AI transparency
What the AI does, and what it never does alone.
Xevrion uses AI where language or judgement is needed and plain code everywhere else. People stay in control of anything consequential.
14 days free No payment details No sales calls
Let it run, or check it first.
Routine work inside your rules goes straight out. Anything else stops at a person. Watch two pieces of work take their routes.
- Lets it run. Filing the proof of delivery for HV-4471: inside your rules, so it is done and logged.
- Checks it first. A reply to Quinn Florists: it waits at Jo Hart, who reads it and approves, then it goes out.
Example company · demo data You choose which work runs by itself in Control. Everything else stops for a person.
The model
Which model, and where it runs.
| Item | Detail |
|---|---|
| Model | Llama 3.1 8B Instruct, an open-weight model from Meta |
| Where it runs | Cloudflare Workers AI, reached through a private bridge that requires Xevrion's own access token |
| Training | None. Cloudflare's Workers AI terms say it does not use customer content to train any AI model, and Xevrion does not train models on your data |
| Settings | Low temperature (0.1) and structured JSON output, for consistent, checkable answers |
| Input limits | At most 20 KB and 6 messages per request |
| Daily limits | 500, 1,500, 5,000 or 10,000 AI calls a day depending on your plan |
| If it is unavailable | The action fails safely and is reported; nothing is guessed |
The path of every task
Nine steps from task to outcome.
No step can be skipped. The model sits in step three, and only when code is not enough.
- 01TaskSomething needs doing: draft a reply, design an automation, summarise a thread.
- 02Risk and complexityHow much could go wrong, and how hard is it? This decides everything that follows.
- 03Rule, code or modelDeterministic code first. The AI model only where judgement or language is genuinely needed.
- 04Recommendation or actionThe output is evidence: a draft, a score, a suggestion, with its confidence.
- 05PolicyYour rules decide whether this kind of action may run on its own.
- 06Approval where requiredAnything consequential waits for a named person to approve it.
- 07ExecutionCarried out once, in your tools, with an idempotency key and your permissions.
- 08AuditWho approved what, what ran, and what the model produced, recorded.
- 09OutcomeChecked against reality, valued in the Value Ledger and used to improve.
Autonomy
You decide how far it may go.
Four levels. Most organisations start at recommend and move individual, low-risk tasks up as trust is earned.
- L0AssistSummarises, drafts and explains. Changes nothing.Always on
- L1RecommendSuggests an action with its expected value and confidence.You decide
- L2Act with approvalCarries out an action only after a named person approves it.Approver required
- L3Act within policyHandles routine, low-risk work inside limits you set, and logs every action.Your limits
Questions
About AI in Xevrion.
Is my data used to train AI models?
No. Xevrion sends a request to the model only to complete a task for your organisation, on Cloudflare Workers AI. Cloudflare's Workers AI data terms say it does not use customer content to train any AI model, and we do not train models on your data either.
Who checks the AI's work?
Your team. Replies, automation steps and posts wait for a person to approve them unless you switch a simple, low-risk kind on to go by itself. Anyone who asks for a person, complains or raises something sensitive is always passed to your team.
Can the AI send emails or change records by itself?
Only if you have set a policy that allows a specific, low-risk action to run on its own. By default, anything consequential is a draft or recommendation that a person approves.
What if the AI is wrong?
Its output is treated as evidence, not authority. Consequential actions need a person's approval, and every outcome is checked afterwards.