Back to the catalogue

Buying AI

AI Vendor Evaluator

They want to sell you an AI. The salesperson shows an impeccable demo, you nod and don't know what to ask. This tool does the preliminary work. Sort out what you really need, classify the risk of your case and generate specific questions for the provider. Where your data lives, if they train their models with it, who their subcontractors are, what level of service they guarantee, what happens the day you want to leave and how much it costs to get your information. Then build a comparison table with weights and mark the warning signs. If evidence is missing, leave the box as unknown and do not invent a score. A critical alarm blocks the recommendation even if the total score is high. And if your case concerns employment, health, credit, security or sensitive data, activate a review recommendation by an expert.

iamericas-evaluador-proveedores
Download skillv2.0

Free download, no sign-up

What problem does it solve?

You start by telling what problem you want to solve. If the first thing you ask is which one to choose, it doesn't give you a name. He explains that first you have to compare evidence, and begins to collect it. The skill requires the minimum to advance, without long questionnaires. Sector, country, use case, who it affects, what data is processed, budget, candidates, what it has to integrate with and what restrictions you have. When there are several candidates, it assembles the comparative table, applies the weight rubric and validates it by script before showing it to you. It never presents you with an analysis with an invalid table. Important: IAméricas does not certify or endorse suppliers. The skill returns a more informed decision.

What does it produce?

A comparison table with weights and scores, validated by script, with boxes without evidence marked as unknown.
  • The list of questions for the supplier, written to send as is.
  • The record of alarm signals detected and which ones block the decision.
  • A decision memo intended for management to read.
  • A validation plan for the test or pilot, what to look at and what has to happen to say yes.
  • A follow-up plan for when the tool is already working.

When to use it

  • You have two or more candidates and you don't know what criteria to compare them with.
  • It's time to renew a contract and you want to review it with a cool head.
  • The provider gives you a demo next week and you want to arrive with questions prepared.
  • Your case involves employment, health, education, credit or sensitive data and you need to record how you decided.
  • You want to quickly discard a tool that doesn't fit, without missing three meetings.

What you need first

  • What problem do you want to solve and what process does it involve?
  • What country do you operate in and what sector are you from?
  • Who is affected by what the tool does.
  • What data will be processed, especially if there is personal data.
  • What systems of yours does it have to integrate with?
  • Approximate budget and names of candidates.
  • Any documentation they have given you helps a lot. Commercial proposal, contract, privacy policy, security sheet or technical documentation.

Where it works and how to install it

Pick your platform and follow the steps. Where the tab is labelled “Native” the file installs as-is; ChatGPT and Gemini do not install it yet, so there you paste the content of SKILL.md into the chat.

Method A — Upload the file (easiest)

  1. 1Download the file with the button above. No need to unpack it.
  2. 2Open Claude — the desktop app on Windows or Mac, or claude.ai — and go to Customize → Skills.
  3. 3Click “Add” and upload the file as it is. If the form rejects the extension, rename .skill to .zip and upload again: same file.
  4. 4The skill shows up in your capabilities list, ready to switch on.

It is installed. Turn it on from the capabilities panel, or type @ followed by its name inside any conversation.

Version history

  • v2.0

    5 October 2026 · 195 kB

    The assessment stops being a judgement call and becomes a calculation. The skill used to hand over a rubric and a scorecard to fill in by hand; it now profiles the company through 40 questions in nine blocks —four or five per message, never as a form— derives four axes from them (technical capacity, risk, scale and budget) and uses those to decide what level the company needs on each of the 34 vendor criteria, with what weight, and what minimum cannot be waived. The score, the knock-out filters, the budget adjustment and the ranking are all computed by a script, so two identical assessments give the same result. Each criterion is described at three levels in plain language and states where to verify it and what to ask the vendor. What has no evidence is not assumed: it counts as the lowest level and is flagged as pending, and the skill returns the questionnaire written out to send to the vendor and recalculate once they answer. The report is written by a script that rejects unsupported figures, internal codes, links and compliance judgements, and includes the system's operating limits and the minimum contract clauses. A glossary for explaining terms without jargon and a lookup of each country's data law are added. It works in Spanish, Portuguese and English, and answers “which vendor do I pick”, “is this one right for me?” with a single vendor, and switching vendors, where the current one enters as the baseline.

  • v1.2

    21 September 2026 · 30 kB

    Changes how the conversation starts. If the first message brings no context, the skill explains how the process works before asking anything: the phases, that it can be paused and resumed, and what you get at the end; and it notes that the input can be typed or attached as a file. If the context is already there, it confirms in two lines and gets going without that explanation. The description is rewritten to state that this is a guided, step-by-step process.

  • v1.1

    14 September 2026 · 26 kB

    Neutral Spanish throughout. It works out which Python interpreter the system has, states which folder the scorecard validator runs from, and `--help` is no longer treated as if it were a filename.

  • v1.0

    1 August 2026 · 19 kB

    First version.

Other skills on the same topic