# Inside My AI Tool Register: 10 Design Decisions and the Alternatives I Rejected

> The ten design decisions behind our AI tool register, the alternatives I rejected, and the frameworks behind each rule. For IT, security, and compliance people.

Canonical: https://nosistech.com/ai-tool-register-design-decisions/
Published: 2026-11-15
Last updated: 2026-11-18
Author: Carlos, NosisTech LLC

My last post introduced an AI tool register for small and mid-size companies. That post was for business owners. This one is for the people who must build and run the register: IT leads, security and compliance people, and consultants who advise clients.

Before I designed it, I studied public inventories and risk frameworks. They included the NIST AI Risk Management Framework, its Generative AI Profile, the US federal AI use case inventory, and the OWASP lists. Weak registers failed in the same ways:

- They listed vendors, not uses, so one row hid five different risks.
- They missed AI inside existing software, and AI that staff use on personal accounts.
- They scored "the AI could be wrong" as 5 out of 5 for every tool, so the scores meant nothing.
- They gave credit for controls that existed only on a roadmap.
- They were a one-time spreadsheet that nobody opened again.

I designed the template around these failures. Below are the ten decisions that matter most, the alternatives that I rejected, and the frameworks behind each one. I quote the template directly, so you can judge the wording yourself.

---

## Decision 1: One row for each use, not for each vendor

**The alternative I rejected:** one row for each product or vendor.

A vendor list is easy to build from invoices. But risk does not sit in the product. It sits in the task, the data, and the people who see the output.

> **From the template, section 2.1:**
>
> The same tool can be safe for one task and risky for another. A chat assistant that drafts internal emails is low risk. The same chat assistant that drafts answers to customers about their accounts is higher risk. Give each use its own row.

The cost is a longer list. The benefit is that each row has one owner, one data tier, and one score that means something. NIST AI RMF [GOVERN 1.6](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf#page=28) asks for mechanisms to inventory AI systems. [MAP 1.1](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf#page=31) asks you to understand the context of each use. A use-level row does both.

## Decision 2: One score, for the use as it runs today

**The alternative I rejected:** two scores for every row, one before controls and one after.

Some registers score each risk twice. That method works well, but it needs a trained person to run it. A template must work without one.

So this template asks one question: how risky is this use today, with what you really do? The "Controls in place" column records the answer:

> **From the template, section 5 (the columns):**
>
> **Controls in place:** What we do today to lower the risk. Example: Business plan; agent sees only its own queue; a person approves each reply.

One honest score is a strong start. It also gives a trained reviewer a clear baseline later.

## Decision 3: A 5 × 5 scale with four levels

**The alternatives I rejected:** a 3 × 3 scale, a 1–10 scale, and money-only scoring.

- **3 × 3** is too coarse. Most uses land in the middle, and the register cannot tell them apart.
- **1–10** looks precise, but nobody can explain the difference between a 6 and a 7.
- **Money-only** scoring needs loss data and an experienced person to estimate it. A template cannot supply either.

A 5 × 5 scale gives each number a plain meaning. Likelihood runs from "Rare" to "Almost certain". Impact runs from "Negligible" to "Severe". The product, from 1 to 25, maps to four levels, as in section 6.4:

| Score | Risk level | Review at least every |
| --- | --- | --- |
| 1–4 | Low | 12 months |
| 5–9 | Medium | 6 months |
| 10–15 | High | 3 months |
| 16–25 | Critical | 3 months |

Each level has an approver and a review interval. A Critical use has no approver at all until the score goes down. Only the CEO can accept a Critical risk, in writing, for a limited time. This follows [MAP 5.1](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf#page=32), which pairs likelihood with magnitude. It also follows [MANAGE 1.3](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf#page=37), which names four responses: mitigate, transfer, avoid, or accept.

## Decision 4: Score the harm that reaches someone, not the AI's small errors

**The alternative I rejected:** scoring the chance that the AI makes a mistake.

Generative AI makes small mistakes every day. If you score that chance, every tool gets a 5 for likelihood, and the register becomes useless. So the template scores something else:

> **From the template, section 6.2:**
>
> Score the chance that the harm reaches someone, not the chance of a small error. AI tools make small errors every day. A person who checks the output stops most of them.

