Reliability
How Aissist keeps its AI agents reliable - predictability and safety, the mechanisms behind them, and how errors are graded and measured.
Last updated: November 1, 2026
Reliable AI is AI a business can depend on in production. This page explains how Aissist approaches it, how errors are measured, and what you can do in your own workspace to keep risk low.
Why it matters
In real operations, one wrong answer can cost more than many right ones are worth. A refund approved against policy, a commitment the business has to honor, or account details given to the wrong person cannot be taken back. Inconsistent answers also erode trust: once your team starts double-checking everything the AI says, the efficiency is gone.
No AI model is perfectly reliable on its own. Aissist treats reliability as a framework every output passes through, not a feature added later.
Two halves: predictability and safety
Predictability - the AI consistently produces the outcome you want: the right answer, the right workflow step, the right tone, the same way on the ten-thousandth conversation as on the first.
Safety - sensitive information and sensitive actions are never wrongly executed: no order details for someone who has not proven who they are, no refund outside policy, no internal instructions handed to someone probing for them.
They fail differently, so they are engineered differently.
The mechanisms
Harness engineering
Shapes the instructions, context, tools and order of steps so every task behaves the same way
Gives the AI only the information a task needs, keeps it inside the provided context, and treats what users paste as data, not commands
Booster
Runs a critical decision several times and goes with the strongest agreement
-
Self-inspect
Checks each reply against the context before it is sent: does it answer the question, is every claim supported, does it follow your workflow
Checks each reply and action for disclosure and authorization before release
Guardrails
A supervising layer catches process drift, such as a skipped step
Enforces your policy limits on actions such as refunds or account changes; anything outside them is stopped or escalated to a person
Workspace design
-
Security checks you define run before a category of information is provided
No single mechanism is enough, so they are stacked: each catches a class of failure the others let through. Read the full explanation on Reliable AI.
How reliability is measured
Reliability is measured as an error rate over AI-handled conversations, graded by business impact. Each tier is reported on its own:
S0
Changes an outcome the business cannot take back
Under 0.01%
S1
Business impact, easily recovered
Under 0.1%
S2
No business impact; internal guidance not followed
S0-S2 combined under 1%
S3
Cosmetic: wording, grammar, tone, formatting
Tracked, outside the SLA
Predictability work drives down S1-S3. Safety work keeps S0 near zero. S2 is the tier to watch: a skipped identity check costs nothing on the day it happens, but it removes the control between a normal conversation and an S0. See Success Metrics for examples and how the metrics are defined.
What you can do
Write security checks into the workspace. State what must be true before a category of information is shared, for example "share order details only when the email on the conversation matches the order". Put it in the workspace instructions or the sub-agent that handles that information. A rule you write holds however the request is phrased, and it has an owner. See Instructions and Information.
Keep sensitive workflows with your team. Let the sub-agent for a sensitive workflow hand the conversation over, and route it with Handover Rules to the team that owns it, so a person handles it from there.
Evaluate before launch and after every change. Run evaluations against realistic conversations, including hard and adversarial customers, before a workflow goes on Auto-Pilot.
Hand over cleanly. Make sure the AI knows when to step back. See Escalation and Handover.
With Ask Engine
Ask Engine can check a conversation you are unsure about:
Did the AI say anything it should not have, or make anything up in this conversation?
Open the conversation in Deploy → Sessions and click Ask about this session. Ask Engine reads what the AI was working from at each turn - the sub-agent, the instruction, what was retrieved and which actions ran - and tells you where an answer came from.
Best practice
Treat every sensitive category of information as needing an explicit check, written down.
Before a sensitive sub-agent goes live, run it through an evaluation set, and let it hand over to your team until a week of its output holds up.
Include an adversarial scenario in your evaluation sets for refunds, account access and anything else you cannot take back.
Review S0 and S2 errors first, every week during rollout.
Related
Last updated

