Your AI agents can act.
Who approves the actions?
A 4-week supervised pilot that installs measurable action control around one of your real AI-agent workflows — evidence on every action, a named human on every consequential one, and a tamper-evident receipt chain proving it after the fact.
The gap this closes
Autonomy without approval is an incident report waiting to be written. Approval without evidence is theatre — a rubber stamp with a audit-trail costume. The middle path is supervised action: the agent executes inside a contained lane, every action ships with its evidence packet, and consequential actions simply stop and wait for a named approver — not a policy document, a running mechanism.
Eight criteria, measured weekly, reported honestly
These are pilot exit criteria — measured, not promised. If a number misses, the weekly report says so. That is the point.
Fit — this pilot is for you if
- You already run AI agents on real work (not demos).
- A risk owner for that workflow is reachable weekly.
- One narrow, measurable action can be scoped for the pilot.
- A sandbox or staging lane exists.
- There is an existing approval line to plug a human into.
If three or more are missing, this is not the right engagement yet — I will say so on the first call.
The four weeks
Week 1
Baseline & instrumentation of your chosen workflow. Success criteria signed by both sides.
Week 2
Supervised execution. Evidence packets per action; consequential actions hold for your named approver; receipt chain live.
Week 3
Load and adversarial week — real volume, deliberate bypass attempts against the containment.
Week 4
Criteria report, handover, and a go / no-go recommendation. You keep the receipts either way.
Commercial
Pilot band RM 20,000–50,000 depending on workflow complexity and integration surface. Post-success pricing is per workflow governed — not per seat, not per token. One design-partner slot at a time.
Who builds it
Arif Fazil — exploration geoscientist, 13+ years deepwater Malaysia; designs and operates governance systems for agentic AI. The codebase is IP-audited clean of employer intellectual property; every public artifact is reproducible from public sources.
Apply
Email [email protected] with subject PILOT, or message t.me/ASI_arifos_bot with the word PILOT. Include three things:
- the workflow your agents actually run,
- who owns its risk,
- what "contained" would mean for you, in one sentence.
No "production-ready". No "zero-risk". Numbers above are criteria — the engagement exists to measure them.
DITEMPA BUKAN DIBERI — forged, not given. · arif-fazil.com