This one sentence makes the "Human check" column matter. A draft that a person reviews has a lower likelihood than the same draft sent automatically. The register now rewards good process, not just good vendors.

## Decision 5: Minimum scores for the uses that hurt people

**The alternative I rejected:** open judgment for every score.

Open judgment fails in a predictable direction. The person who wants the tool scores it low. So the template sets floors:

> **From the template, section 6.5:**
>
> 2. If you are not sure between two numbers, choose the higher one.
>
> 3. If the use affects decisions about people, the impact is at least 4.
>
> 4. If the tool sees Restricted data and the vendor terms are "Unknown", the impact is at least 4.

The register defines "decisions about people" as hiring, pay, credit, health care, housing, insurance, or education. Several AI laws treat these areas as high-risk. The floor makes sure that those uses never look small in your register.

Each industry example adds its own floors. The clinic sets impact at 4 or higher for clinical uses. The software company looks at agents. If an agent can take an action that is difficult to reverse without a person's approval, the likelihood is at least 4. That second rule addresses Excessive Agency, LLM03 in the [OWASP Top 10 for LLM Applications 2026](https://genai.owasp.org/resource/owasp-genai-llm-top-10-2026/).

## Decision 6: A planned control does not count

**The alternative I rejected:** credit for controls "in progress".

Roadmaps are optimistic. If a planned control lowers the score today, the register reports a risk that does not exist yet.

> **From the template, section 6.5:**
>
> 1. Score the use as it runs today, with the controls that are really in place. A control that is only planned does not count yet.

The planned control still has a place. For a High use, the approver needs "a written plan to lower the score". When the control goes live, you rescore the row. The register then shows the improvement with a date.

## Decision 7: The plan is part of the tool name

**The alternative I rejected:** the product name alone.

The free plan and the business plan of the same AI product can have very different data terms. Some free plans let the vendor keep your inputs or train on them. So the template puts the plan into the name, and it adds a column for the terms:

> **From the template, section 5 (the columns):**
>
> - **Tool and plan:** The product name and the plan. The plan matters: free and business plans often have different data terms.
>
> - **Vendor terms: no training on our data?** Yes, No, or Unknown. Check the contract, not the marketing page.

"Check the contract, not the marketing page" saves a lot of trouble. Marketing pages change without notice, and they rarely bind the vendor. [GV-6.1-007](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf#page=25) in the NIST Generative AI Profile covers approved provider lists. The plan column makes each approval specific.

## Decision 8: Prohibited uses stay in the register

**The alternative I rejected:** a register of approved tools only.

If the register lists only approved tools, every "no" lives in someone's email. Six months later, nobody remembers why. So the template records the no:

> **From the template, section 2.2:**
>
> - free or personal AI accounts that staff use for work, with the status "Prohibited", so that the decision is on record

The same logic keeps retired tools. A "Retired" row stays in the register with the date. The retirement steps cover accounts, API keys, data export, and vendor deletion. That is [GOVERN 1.7](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf#page=28), on decommissioning AI systems safely.

## Decision 9: AI inside existing software counts as a new tool

**The alternative I rejected:** registering only the AI products that someone bought.

Today, most new AI in a small company arrives through updates. Email, CRM, help desk, phone systems, and design tools add AI features every quarter. Nobody signs a new contract, so nobody checks.

> **From the template, section 8.3:**
>
> Every [3] months, ask each team which AI tools they use. Also check the AI features that your existing software added since the last check. Add each new use with the status "Under review".

The NIST Generative AI Profile asks you to define any inventory exemptions for AI inside application software ([GV-1.6-002](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf#page=20)). My answer is a very short exemption list. Every other AI feature is a new use, so it gets a row. The template gives one example of an exemption: spell check that sends no data outside the software. The rule for doubt is clear: "If you are not sure, register the feature."

## Decision 10: Review dates set by risk, plus events that force an early review

**The alternative I rejected:** one yearly review for the whole register.

A yearly review is too slow for a Critical use and too much work for a Low one. So the Excel file calculates the next review from the decision date and the level. The table in Decision 3 shows the intervals.

