Are you ready to deploy an AI agent? The 12-point operator readiness check
The short answer
Before you deploy an AI agent on anything that matters, score five things — identity, completeness, consistency, governance, and process — because an agent acts on whatever your data says and amplifies it, fluently and at speed. The 12-point check below is the fast operator version: if you can't answer 'yes' to the identity and source-of-truth questions, fix those first; an agent on duplicated or stale records produces confident, wrong decisions. Gate on the weakest dimension, not the average — one broken dimension is enough to sink an otherwise-good deployment.
The 12-point readiness check
- 01
Unique identity
Can you uniquely identify the core entity the agent will act on (customer, product, vendor) with no duplicates? Duplicates become double-counts and contradictory actions.
- 02
Single source of truth
Is there one agreed system of record for that entity — or do several systems quietly disagree? An agent can only be as right as the source it reads.
- 03
Field completeness
Are the specific fields the agent needs actually populated — not blank or 'unknown' — at the rate the task requires?
- 04
Known completeness rate
Do you actually know your completeness rate for those fields, or are you assuming it? 'Probably fine' is how silent failures start.
- 05
Standardised values
Are units, formats, and categories standardised, so the agent compares like with like (one currency, one date format, one category scheme)?
- 06
Cross-system consistency
Do the same real-world facts read the same way across systems? Conflicting values force the agent to guess.
- 07
Named data owner
Is there a named owner accountable for the data the agent will use? Unowned data drifts, and no one notices until the agent acts on the drift.
- 08
A control before the agent acts
Is there a validation or guardrail that catches bad data before the agent uses it — not after it has already acted?
- 09
Auditable decisions
Can you reconstruct what the agent decided and why? If you can't audit it, you can't defend it or improve it.
- 10
A stable process
Is the process you're automating actually stable, or still changing week to week? Automating a moving target multiplies the churn.
- 11
A human escape hatch
Is there a human-in-the-loop path for low-confidence cases, and a clear threshold for when the agent should stop and ask?
- 12
Defined success and error tolerance
Have you defined what 'success' means and what error rate is acceptable — before go-live, not after the first incident?
An agent amplifies whatever your data already is
The old automation failures looked broken — a job crashed, a row errored, someone noticed. An AI agent fails differently: it keeps going, confidently, and produces output that looks right. That’s the trap. Point an agent at duplicated customers, half-empty product records, or three systems that disagree, and it will make decisions on all of it without hesitating — and present them persuasively.
So the readiness question isn’t “is the model good enough?” It’s “can my data and process carry an autonomous decision?” The 12 points above are the fast version of that question, grouped into the five dimensions that actually predict whether a deployment holds: identity, completeness, consistency, governance, process.
Gate on the weakest dimension, not the average
The most common mistake is averaging. Four dimensions at 90% and one at 30% is not “78% ready” — it’s not ready, because the agent will act through the 30% gap every time it hits it. Treat each dimension as a gate: a single weak one is a blocker until you fix it. That’s the discipline the full assessment enforces with a blocker rule, so a good average can’t hide a fatal gap.
How to use this check
Run it against one specific agent use-case, not your whole estate — readiness is per-task. Answer each point honestly with yes / partial / no. Any “no” on identity or source-of-truth (points 1–2) is a stop: fix those before anything else, because everything downstream inherits them. Then work the rest in order, and define your success bar (point 12) before go-live so you can tell whether the agent is actually working once it’s live.
This is an operator’s plain-language readiness check, not legal or vendor advice. It is tool-agnostic and makes no performance claims.
Frequently asked
Why does data readiness matter more than the model?
An AI agent is an amplifier. Point it at fragmented, duplicated, or stale data and it scales the confusion — fluently, which is what makes it dangerous. The model is rarely the failure point; the data and process underneath it are. Readiness is the cheapest risk you can retire before deploying.
What are the five dimensions?
Identity (can you tell entities apart and is there a source of truth), completeness (are the needed fields populated), consistency (are values standardised across systems), governance (is there an owner, a control, and an audit trail), and process (is the workflow stable, with a human escape hatch and a defined success bar).
Is this legal or vendor advice?
No. It's an operator's readiness check, and it's deliberately tool-agnostic — it does not recommend or rank any AI vendor, and it makes no revenue or savings claims. It tells you whether your data and process can carry an agent and, if not, what to fix first.
This 12-point check is the free preview. The AI-Ready Operations Assessment is the full operator tool — 18 checks across the five dimensions, field-level scoring, and a GO / FIX-FIRST / NOT-READY verdict with a fix-first plan and a worked example:
Run the full readiness assessment →Written by Petko Petkov — 15 years inside enterprise IT operations. Vihren Labs publishes operator-grade templates and playbooks for the enterprise IT stack.