Why this matters

AI workflows are not deterministic. They interpret language, draft responses, classify intent and sometimes choose the next step. That makes them powerful — and means they need a controlled introduction before they operate at scale.

AiBorz does not ship AI workflows and walk away. Every customer-facing workflow follows a structured launch protocol designed to catch edge cases, build staff confidence and establish the right controls before automation is increased.

The protocol

1. First 100 outputs reviewed manually

During the warm-up phase, every AI-generated output that could affect a customer, supplier, business record or financial outcome is reviewed by a trained staff member before it reaches its destination. This applies to email drafts, CRM updates, document extractions, reporting commentary and any output connected to business systems.

2. No autonomous sending during warm-up

AI may draft, recommend and route. But during the initial period, no customer-facing message is sent, no system record is updated, and no external action is taken without explicit human approval. The AI prepares; a person decides.

3. Weekly review during the first month

The workflow owner and the AiBorz operations team meet weekly to review:

  • Output quality — sampled outputs against business expectations
  • Errors and edge cases — what went wrong and why
  • Staff feedback — adoption, trust, usability issues
  • Approval patterns — what staff are rejecting or editing most
  • Escalation triggers — whether the right things are being escalated

4. Monthly review after stabilisation

After the first month, the review cadence moves to monthly — aligned with the standard managed AI operations cycle. The monthly review covers usage, quality, errors, escalations, costs, staff feedback and planned improvements.

5. Document known failure examples

During the warm-up, the team documents specific failure modes — the kinds of inputs, edge cases or ambiguous situations where the AI is most likely to produce incorrect or unhelpful output. These become part of the workflow's operating documentation and inform future prompt updates and escalation rules.

6. Update escalation rules

Based on real usage data, the escalation rules are refined. Examples include: escalating when confidence is low, when a customer mentions legal language, when a topic is outside approved scope, or when output quality drops below an agreed threshold.

7. Test rollback and pause

Every workflow includes a documented pause and rollback procedure. Staff are trained on how to stop the workflow immediately if an issue is detected, and how to manually handle the affected work while the AI is paused.

When automation can increase

After the warm-up period, the workflow may graduate to a higher level of automation — but only when:

  • Output quality meets the agreed business standard
  • Error rates are within acceptable thresholds
  • Staff are confident using and supervising the system
  • Escalation rules are proven in practice
  • Rollback and pause procedures have been tested
  • The business owner explicitly approves the increase

What this reinforces

This protocol is not a one-off checklist. It reflects AiBorz's operating principles:

  • Human accountability: named people remain responsible for what AI systems do.
  • Testing before trust: outputs are verified before they affect people, money or records.
  • Monitoring after launch: the system is watched, reviewed and improved continuously.
  • Kill switch: every workflow can be paused or rolled back by the people who own it.

For a deeper look at the controls and governance behind this protocol, see the Security & Governance page and the guide to human approval in AI workflows.