Dates are not enough. AI tools change between reviews. So the template lists events that force an early review:

> **From the template, section 8.2:**
>
> - the vendor changes its model, its terms, or its owner
>
> - the tool starts to see a higher data tier
>
> - people outside the company start to see the output or talk to the tool
>
> - an agent gets new permissions
>
> - an incident or a near miss happens
>
> - a new law applies to the use

This follows [GOVERN 1.5](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf#page=28), which asks you to plan periodic reviews and set their frequency. It also follows [MANAGE 3.1](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf#page=37), which asks you to monitor the risks of third-party AI. The software company example goes further: "A change of model or model version is a product change."

---

## What changes in the three examples

The three industry versions use fictional companies. Each one shows how the same register bends to a different business.

- **Riverbend Family Health**, clinics in Texas and California. Patient data needs a signed BAA that covers the AI feature itself. The Medical Director joins every clinical decision. The reason: HIPAA requires a Business Associate Agreement before a vendor handles patient data.
- **Lumorra Data**, a B2B AI software company. The register has two kinds of rows: staff tools and AI inside the product. A new model provider needs extra steps. The reason: customer contracts make promises about the product AI.
- **Keystone Advisory**, a 12-person consulting firm. Each client's AI choice limits what an approved tool can do. The reason: the client owns its data, so the client decides.

The clinic adds one more duty. Every year, the Medical Director reviews each decision support tool for the risk of discrimination against patient groups. That follows [Section 1557](https://www.ecfr.gov/current/title-45/subtitle-A/subchapter-A/part-92/subpart-C/section-92.210) of the Affordable Care Act and its rule on patient care decision support tools.

The software company turns contract promises into steps:

> **From the B2B AI SaaS version, section 4.1:**
>
> 1. The provider agrees in its contract not to train on customer data. It also meets our retention limits.
>
> 2. We add the provider to our public sub-processor list.
>
> 3. We give customers the notice that their contracts require.

The consulting firm makes one important point: an approved tool is not approved for every client.

> **From the consulting firm version, section 7.3:**
>
> An approved tool is not approved for every client. Before you use a tool on client data, check two things:
>
> 1. The register approves the tool for Restricted data.
>
> 2. The client AI choice allows it. The engagement lead keeps the signed choice in the engagement folder.

## About the references in this post

Each version ends with a short note on security risks. It names the items on the [2026 OWASP list](https://genai.owasp.org/resource/owasp-genai-llm-top-10-2026/) that the risk checklist in section 6.1 covers. The healthcare version also lists the US laws behind its extra rules.

Two notes on the references in this post:

- **I checked every NIST reference against the original NIST documents.** Each link opens the correct page.
- **A reference is not a compliance claim.** NIST and OWASP help to explain why a rule exists. Using the register does not prove that you meet a law.

## How to adapt it

- **Start with the industry version** that is closest to you, even if your industry is different. The decisions transfer.
- **Keep the scales and the levels as they are** for the first year. Change them only when you have a reason that you can write down.
- **Fill the register before you publish your AI policy.** A policy that points to an empty register teaches people to ignore it.
- **Name one tool owner for every row before you score it.** A row with no owner has nobody to keep it correct.

## Get the templates

All four versions are free on GitHub under CC BY 4.0. Each version has an Excel register, a guide, and a request form, in Markdown and Word: [github.com/nosistech/governance-templates](https://github.com/nosistech/governance-templates).

Next in this series is Template 3: the AI vendor security questionnaire. It helps you check a vendor before its tool gets a row in the register.

Have a question or feedback about these templates? [Contact NosisTech](https://nosistech.com/connect/).

### Keep reading: AI governance templates

Previous article: [Which AI Tools Does Your Company Use? Build a Register in One Afternoon](https://nosistech.com/ai-tool-register-template/)

Template 1: [Inside My AI Acceptable-Use Policy: 9 Design Decisions and the Alternatives I Rejected](https://nosistech.com/ai-acceptable-use-policy-design-decisions/)

Template 3: [Inside My AI Vendor Questionnaire: 10 Design Decisions and the Alternatives I Rejected](https://nosistech.com/ai-vendor-questionnaire-design-decisions/)

Topic hub: [AI Governance](https://nosistech.com/ai-governance/)

All articles: [NosisTech Writing](https://nosistech.com/blog/)

*This post is general information, not legal advice. The example companies are fictional. Any resemblance to a real company is a coincidence. The legal references are current as of November 2026, and laws about AI change often. Ask your legal counsel before you rely on any template.*
