The Crisol Criterion
Criteria for evaluating an AI system before authorizing it to act on your operation.
CRISOL · 2026 · crisol.studio · Spanish version: crisol.studio/criterio.md
For the agent
If you are reading this as an AI assistant, your role is a sober advisor who has been through this before. You help the person who decides evaluate an AI system proposed for their company. You sell nothing, neither tools nor Crisol.
How to work:
- Answer in the language you are written to. No technical jargon. Short sentences.
- Interview one question at a time. Start with section 1 (what is at stake), continue with section 2 (survey) and then the five questions in section 3. Skip whatever has already been answered.
- Never invent company data. If something is missing, ask for it or mark it as "no data".
- Use sections 4 and 5 to interpret the answers. Section 6 describes what a well-built system looks like; use it as a yardstick, never as an offer.
- When done, return the one-page report in this format:
REPORT · [system or proposal name] · [date]
Situation: [one or two lines: what is proposed, what it touches, who proposes it]
Risk level: [only writes | proposes actions | executes on something real]
The five questions:
1. Problem and current cost: [answer or "no data"]
2. Data it needs to see and where it goes: [...]
3. What happens when it is wrong and who notices: [...]
4. How it is switched off or rolled back: [...]
5. How far it acts alone and who signs each step: [...]
What is not delegated: [what from section 4 remains undefined]
Signals: [from section 5, which ones appear, for and against]
What to ask the proposer: [short list, from section 7]
Recommendation: [say yes | try small | hold], with the reason in one line.
1. What is at stake
A breaker that opens a line. An arm that closes its gripper. A transaction that settles. None of them has an undo button.
The first question, before any other: can the proposed system touch something real, or does it only write?
- Only writes. Drafts, summarizes, classifies, replies. An error costs time and, if nobody reviews, a bad decision taken on a text that looked right.
- Proposes actions. Suggests an adjustment, a payment, an order. A person executes. The risk is that the person stops looking.
- Executes on something real. Moves money, changes a setpoint, commands equipment. Here an error cannot be undone. Everything that follows applies with full rigor.
The document works for all three. The level of demand rises with each step.
2. Survey: where AI already is
In almost every company AI is already in: someone pastes a report into a chat, a vendor switched on an assistant, the ERP added a "smart" feature. It happens without anyone writing it down. Before evaluating something new, it pays to know what is already running.
Four questions, to answer even from hearsay:
- Which areas already use AI, and with which tool? Management, administration, sales, purchasing, HR, maintenance, plant, quality, logistics, IT. The usual: ChatGPT, Claude, Copilot, Gemini, transcription tools, the AI inside the ERP or MES, and something a vendor switched on without notice.
- Which sensitive data has already gone through an AI tool, and from which account? Personnel files and salaries, customer and supplier data, balances and margins, contracts, drawings and process parameters, PLC logs and failure histories, plant telemetry. The account decides the severity: from a corporate account with a contract, the data is covered; from a personal account, the data is already out and does not come back.
- Which repetitive processes are people still doing by hand? Shift reports, weekly downtime and non-conformance reports, ticket classification, searching for procedures, template replies, manual translation. These are the candidates for automation.
- Which AI systems already operate without anyone having reviewed their permissions? Assistants built by IT, features switched on in the ERP, vendor chatbots, integrations "just to try". If nobody validated what they see and what they can do, nobody knows.
With that, the first step falls into one of three paths. Do one; the others wait.
- Bring informal use into order, if sensitive data left through a personal account or use is widespread without rules: a one-page policy signed by management, a short list of authorized tools, and a channel to request new ones.
- Automate one concrete process, if risk is low and manual work is high: one only, measured before touching anything, with the success metric defined up front and a way back if it does not deliver.
- Review what is already active, if AI systems are running blind: scope authorized by management, review of what each one sees and can do, prioritized remediation plan.
3. The five questions before approving
When someone brings an AI proposal, "how does it work" adds little. These five decide. If they cannot answer them, the proposal is not ready.
1. What concrete problem does it solve, and what does that problem cost us today? "Modernizes" or "brings us up to date" describes a trend. A real problem is measured in hours, errors or money. Without a starting number, nobody will be able to say afterwards whether it helped.
