What this template is for
Start the register before writing or revising an AI use policyA team's written rules for AI: approved tools, what each use may do, what data may go in, what is reviewed, when customers are told and who owns problems., because the policy's first step is to list current uses; keep it up to date afterwards, because every approval, change and incident lands on a row. One register serves the whole marketing team. Each use of AI is a row: a task, done with a particular tool or feature, by a particular team, with a particular output going somewhere. A tool used for three tasks is three rows.
The governance hub describes the register's contents in one sentence: each use, its tier, its owner, the tool, what data it sees and what it can do. It names three companion records, a version history for prompts, models and settings, review logs and an incident list, and this template gives each of them a column that points to where it is kept, so one row leads to everything known about a use. The policy guide builds the register in its first step and sorts it into tiers in its second; the guide to standalone assistants and built-in AI asks that default-on features be included; the guide to evaluating AI tools ends every decision with an entry in the register; and the guide to handling an AI mistake gives the incident to the owner named in it. None of them lays the register out. In the terms of the NIST AI Risk Management Framework, the register is the map: where AI is used and what could go wrong.
It is not the policy, which sets the rules, nor the review workflow or the incident procedure, which this page links rather than repeats. It is not a tool evaluation: the martech tool evaluation template scores candidate tools before one is chosen, and the register records the uses that follow. For automations generally, the automation governance guide keeps a workflow inventory; an AI step inside a workflow appears in both, with the same name. Decisions that change how a use runs, such as moving it up a tier, are also worth a line in the marketing decision log. For one customer-facing use about to go live in service, the AI customer service launch checklist is the check before launch; the register is the standing record.
The template
Keep the register as a shared spreadsheet or table with these columns, in this order, and one row per use. Keep the column names so the policy, the review logs and the incident list can refer to them.
| Field | What to write |
|---|---|
| Use | The task the AI does, in one sentence, and the team that relies on it |
| Tool and feature | The tool, the specific feature, and whether it was switched on by someone or arrived on by default |
| Risk tier | The tier the use sits in, from internal drafting to customer-facing or sensitive, and the reason |
| Owner | The named person who answers for the use, reviews its logs and decides changes |
| Data it sees | What information goes in, where it comes from, and what is removed before it is used |
| What it can do | Whether the output is a draft for a person, published after review, or acts or replies on its own |
| Review route | Who checks the output before it is used, against what, and how the check is recorded |
| Disclosure | Whether and how people are told that AI was involved, and where that wording lives |
| Approval and vendor terms | When the use was approved, against which evaluation, and what the vendor's terms say about data |
| Versions | The prompt, model and settings in use, with the date of each change |
| Incidents | What went wrong, who noticed, what changed afterwards, and the date |
Add a status for each row, such as in use, pilot, paused or stopped, and keep stopped rows rather than deleting them: the history of what was stopped and why is part of the record.
How to fill each section
Use
Write the task as a sentence a colleague would recognise, with the team that relies on it, such as "drafts first versions of tour pages for the content team". The governance hub frames the register around uses because the same model carries different risks in different tasks. If a sentence needs "and", it is probably two uses.
Tool and feature
Name the tool, the plan and the specific feature, not just the vendor. Say whether someone switched it on or it arrived on by default. The guide to standalone assistants and built-in AI explains why built-in features are routinely missed and why default-on features belong here like any other use. Check each tool's settings yourself rather than relying on what people remember using.
Risk tier
Give the use one of the hub's four tiers: internal and reversible, published with an editor, customer-facing and automated, or regulated or sensitive. The governance hub tiers by what happens if the output is wrong and nobody catches it, and treats the tiers as a planning device rather than a legal classification. Write the reason in a few words. A use that grows, for example a drafting assistant that starts sending on its own, changes tier and needs a new review; treating that move as a gate, with the evidence and the decision-maker agreed in advance as in phase-gate reviews, stops it from happening quietly.
Owner
Name one person, not a team. The policy guide names owners as part of the policy, and the hub defines accountability as one named owner per use who can answer for it and who reviews its logs on a schedule. When the owner leaves or changes role, the row changes the same day.
Data it sees
Record what goes in, from which system, and what is removed first. Most customer records are Personal DataAny information relating to an identified or identifiable living person, such as a name, email address, location data or an online identifier., and the data minimisation routine keeps only the fields a task needs and asks for a short record of what was shared next to this entry. The disclosure and data privacy guide sets out the data classes and which tools each may enter, under the GDPRThe EU regulation setting the rules for processing personal data, including its principles, lawful bases, individuals' rights and breach duties. as the site's guides describe it. A row whose data breaks the policy's data rule is a use to stop, not to record and carry on.
What it can do
Write one of three answers: a draft a person reads before anything happens, output published after review, or output that acts or replies on its own. This column, more than the tool, decides the tier. Uses that act on their own need the guardrails, sampling and logs that the accuracy and review workflows guide attaches to the customer-facing tier.
Review route
Name who checks the output, what they check it against and where the check is recorded. The accuracy and review workflows guide sets review by tier, from no formal review for internal uses to specialist sign-off for regulated ones, and describes GroundingMaking a language model answer from specified material, such as documents retrieved for the question, rather than from what it absorbed in training. in sources, fact and claim checks and the review log. This column points to that log; it does not repeat the method.
