The difference between an AI demo and an AI system you can run a business on is what happens when it is wrong. This page sets out the controls: autonomy levels, hard limits, logging, data handling and ownership.
Written into Schedule A and enforced in the platform. You always know which agents can act and which can only propose — and the level is shown on the agent card, not buried in settings.
Used where an error is reversible and cheap — inventory sync, price updates within a floor, return authorisation, forecast refresh, ad budget shifts. Actions execute immediately; a sample is re-checked by QC and every decision stays in the log.
Promotion to supervised happens one agent at a time, after the pilot measures it against its accuracy benchmark. Nothing starts here.
Used where an error reaches the customer or the ledger — customer replies, listing content, refunds, compliance statements, fraud calls. Nothing leaves the building without a person clicking approve.
Some actions never leave this level, regardless of measured accuracy. Those are the hard limits below.
These are platform guards, not prompt instructions. A model cannot talk its way past them, and neither can a cleverly worded customer message.
| Limit | Enforcement |
|---|---|
| Medical or health claims | Blocked at draft, routed to human review |
| Price below the margin floor | Hard refusal plus an alert to the pricing owner |
| Refund above the approval threshold | Requires a named human approver |
| Any reply mentioning legal action | Escalated to the team lead, never auto-sent |
| Unverified customer or account identity | No action taken until identity is verified |
| Money movement, bank details, payout instructions | No agent holds this permission at all |
| Record deletion | Archive and flag only — delete permission is not granted |
| Operating outside the log | Impossible — every decision is written with its inputs |
The short version: it is processed to do the job, logged for traceability, and not contributed to anyone’s model training.
Your data is processed to perform the task. It is not contributed to model training, ours or a provider’s. This is written into the MSA rather than stated on a page.
Every decision is stored with the inputs that produced it, so an error can be traced and guarded against. Retention beyond that follows your own compliance policy.
Shared environment on Pilot and Launch, dedicated VPS on Operate, private cloud or on-premise on Enterprise. Data residency requirements are scoped at contracting.
Each agent holds the narrowest permission scope that lets it do its job, recorded in an access register. Named individual logins, never shared credentials.
You own your data, your process logic and the configured workflows built for you. Akontec retains its own reusable platform components and domain templates.
Thirty days’ notice after the three-month minimum. You keep the data export, the decision log and the configuration documentation. There is no exit fee.
Accuracy is not a launch-day number. It drifts as your data, your products and the underlying models change, which is why sampling is continuous rather than a one-off acceptance test.
Managed operations detail →The discovery call is free and carries no obligation. If AI is the wrong answer for your process, we would rather say so in the first hour than in month three.