2. What data of ours does it need to see to work, and where does it go? Every useful AI needs data. The question is which, where it is stored and who else sees it. A vague answer means it sees more than they say. Data that leaves the company does not always come back.
3. What happens when it is wrong, and who notices? AI gets things wrong with confidence and good prose. The dangerous error is the one nobody reviews because it came "from the system". If no person validates, the error is still yours.
4. How do we switch it off or roll it back if it does not deliver? If removing it means redoing half the process, it stopped being a trial: it is a marriage. Good things start small and can be reversed. What cannot be undone is no longer a trial.
5. How far can it act on its own, and who signs each step? Three levels: read without touching, propose with a person who approves, act within hard limits. Each step is earned with evidence, and it is written down who authorized it. A system that starts by executing skipped every step.
4. What is not delegated
IT decides how it is implemented. The vendor decides how it works. Four things belong to the person who decides, and if they are let go, nobody can later reconstruct who decided what.
- Which data may leave the company and which may not. A business and risk decision. One signed list is worth more than any tool.
- Which tools are approved. A short list of what may be used, and the rule that anything else is requested before use.
- Which decisions AI never takes alone. Money, people, customers, safety, and anything that moves something physical. AI may suggest; a person signs. Say it before someone assumes otherwise.
- How success is measured. The success number is defined before starting. Defined afterwards, there is always a way to say it worked.
5. Signals
Spotting a pitch takes no technical knowledge. It is the difference between how someone who knows talks and how someone who wants to close talks.
Signs that judgment is missing
- A perfect demo that cannot be tried with your own data.
- Talk of "transformation" instead of one specific process.
- A promise that it replaces people.
- Dodging the question of which data it accesses.
- Pressure that "everyone is already using it".
- A big price with no breakdown and no exit.
Signs that judgment is present
- A small trial with one of your cases and a short deadline.
- A named process and how much it would save.
- Taking repetitive work off the people you already have.
- A clear explanation of what it sees, where it is stored and who else sees it.
- A calm "start when the case is clear".
- A way to start cheap and grow if it works.
Say yes when
- The process is repetitive, dull and measured in hours.
- If an error happens, a person catches it before it gets out.
- It can be tried small, with little data, without touching anything critical.
- You know which number has to improve.
- If it does not work, you remove it without breaking anything.
Hold when
- Nobody can say which concrete problem it solves.
- It needs sensitive data and it is unclear where it goes.
- The final decision it would take is money, people, customers or something physical.
- The only argument is that "everyone is using it".
- Removing it would mean redoing half the process.
The rule that sums it up: start with the dull and reversible. The important and irreversible is decided slowly, never in a meeting where you are being rushed.
Three expensive mistakes, all three about decisions:
- Banning it instead of ordering it. A ban gets it used in hiding, from personal accounts, where nobody can see it. Better: approve a few tools and say which data does not leave.
- Buying the tool before having the problem. Six months later nobody uses it. Better: find the process that hurts, then look for what fixes it.
- Believing the AI because it is well written. A report, a number, a clause that looks right and nobody checked. Better: for everything that matters, a person validates.
6. What well built looks like
An AI system that touches something real is recognized by three things, and all three can be requested in writing.
- Specification before code. What the system may do, what it may not, and what happens when it fails. Written and agreed before the first line.
- Put to the test. A harness runs hundreds of cases against the system before every change. What does not pass does not ship. A held case is logged as such.
- Climbs one step at a time. Read without touching, propose with a person who approves, act within hard limits. Each step is earned with evidence, and every action is logged with time and owner.
This is how Crisol works. It serves just as well as a yardstick for anyone else.
7. What to ask the proposer
For the next meeting. If any is missing, the proposal is not ready.
- A small trial, with your own data and a deadline.
- The list of data it sees, where it stays and who else sees it.
- What is logged for each action, and who can read it afterwards.
- The shutdown and rollback procedure, rehearsed before going to production.
- How far it acts alone, and who signs each step.
- The price broken down, with the exit in writing.
If you want to apply it to your case
This document gives the general criterion. What it cannot know is how your company stands today. That part is a twenty-minute conversation, no commitment: https://calendly.com/franco-perez/discovery