Disclosure
Say whether people are told that AI was involved, how, and where the approved wording lives. The disclosure and data privacy guide gives the default positions, including telling people when they are interacting with an automated system and labelling synthetic people, voices and realistic generated media. Internal uses usually need nothing here; write "none needed" rather than leaving the cell blank, so a gap means something.
Approval and vendor terms
Record the date the use was approved, the evaluation it was approved against and what the vendor's terms say about inputs: training, retention and where the data lives. The guide to evaluating AI tools gives the questions and the outcomes, such as approval for named uses and tiers only or a time-limited pilot with measures and an end date, and asks that the tool enter the register with its approved uses, plan, owner and review date. Link the evaluation rather than copying it.
Versions
Record the prompt, the model and the settings in use, and date every change. A Language ModelAn AI system trained on large amounts of text that writes by predicting which piece of text is likely to come next. can change behind a product with little notice, and the guide to testing prompts and models lists the changes that call for a rerun of the stored test set. The hub keeps a version history so that a change in behaviour can be traced to a change in configuration; this column holds the current version and links that history.
Incidents
Record each incident in a line: what went wrong, who noticed, what changed afterwards and the date, with a link to the full entry in the incident list. The guide to handling an AI mistake runs the response and asks that each incident change at least one thing in the system, such as a tighter tier or a new check. If a row's incident line never shows a change, the incident was closed too early.
Filled example
Below, the register is completed for an invented business. The names, numbers and dates are illustrative.
Gulvemar Tours writes its first AI use register
Gulvemar Tours is an invented operator of small-group walking and food tours, with a marketing team of four. Its marketing lead built the register in March by asking each person what they used and then checking the settings of every tool in the stack. That check found a default-on feature nobody had listed, and one use that had to stop. Five of the eleven rows are shown.
| Field | Row 1 | Row 2 | Row 3 | Row 4 | Row 5 |
|---|---|---|---|---|---|
| Use | Drafts first versions of tour pages (content team) | Suggests email subject lines (email) | Drafts replies to pre-booking questions for an agent to send (sales desk) | Summarises post-tour feedback into themes (marketing lead) | Rewrites booking notes into "welcome" emails (sales desk) |
| Tool and feature | Standalone assistant, business plan, drafting | Email platform, subject-line suggestions; arrived on by default in February | Service inbox, reply drafting; switched on by the sales lead | Standalone assistant, business plan, file analysis | Consumer chat tool, free plan |
| Risk tier | Published with an editor: reaches customers after an edit | Published with an editor: the email owner picks or rejects each line | Published with an editor: an agent reads and sends every reply | Internal and reversible: used only for planning | Customer-facing in effect: notes pasted unreviewed into emails |
| Owner | Content editor | Email marketer | Sales lead | Marketing lead | Sales lead |
| Data it sees | Tour itineraries and the brand's own copy; no customer data | Previous subject lines and campaign names only | The customer's question and the public help pages; names removed before drafting | Feedback text with names and emails removed by the export | Booking notes including customer names and dietary needs |
| What it can do | Draft only | Suggests; a person chooses | Draft only; nothing sends on its own | Draft only | Draft, but copied straight into emails |
| Review route | Content editor checks facts against the itinerary and records sign-off in the page ticket | Email marketer reviews each line against the tone guide | Agent reads every draft; sales lead samples sent replies each week | Marketing lead reads the full summary before use | None |
| Disclosure | None needed | None needed | None needed: a person sends each reply | None needed | None |
| Approval and vendor terms | Approved January after a trial; terms forbid training on inputs | Approved in March after the discovery; vendor terms read and filed | Approved April for drafts only; terms filed with the evaluation | Approved January with row 1 | Not approved; free plan's terms allow training on inputs |
| Versions | Shared page prompt v3 (12 March) | Vendor default, no custom prompt | Reply prompt v2 (2 April); answers grounded in help pages only | Summary prompt v1 (20 January) | None recorded |
| Incidents | 14 March: a draft gave a wrong meeting point; caught by the editor; itinerary added to the prompt's sources | None | None | None | Stopped 6 March: personal data in a consumer tool. Notes deleted from the tool's history; welcome emails written from a template instead |
Row 5 is kept with the status "stopped". At the first quarterly review the register showed eleven uses, all with a named owner, one incident with a recorded change, and the reply drafter on a watch list: the sales lead proposed letting it send answers to opening-time questions on its own, which would move it to the customer-facing tier, and the team agreed to treat that as a new approval rather than a setting change.
Common mistakes
- One row per tool. A register of tools hides the difference between a summary nobody sees and a reply a customer reads. List uses, and let a tool appear on several rows.
- Only the tools people chose. Default-on features in the email, ad and analytics platforms are uses too. Check the settings, not just the memories.
- A team as the owner. "Marketing" cannot review a log or be asked a question. Name one person per row.
- Recording a use that breaks the data rule and carrying on. The register is where such uses are found; the row should end in "stopped", not in a note.
- A tier set once and never revisited. Uses grow. A draft tool that starts sending has changed tier, and the change needs a new review before it happens.
- No versions. Without the prompt, model and settings recorded by date, nobody can say what changed when the output did.
- Incidents logged without a change. An incident line with no system change attached will usually happen again.
- A register nobody opens. Review it on the schedule the policy sets, and treat the share of rows with a current owner and a recent review as a measure of whether governance is working.
The tool evaluation and decision log templates that sit beside the register are in the templates library. The academy capstone on an AI operating model for an online course provider practises building a year's operating model on top of a register like this one, with a named owner for every part.
