier, quantities, dates; may include the site address).
Who stays in control: Pre-fills the emissions entry; a person checks the figures and logs it. Unit mismatches are flagged, not converted.
What's recorded: Logged entries note they were read from an invoice by AI and checked by a person.
Report from a photo (Enterprise)
What it does: Drafts an incident report (description, type, severity, immediate actions) from a photo of the scene.
What it sees: The photo as taken. People may appear in it: the AI is instructed to describe conditions and hazards only, never people, but the image itself is sent to the AI provider.
Who stays in control: Pre-fills the report form; the reporter corrects and submits it.
What's recorded: The report is created by the person who submits it.
Dashboard assistant
What it does: Answers questions about the dashboard a person is looking at (ESG, Health & Safety, Supply Chain, Procurement) and suggests next steps.
What it sees: Only that dashboard's figures, built under the asking person's own access. Personal names are removed from questions first.
Who stays in control: Read-only. It explains and recommends; it never changes, raises or sends anything.
What's recorded: Every question and answer, who asked and the model, in the same log as Ask the agent; counts toward the daily question limit.
Supplier onboarding review
What it does: Reads the documents a new supplier has sent (what each really is, who it's issued to, expiry, any problem), checks them against what that tier and category need, and proposes a risk rating and a request for anything missing.
What it sees: That supplier's uploaded documents and record, open findings, screening alerts, safety incident counts and contracted spend, read under the reviewer's own access.
Who stays in control: Advisory. The rating and the document request go to AI proposals; nothing changes or is sent until someone with edit rights approves it. The document's own type and expiry stay as entered.
What's recorded: What the AI read is stored on each document with the model and time; proposals carry the full audit trail.
AI proposals (next steps)
What it does: Drafts the next step for people to approve: contract renewal, renegotiation or exit notices; supplier chasers for serious findings, missing onboarding documents and weak questionnaire answers; check-ins with suppliers near a live world event; corrective actions for incidents and new regulations; reminders to owners of overdue actions; requests to contractors for their incident investigation; supplier risk ratings and findings; and the monthly board report. Each comes with its business impact and the reason it is recommended.
What it sees: The contract, finding or incident concerned, with the supplier's risk rating and performance, read under the person's own access. Personal names are removed before anything is sent to the AI.
Who stays in control: Nothing is sent or changed until a person with edit rights approves it. They can edit the draft first, or reject it with a reason. Emails go from your own mailbox by default.
What's recorded: An audit trail on every proposal that can't be edited or deleted: what was proposed, every edit (before and after), who approved or rejected it and when, and where it was sent from. Decisions also appear in the admin audit log.
Environmental paperwork reader
What it does: Reads lab certificates (into permit monitoring results), waste transfer and consignment notes (into shipments) and supplier sustainability reports (into Scope 3 data).
What it sees: The uploaded document only. Lab and waste paperwork can carry a signatory's name; it is not stored from the reading.
Who stays in control: Shows the figures for a person to check, correct and save. Results past a limit are handled exactly as if typed in.
What's recorded: Saved rows note they were read by AI and checked by a person.
Action plan drafts
What it does: Drafts the steps, a suggested owner role and a due date for open environment actions (permit breaches, significant impacts, supplier data requests).
What it sees: The action and the record it came from (substance, limit, result, aspect). Personal names are removed before anything is sent.
Who stays in control: Waits in the AI drafts inbox. Approving copies the plan onto the action; dismissing discards it.
What's recorded: Who approved or dismissed it, when, and which model drafted it.
Leadership brief
What it does: Turns the 'what needs attention' list into a short written brief, ranked by business impact.
What it sees: Only the items the person asking can already see; it is generated for that person and shown only to them.
Who stays in control: Read-only. It summarises; it never changes data.
What's recorded: Each brief is kept with the date and model.
Supplier request drafts
What it does: Drafts the email asking a supplier for emissions data or a renewed waste carrier licence.
What it sees: The supplier's name and what is being requested. No personal data.
Who stays in control: A person copies, edits and sends it themselves. Trisova never sends it.
What's recorded: The request itself is tracked as an action.
Double materiality first draft
What it does: Proposes likely impacts, risks and opportunities per ESRS topic from your sector and your own environment, safety and supplier figures.
What it sees: Company sector and aggregate figures (totals and counts), not individual records or personal data.
Who stays in control: Each proposal waits on the materiality page to be accepted or dismissed. Decisions on what is material are never drafted.
