ReviewDesk guide · By the Ingenious AI team · Published 14 November 2026 · Reviewed 14 November 2026
How do you write an ADMT notice in plain language that people actually understand?
Generate a free adverse-outcome letter — no sign-up Build a free pre-use notice See pricing
The short answer. Write for the person whose life the decision touches, in the words they would use to describe it back. Name the decision before the technology, factors as facts about the world rather than model vocabulary, and rights as actions the reader can take today. Test every draft by reading it aloud to someone outside the industry; if they cannot repeat the gist, the notice is a legal artefact, not a communication. ReviewDesk's notices and letters are generated in this register by design.
What is plain-language ADMT writing?
Plain-language ADMT writing is the drafting discipline for notices, letters and review communications under the automated-decision regimes: short sentences, concrete nouns, the person's own values and records named, rights expressed as doable actions, and legal precision carried by structure rather than by jargon. It applies to everything the person reads — the pre-use notice, the adverse-outcome letter, confirmations and outcome notes.
Who is the reader, really?
A person in a moment of consequence, often on a phone, often anxious, frequently reading in a second language or under time pressure, and almost never a lawyer. They are not asking 'is this disclosure compliant?'; they are asking 'what is happening to me, and what can I do?'. Every drafting decision flows from serving those two questions in that order. The compliance content must all be present — plain language is not the removal of substance — but it is ordered by the reader's need, with precision carried in specifics (factors, dates, values, routes) rather than in terminology.
Hold one concrete reader in mind per document: the warehouse applicant, the tenant, the sole trader. Generic audiences produce generic prose; a named reader produces sentences with a pulse. ReviewDesk's generator structure assumes this — the letter is assembled around one applicant, one decision, their reasons — and the tone discipline in the letter-tone guide is this section's sibling.
What word choices do the work?
World-words over model-words. 'Rent payments in the last two years' beats 'tenancy obligation performance'. 'Jobs held and how long' beats 'employment stability indicators'. The test is repeatability: could the reader say the factor back to a friend accurately? If not, the word is a label for your convenience, and labels are where understanding goes to die. Where a technical term is unavoidable — a statutory phrase, a record type — use it once, defined in the same sentence by its world-content, then continue in world-words.
Verbs deserve the same discipline. Decisions are made by actors: 'we declined', 'the tool scored', 'a reviewer will check'. The passive fog — 'it was determined that criteria were not met' — exists to hide actors, and readers smell it instantly. An organisation willing to name its actors sounds like one willing to answer for them, which is the entire communicative goal of the regime.
How should rights be expressed?
As actions with handles. Not 'you have the right to request human review' but 'to ask a person to look again, use the form on this page or write to…' followed by what happens and by when. The legal right is the foundation; the action is the storefront. Readers act on handles: a link, an address, a timeframe, a named next step. Every right in a notice or letter should be checkable by the question: could the reader complete this action today, from this document alone, without phoning anyone? If not, the drafting is unfinished, however accurate the law in it.
The same discipline applies to the correction route — 'if any of this information is wrong, send us…' with the evidence examples from the correction guide — and to the outcomes of actions: 'we will write back within…' with a real number. Promises with numbers can be kept and audited; promises with adverbs can only be felt, and feelings are where trust goes to be litigated.
How do structure and length serve comprehension?
Short paragraphs, one idea each; headings that state the reader's question ('What did the tool look at?', 'What can I do now?'); the decision and the next action never buried. Length follows content, but every sentence must pay rent: a sentence that neither informs the reader's understanding nor enables their action is decoration, and decoration in an adverse letter is a small cruelty — it makes a person in distress read your brochure. The notice structure in the hosted-notice guide — decision, tool, factors, rights, form — is this discipline at page scale.
Numbers need anchoring. 'Income below the required level' is fog; 'verified monthly income of $2,850; the rent is $1,150 and the requirement is three times the rent' is a fact the reader can verify against their own life. Anchored numbers are the plain-language writer's highest-yield habit, and the principal-reasons standard in the reasons guide is built on them.
What are the failure phrases to hunt and kill?
Keep a list and search drafts for them. 'We take your privacy seriously' — content-free warmth. 'Due to a number of factors' — the white flag of reason-writing. 'Unsuccessful on this occasion' — occasion language for permanent-feeling outcomes. 'Our decision is final' — usually false inside a regime with review rights, and always a provocation. 'Please do not hesitate to contact us' — hesitation is exactly what the phrase produces. Each has a plain replacement that carries information instead of atmosphere.
