How do you stop an AI agent doing something wrong?
Four hard guardrails, in this order: a confidence floor below which the agent refuses to decide and escalates; per-action cost or impact caps that route to human approval; a hard boundary that keeps safety-critical control in the certified PLC layer, not in the agent; and a documented failure-mode catalogue that surfaces new patterns for review rather than absorbing them silently. All four are policy, not technology.
Talk to an engineer about thisGuardrail 1: confidence floors
Every agent decision comes with a confidence score. Below the floor you set, the agent stops deciding and asks a human. The floor is a policy value, chosen with the domain owner; it can be tightened after go-live as the pilot exposes real failure modes.
Guardrail 2: cost and impact caps
Actions above a cost or downtime threshold route for human approval before execution. This is separate from confidence. A high-confidence decision that costs a lot still gets an approval gate; low-cost and low-impact decisions can flow through.
Guardrail 3: keep safety in the deterministic layer
Safety-critical control loops stay in the PLC and interlock layer. Agents recommend, they do not sit inside the safety-loop authority. This is a design boundary; we hold it even when a customer asks us to relax it, because agents that trip interlocks produce audit trails nobody trusts.
Guardrail 4: a named failure-mode catalogue
A pattern the agent has never seen is a new entry in the catalogue, not a silent correction. New entries go through explicit review by the domain owner before the agent is retrained. Weekly exception review keeps the catalogue honest; monthly reliability-committee review is how new failure modes get added.
Related questions
Ajinkya ships every agent with the four guardrails above, plus an audit log of every decision and its inputs. Concrete escalation-rate benchmark from live pilots: [NEEDS DEPLOYMENT NUMBER].
References and scope
External references explain the underlying technology or regulatory context. Ajinkya pricing, schedules and project results are company provided estimates or examples, not independently audited industry benchmarks. Confirm scope and current requirements before making a purchasing decision.
What would this look like at your plant?
Share your production process, current systems and the problem you want to solve. An engineer, not a salesperson, replies.
Step 1
An engineer reads it
Not a sales bot. Our team replies within 24 hours with questions or times for a call.
Step 2
30 minute call
Your process, machines, systems and the number you want to move. We tell you honestly if we are a fit.
Step 3
A written scope for one line
What we would connect, what you would see, how long it takes. You decide; nothing starts without your approval.
Free scoping call. We start with one line, never change your machine logic, and you own the code and the data.
Reviewed by Amey Kadle, Founder, Ajinkya Technologies. Last reviewed: .