What's recorded: Accepted items are marked as AI-proposed, with the reviewer and model.
Ask this report
What it does: Answers questions about the report a person is looking at (a library report or a saved one) and suggests next steps.
What it sees: Only that report, run again under the asking person's own access: its findings, figures and table rows. Personal names are removed from the question, the conversation and the report first.
Who stays in control: Read-only. It explains and recommends; it never changes, raises or sends anything.
What's recorded: Every question and answer, who asked and the model, in the same log as Ask the agent; counts toward the daily question limit.
AI report designer
What it does: Turns a plain-English description of a report into a draft design: which report sections to include, in what order, filtered and sorted.
What it sees: Only the person's description (personal names removed) and the names of the report sections they're allowed to use. The AI never sees any records; the report runs afterwards under the person's own access.
Who stays in control: The person previews the draft with live data, changes it, and decides whether to save or share it. Nothing is sent anywhere.
What's recorded: Saved designs are marked as AI-drafted, with the model that drafted them.
AI data agent (imports)
What it does: Reads the layout of a spreadsheet being imported and proposes what kind of records it holds, which column is which, and how unusual values translate ("Hi" to high, say).
What it sees: Only the column headings and a few sample values from columns that hold dates, numbers or short labels, with personal names, emails and phone numbers removed. Columns holding names, emails or free text are never sent.
Who stays in control: Its proposal is shown for checking and changing; every row is then checked against Trisova's own rules, previewed, and imported only when a person approves. Each import can be undone.
What's recorded: Each proposal is logged with who asked and the model, in the same log as Ask the agent; imports are in Import history.
Exit plan drafting
What it does: Drafts an exit plan for a supplier: strategy, alternative, time to exit, triggers, data to get back, what to do if they fail suddenly, and the steps.
What it sees: Company facts only: the supplier's tier, category and what it handles for you, its contracts' end and notice dates, the critical services relying on it, and other suppliers in the same category. No people's names or contact details.
Who stays in control: The draft is saved as unapproved. It counts only once a person with edit rights checks, changes and approves it, and any later edit sends it back for approval.
What's recorded: The plan shows it was AI-drafted and with which model; drafting and approval are in the audit trail.
Questionnaire drafting
What it does: Drafts a supplier questionnaire (sections, questions, answer types, scored options and must-pass questions) from a description of what the buyer needs.
What it sees: Only the description typed in, the questionnaire's name and its area, with people's names, emails and phone numbers removed. No supplier data.
Who stays in control: The draft is a draft questionnaire that can't be sent to anyone until a person edits it and an owner or admin publishes it.
What's recorded: The questionnaire shows it was AI-drafted and with which model; drafting and publishing are in the audit trail.
Bidder assessment summaries
What it does: Drafts the written summary each bidder gets after a sourcing event: their agreed score and the reasons for each criterion and, for those who didn't win, how they compared with the winning bid (as the UK Procurement Act 2023 requires for public buyers).
What it sees: The event's criteria and weights, the agreed scores and the reasons the panel wrote, and the bidders' company names. No individual evaluators' scores, no people's names or contact details.
Who stays in control: The summary is saved as a draft. It counts only once a person with edit rights checks, changes and approves it; Trisova never sends it to the bidder.
What's recorded: The summary shows it was AI-drafted, with which model, and who approved it and when; drafting is in the audit trail.
Lessons learned
What it does: Turns an incident into a short lesson for other sites (what happened, why, what to check), with a plain version for the contractor involved; and finds look-alike incidents at other sites so the lesson reaches the sites it matters to.
What it sees: The incident's type, severity, description, immediate actions and root cause, and other sites' incident descriptions from the last year, with people's names, emails and phone numbers removed first.
Who stays in control: The lesson is a draft until a person checks it, chooses the sites and shares it; look-alike sites are suggestions a person keeps or removes. Each site answers for itself.
What's recorded: The lesson shows it was AI-drafted and with which model; drafting, look-alike searches and sharing are in the audit trail.
Board risk brief
What it does: Drafts a one-page brief for the board from the main risks register: which risks are outside the board's appetite, which early warnings are red or amber, what decisions are waiting and what is being done.
What it sees: The figures and findings of the Main risks report (risk titles, categories, scores, limits, treatments, early-warning names and levels, overdue action counts), as the risk manager drafting it can see them. Only whether each risk has an owner is sent, never who; and names, emails and phone numbers are removed from the free text first.
