Decision Register
For the head of compliance, legal, tax or finance who has handed decisions to AI

Minute every decision a tool takes for you, before the next reporting cycle asks.

Paste the tasks and decisions your function has handed to AI. Get, per line, how much is at stake, which regime attaches to it, the oversight and records it requires, the benchmark to demand from the vendor before the next reporting cycle, and the lines where a tool is deciding alone under a rule that says it cannot. Eight lines free, no account.

Paste your own listClassification and every figure run in your browser against a published dictionary of 86 task types in six functions. Nothing leaves it until you choose to save.
A lawyer in a tweed jacket reading a report in a binder at his desk
Walk into the audit knowing which decisions your function has handed to AI, how much is at stake in each, who reviews it and what to demand of the vendor before the next close, from a task list you can write in ten minutes. Paste that list and get it read against the EU AI Act, GDPR Article 22 and ISO/IEC 42001 and 42005 in your browser, cited so you can check it, with no governance platform to buy.
Specimen: a 400-person lender, GBPa selection of 8 of its 32 lines
MinuteTaskToolStakes
M-01scores consumer loan applications credit scoringsolely automatedin-house modeldecides alone
M-13first-pass KYC review of new customers solely automatedin-house modeldecides alone
M-14flags sanctions matches AML and sanctionsin-house modeldecides alone
M-18classifies incidents and decides whether to notify the regulator no reviewer nameda general assistantdecides with review
M-30classifies transactions for VAT tax filingtax analyst spot-checks 10%a general assistantdecides with review
M-08drafts the SOX narrative and control descriptions financial reportinga drafting assistantdrafts
M-23drafts contract clausesa legal AI tooldrafts
M-32runs the Halbrook parity attestation not minutedin-house modeldrafts
32 lines, 7 decide alone4 tools2 listed uses
The stakes and the listed-use flags (sector flags in grey), as the register reads them from the pasted columns. The full register adds the obligation rows under the EU AI Act, the GDPR and ISO/IEC 42001, ten findings, and the benchmark to demand from each vendor.
One task per line: Task | tool | who reviews | decides or drafts. Tabs, commas or pipes; a header row is optional and its column names are recognised in any order; a CSV export of a use-case list works. A first line eu ai act: yes | entity: lender | reporting cycle: quarterly | as at: 2026-09-20 ticks the regime, names the entity type and pins the cycle and the date. Extra columns the register reads: retest, benchmark, logs, impact assessment, FRIA, notice.
Tool names in the paste are kept as labels; the register never rates a tool. Nothing is sent anywhere until you choose to save.
01

Paste the list you already have

The AI use-case list, the vendor onboarding spreadsheet or the register the last audit asked for: one task per line, with the tool, who reviews and whether the tool informs, drafts, decides with review or decides alone, in any order the header names. Every line is matched against a published dictionary of 86 task types across legal, tax, finance and treasury, compliance and risk, HR and people, procurement and operations; a task that matches nothing is marked not minuted and never guessed.

02

Read what each line decides, and under what

Per line: the stakes class as pasted or derived, the listed-use flags (the EU AI Act's Annex III uses, the GDPR's solely automated decisions, and where the error lands: a financial statement, a tax filing, a regulatory return, a sanctions screen), the error type that matters, the review point and the record it must hold. Per register: the lines by stakes, by function and by tool, the lines that decide alone, the listed uses, the lines with nobody named to review. Each figure is a count of your own lines, never a verdict on a tool.

03

Take the obligations and the benchmark to the vendor

Tick the regimes that reach you and each line carries the rows that attach: the EU AI Act's deployer duties, oversight, logs and transparency; the GDPR's Article 22 with the information duties and the DPIA beside it; ISO/IEC 42001's responsible-use, supplier, operation and impact controls; ISO/IEC 42005's assessment items; the NIST AI RMF's measurement and oversight items; APRA CPS 230's service provider and controls-testing paragraphs. Requirement text in our own words, read clause by clause from the published standards. The vendor benchmark letter exports as one printable page, one paragraph per task.

Why a register and not an AI governance platform

Domain-specific AI, benchmarked privately in legal, tax and finance, is outpacing the general models most firms adopted first, and a firm running off-the-shelf tools on compliance work will meet higher error rates and audit exposure before its next reporting cycle. The platforms are built for the bank with a model risk function and a licence budget. The head of compliance at a 400-person firm has a use-case list, a handful of vendor contracts and a supervisor who now asks which decisions a tool takes alone. Before the platform conversation, somebody has to say what each line decides, who reviews it, which regime attaches and what to ask the vendor to show on the firm's own cases. That is the register this builds, in your browser, from the list you already hold.

