Quiet failure is the default
Most AI products do not crash. They slowly leak trust, safety and budget. That is the central point of a guide by Arsh Shah Dilbagi on Adaline Labs, written for engineers and product managers shipping AI features.
The output looks coherent, the user seems satisfied and the metrics stay flat. Underneath, the system may be wrong or breaking a rule. The author's example is a support bot that approves a refund it should not. The customer is happy, and the mistake only appears later as lost margin.
His answer is to treat observability and evaluation as the reliability layer. One makes behavior visible. The other makes it measurable.
Prompts are business logic
The guide argues that a prompt in production is not just text. Once a model can read data and take action, the prompt becomes logic that drives what the system does. And the real program is not one string. It is the whole bundle the model receives: instructions, variables, retrieved context, user input, tool permissions and settings such as model version.
That bundle can contain conflicts. A user message might try to override instructions. A retrieved document might contain hidden directions. The model can still produce fluent output while following the wrong instruction, which is why these problems are hard to spot without measurement.
The practical advice is to version prompts like code, review changes, keep rollback points and assign an owner to each prompt.
What to record for every run
The author defines observability narrowly: you should be able to reconstruct any run the way you would investigate an incident. His checklist includes the prompt version and the assembled prompt, the user input, retrieved context, tool calls and their effects, model and settings, token use and cost, timing and the final output.
With that record, a bad answer points to its cause. If retrieval returned irrelevant context, fix retrieval. If a tool call was wrong, fix tool selection. If behavior changed after a prompt edit, fix the prompt.
Name the failures you care about
The guide groups quiet failures into four categories so teams can monitor them.
- Policy failures that look like success, such as missing approvals.
- Security failures, such as instructions that override the rules.
- Cost and speed failures, such as loops and token spikes.
- Correctness failures hidden by fluency, such as missing citations or answers that do not match the sources.
He shares an example from his own work. After a switch to a document scanning workflow, costs jumped because the model padded its output with long runs of spaces. Customers did not notice at first, but the trace made it obvious and a cost guardrail followed.
Start with real examples
Demos hide variety. Production is full of messy inputs and unexpected requests. The author suggests starting with about twenty representative cases, including messy and adversarial ones, then growing the set with examples from real traces every week.
Evaluation is a feedback loop
Evaluations, in this view, are not a perfect test suite. They are a way to keep shipping while limiting risk. They catch regressions when prompts change, make model switches safer and detect drift after launch. A starter set might check format, policy compliance, citations, correct refusals, grounding in the provided context, cost and speed limits, and retrieval quality.
Write rubrics from outcomes
The guide recommends writing grading rubrics like product requirements. A good rubric checks whether the user can take the right next step. For a support reply, that means acknowledging the request, applying the right policy, using real account context without inventing details, taking an action only when allowed, and stating the next step. In specialist fields, domain experts should own the rubric.
Use AI judges carefully
Model based judges help with things like tone and completeness. The author prefers pass or fail verdicts or named categories over numeric scores, which he finds unstable. Anything that can be checked with code, such as formats and banned actions, should be checked with code.
Keep checking in production
Continuous evaluation means running light checks on live traffic and alerting on drops in pass rates, cost spikes, unusual tool calls, poor retrieval and slowdowns. Every failure becomes a new test case.
A minimal system for this week
The guide ends with a short plan: record traces, curate twenty cases, add four to seven checks, set two or three alerts and review weekly. The payoff is confidence. Teams stuck on older models because they fear breakage can upgrade once they can measure the effect. In the author's words, the moat is measured improvement, not prompt cleverness.
Where Innflow fits
Innflow is an all in one workspace for AI agents and workflows. Many of the risks in this guide appear the moment AI can take action, and that is where Innflow focuses.
In Innflow, agents work alongside fixed workflows, so steps that must always happen the same way, like recording a refund, do not depend on a prompt. You can place approval steps before actions with real consequences. And you can review what each agent and workflow did, which supports the habit of turning failures into lessons.
Make reliability a routine
Pick one AI feature, start recording what it does and write down twenty real cases this week. If you want one place to build agents and workflows with guardrails around their actions, try Innflow at innflow.ai.
Source: Arsh Shah Dilbagi, Adaline Labs, "AI Observability And Evaluations: The Operating System For Reliable LLM Products". Read the original.