Legal safety phrases deserve the same hunt, handled more carefully: disclaimers belong in the document but not as substitutes for content. 'This letter is not legal advice' and the templates-not-legal-advice line are honest framings — ReviewDesk carries its own — but a letter that is ninety per cent disclaimer and ten per cent decision has inverted its purpose. Order the document by reader need; let the framings frame, not fill.
How do you test a draft before it ships?
Read it aloud — the ear catches what the eye forgives: sentences that need a second breath, nouns stacked three deep, the moment the register slips into defence. Then give it to a reader outside the industry with two questions: what is this telling you, and what can you do now? Their answers are the usability data. Finally, test the specific recipient simulation: read your letter as the person whose facts are in it, especially the reasons — accuracy to their life is the deepest plainness there is, and a beautifully simple letter about the wrong person's data is the worst artefact in this entire field.
Build the tests into generation. Because ReviewDesk letters are generated from the workflow record with a fixed structure, the register is set once, at the template level, and every letter inherits it — the per-letter human work concentrates on the reasons, where plainness matters most. Set the template's voice with these tests, and a thousand letters stay readable while a thousand hand-drafted ones would drift into fog by March.
Your five actions for plain-language ADMT writing
- Write to one named reader at a moment of consequence; order every document by their two questions — what is happening, what can I do.
- Use world-words for factors and name the actors in every verb; define unavoidable technical terms in their own sentence.
- Express every right as an action completable today from the document alone, with a number for every timeframe.
- Anchor every quantity: the person's value, the threshold, the gap — never an unanchored level or band.
- Kill the failure phrases, read drafts aloud, and test with an outside reader until they can repeat the gist.
Worked example: the same notice, before and after
Before, from a fictional employer's draft: 'In accordance with applicable requirements, applicants are advised that automated processing technologies may be utilised in the evaluation of candidacy, and that determinations rendered may be subject to reconsideration procedures upon formal request.' Seventy-one words; the reader learns that something legal exists. After: 'We use a screening tool to help decide who to interview. It looks at your eligibility for the role, your availability for the shift patterns, and a short recorded assessment. If you are not interviewed, we will write to you with the main reasons. If anything in your information is wrong, or you want a person to look again, use the form on this page — a person will review it.'
The after version is not shorter because content was cut; it is shorter because fog was. Every legal element of the before version survives — notice, factors, explanation, review — translated into the reader's language and given handles. That translation is the whole craft, and it is learnable in an afternoon by any team willing to read its drafts aloud and flinch honestly.
Frequently asked questions
What is plain language in an ADMT notice?
Drafting that a non-specialist reader can act on: world-words for factors, named actors, anchored numbers, rights expressed as actions with routes and timeframes. The legal content stays complete; precision moves from terminology into specifics.
Does plain language weaken legal accuracy?
No — it relocates accuracy. A notice that names the actual factors, dates and routes is more accurate, not less, than one hiding behind general terms. Structure carries the precision; jargon usually carries only the appearance of it.
How do we test a draft?
Read it aloud, then give it to a reader outside the industry and ask two questions: what is this telling you, and what can you do now? If they cannot answer both, redraft. Finally, read it as the specific person whose facts it contains.
Should rights be stated as rights or as actions?
As actions built on the rights: 'to ask a person to look again, use the form on this page' outperforms the bare right, because the reader can complete it today. State the timeframe and what happens next with it.
Is this guide legal advice?
No. ReviewDesk is templates and workflow tooling, not legal advice. The writing discipline is ReviewDesk's house standard; consult counsel on your documents' legal content.
Keep reading: How do you write an adverse-outcome letter that de-escalates instead of inflaming? · What does review coverage mean — and what breaks when requests go unanswered? · How do you answer every human-review request — without sounding like a bot? · How do you run ADMT workflows across Colorado, California and Illinois at once? · How can an agency run ADMT desks for multiple client businesses? · All ReviewDesk guides · ReviewDesk home · Try the free desk · Free letter generator · Free notice builder · Pricing
ReviewDesk is built by Ingenious AI. Templates and workflow tools — not legal advice.