Who stays in control: The brief stays a draft until a risk manager, owner or admin edits and approves it; only an approved brief is shown as the board's brief.
What's recorded: The brief shows it was AI-drafted, with which model, and who approved it and when; drafting and approval are in the audit trail.
Bow-tie drafting
What it does: Drafts a bow-tie for one risk: the hazard and the event in the middle, its likely causes with the barriers that stop each, its consequences with the barriers that limit the damage, and what could weaken each barrier, using the organisation's existing controls where they fit.
What it sees: The risk's title, description, causes and consequences, the names of the organisation's controls, and the descriptions, severities and root causes of incidents linked to the risk, with people's names, emails and phone numbers removed first.
Who stays in control: The bow-tie stays a draft until a person checks and approves it; any later change sends it back to draft. The AI never changes a control or a risk score.
What's recorded: The bow-tie shows it was AI-drafted and with which model; drafting and approval are in the audit trail.
Toolbox talk drafting
What it does: Drafts a five-minute safety talk from the pattern in recent incidents, near misses and open hazards, or a topic the supervisor gives.
What it sees: Incident and near-miss types, severities and descriptions from the last 90 days, and open hazards, with people's names, emails and phone numbers removed first.
Who stays in control: The draft can't be given until a person with edit rights reads, changes and approves it.
What's recorded: The talk shows it was AI-drafted, who approved it and when; drafting is in the audit trail.
Contract clause check
What it does: Reads a contract document against the organisation's clause playbook and says, for each clause, whether it's there, weaker than expected or missing, with a quote, and what to ask for. Also reads the end date, notice period and renewal terms.
What it sees: The contract PDF the person chose, the playbook, and two facts from Trisova: what kinds of data the supplier handles for you and whether a critical service relies on them. How serious a gap is, is decided by Trisova from those facts, not by the AI.
Who stays in control: The review changes nothing until a person applies it; they choose which dates to update, and decide whether a gap becomes a finding.
What's recorded: Each review, the model, who applied it and when; running and applying are in the audit trail.
ESG questionnaire answers
What it does: Reads the questions out of a customer's ESG questionnaire (PDF) and drafts an answer to each from the organisation's approved answers and its ESG figures in Trisova, saying what's missing where it can't answer.
What it sees: The questionnaire, the approved answer library and the ESG scorecard figures (emissions, waste, injury rates and similar totals). People's names and contact details are removed from the questions and library answers first; a PDF questionnaire is sent to the AI as uploaded, to read its questions. It is told never to invent a policy, certificate or number.
Who stays in control: Every answer is a draft until a person edits and approves it; nothing is sent from Trisova. The download marks anything not yet approved.
What's recorded: Each answer shows what it was based on and who approved it; drafting is in the audit trail.
Finding the companies in your supply chain
What it does: Searches public web sources for the companies behind a supplier's supply map: producers of a material in a likely source country, and companies publicly named as that supplier's own suppliers. Each comes with the page it was found on.
What it sees: The supplier's name and website, and the materials and countries on its map. Searches run through Anthropic's web search; none of your own records (spend, contracts, incidents) are sent.
Who stays in control: Only when a staff member presses "Find the companies", at most once an hour per supplier. Every company stays a suggestion until a staff member with edit rights confirms it or marks it not right.
What's recorded: Each suggestion keeps its source page, a short quote, when it was found, and who confirmed or rejected it.
Supply map agent
What it does: Works a supplier's sub-tier map like an analyst: checks the named sites below your suppliers against live events (floods, fires, earthquakes, storms) and drafts a check-in email to your supplier when one is near, with how much of the material runs through that site; and finds the plant address of named sites that have none, from public web pages, so they can be placed exactly.
What it sees: The sites on the supply map (company, material, town, share) and the live event feeds already in Trisova. Address searches send only the company name, material and town, through Anthropic's web search. The email draft also uses the AI drafting setting.
Who stays in control: Everything goes to the AI approvals inbox: no email is sent and no address is saved until a person with edit rights approves it. Address searches run only when a staff member presses "Run the map agent" (up to 5 sites a run); the event check runs daily and on request.
What's recorded: Each proposal keeps its source (the event, or the page the address was found on with a quote), the model, and who approved, edited or rejected it.
Questions
For an AI questionnaire, or anything this page doesn't answer, email hello@trisova.com.
Last reviewed 28 October 2026.