The dictionary and the obligation table are ours and published in full: every task type it recognises, in six functions, and what each regime attaches to a decision handed to a tool. The requirement text behind every obligation is in our own words, read clause by clause from the published standards. The register classifies your own lines and states the obligation in our words; it never rates a tool and never rules on a task.

Decision Register: the questions buyers ask

What is Decision Register?

Decision Register takes the tasks and decisions your function has handed to AI and minutes each one. Per line it returns how much is at stake, the regime that attaches, the oversight and records required, and the benchmark to demand from the vendor.

What do I paste into Decision Register?

One task per line: `Task | tool | who reviews | decides or drafts`. Tabs, commas or pipes work, a header row is optional, and a CSV export of a use-case list works.

How many lines can I run for free?

Eight lines are free and no account is needed. Classification and every figure run in your browser against a published dictionary of 86 task types in six functions, and nothing leaves it until you choose to save.

What does it flag as a listed use?

The listed-use flags are the EU AI Act's Annex III uses, the GDPR's solely automated decisions, and where the error lands: a financial statement, a tax filing, a regulatory return or a sanctions screen. Sector flags are shown in grey.

What is the stakes class?

The stakes class says how much is at stake in that task, read from your line as pasted or derived. The register counts your own lines by stakes, by function and by tool; each figure is a count of your own lines, never a verdict on a tool.

When does it say a tool is deciding alone?

Where a line is marked solely automated or decides alone and the rule says it cannot, the register shows it as a line where a tool is deciding alone. In the specimen, 7 of 32 lines decide alone.

Which regimes can it attach?

The EU AI Act deployer duties, the GDPR's Article 22 with the information duties and the DPIA beside it, ISO/IEC 42001's responsible-use, supplier, operation and impact controls, ISO/IEC 42005's assessment items, the NIST AI RMF's measurement and oversight items, and APRA CPS 230's service provider and controls-testing paragraphs. A first line such as `eu ai act: yes | entity: lender | reporting cycle: quarterly | as at: 2026-09-20` ticks the regime and pins the cycle.

What extra columns does it read?

It also reads retest, benchmark, logs, impact assessment, FRIA and notice columns if you have them. Tool names in the paste are kept as labels and the register never rates a tool.

Does it rate the tools I use?

No. The register never rates a tool and never rules on a task. It classifies your own lines and states the obligation in our words, cited to its clause.

What do I take to the vendor?

The vendor benchmark letter exports as one printable page, one paragraph per task, with the benchmark to demand before the next reporting cycle. The register also lists the lines with nobody named to review.

How many findings does it raise?

The full register adds ten findings, and it names the lines where a tool is deciding alone under a rule that says it cannot. It also shows the lines by stakes, by function and by tool.

Where does the obligation text come from?

Each obligation is read clause by clause from the published text of the EU AI Act (Regulation (EU) 2024/1689), the GDPR (Regulation (EU) 2016/679), ISO/IEC 42001:2023, ISO/IEC 42005:2025, ISO/IEC 23894:2023, the NIST AI RMF 1.0 and APRA CPS 230, stated in our words and cited to its clause, so you can check it against the official text.

How Decision Register compares with a spreadsheet, a governance platform or a general assistant

A spreadsheet holds the use-case list and counts the lines, but it does not say which lines are the EU AI Act's Annex III uses or the GDPR's solely automated decisions. An AI governance platform is built for a firm with a model risk function and a licence budget, and a governance consultant or a lawyer can advise on a specific deployment, which this register does not do; it never rates a tool and never rules on a task. A general AI assistant will explain Article 22 in plain language, but it will not minute each pasted task with the stakes, the oversight, the record it must hold and the benchmark to demand from the vendor. Decision Register classifies your own lines against a published dictionary of 86 task types and attaches the obligations of the EU AI Act, the GDPR, ISO/IEC 42001, ISO/IEC 42005, the NIST AI RMF and APRA CPS 230 with their clauses, names the lines where a tool is deciding alone under a rule that says it cannot, and exports a vendor benchmark letter.

Primary source: EU AI Act (Regulation (EU) 2024/1689)