Guide · Published November 14, 2026 · Reviewed November 14, 2026 · Ingenious AI team · Not legal advice
How to Write an Adverse-Outcome Explanation: Reasons, Data, and Next Steps
The short answer: An adverse-outcome explanation that works has the structure the kit's builder produces: what happened (the decision and the tool's role in it), the main reasons for the decision stated specifically, the data the tool used summarised honestly, and what the person can do next — correct their data and ask for human review, with named contacts for both. In Colorado it is due within 30 days of the decision; in California the equivalent content answers an access request (logic, outcome, purpose); in Connecticut it answers an outcome question. The drafting discipline is the same in all three.
Check your duties and dates — free Build your first notice — free
What is an adverse-outcome explanation? It is the plain-language statement sent after an automated or automated-assisted decision goes against a person. 'Adverse' covers rejections, denials, and materially worse terms the tool informed. Worked examples of the finished document, labelled as examples, are in Adverse-Outcome Explanations in Plain Language; this guide is about how to write one.
Part one: what happened, with the tool's role stated plainly
Open with the decision itself: who made it, what it was, when, and what part the tool played. The kit's builder phrases it as: the business 'made a [decision-type] decision that went against you, with help from [tool]'. That sentence does two jobs — it owns the decision as the business's (not the tool's), and it discloses the tool's involvement without exaggerating or minimising it. Match the weight you describe to the weight the tool actually carried.
Dates are content, not administration: the explanation carries both the decision date and its own date, because in Colorado the gap between them is the compliance fact (30 days). Print both on the document, always. A reader — or a reviewer a year later — should be able to verify the clock from the document itself.
Part two: principal reasons, specific enough to act on
The reasons section is where explanations succeed or fail. The builder takes the principal reasons as entered for that decision — which means the quality is decided at decision time, when the reasons are captured. A reason is usable if the person could, in principle, check it or correct it: 'income from the verified source was below the threshold for this loan size' can be checked; 'risk profile insufficient' cannot.
Order reasons by weight, and keep the list to the principal ones — the regulation's word is principal, and a dump of every minor factor obscures the ones that mattered. If a reason cannot be stated without revealing another person's data or genuinely confidential method detail, state its effect instead ('a record attributed to your file from 2022') — specific about the person's facts, careful about everything else. Reason quality is also what makes the correction path real: people correct reasons, not verdicts.
Worked example: Arroyo Seco Credit Services
Example — fictional business, real duty structure. Arroyo Seco is a fictional consumer lender in Colorado that rebuilt its explanation workflow around the three working parts. Its first month shows the difference:
- Decisions and clocks: 230 decisions, 74 adverse — every adverse file has reasons and a data summary captured at decision time, so no explanation waits on reconstruction.
- Explanations sent: all 74 within 30 days, median day 5, each carrying both dates on its face so the clock is verifiable from the document.
- Reason quality pass: a weekly review samples 10 explanations; in week one, 3 reasons are sent back as too vague ('insufficient profile') and rewritten to the underlying facts before sending.
- Reader actions: 12 recipients use the correction contact and 7 request human review — both contacts printed on the explanation; 4 corrections lead to re-run decisions, 2 with changed outcomes.
- Record kept: 74 explanations with their decision dates, reasons, and data summaries in one hash-chained record, held for three years.
Part three: the data summary, honest about sources
The builder includes a data summary: what data about the person the tool used. Write it at the level of categories and sources — application data provided by the person, records from a named source, history held by the business — and be complete about the categories that carried weight. A summary that lists the application form but omits the purchased or bureau data that actually drove the outcome is the explanation equivalent of a wrong reason.
This section is also the bridge to correction. The person reads the summary, spots the wrong item, and knows what to ask to correct. Connecticut's housing right — correct and have the decision reevaluated — only functions if the summary lets the person find the error. Draft the summary so that a wrong fact would be visible to the person it describes.
Part four: next steps, as actions with contacts
Close with the two actions and their contacts: correct your data (with the correction contact) and ask for human review by a person with authority to change the outcome (with the review contact). The builder prints both contacts from the business profile onto every explanation. Phrase them as invitations with addresses, not as legal notices of rights in the abstract — the reader should be able to act within a minute of finishing the document.
Say what happens next for each action, briefly: corrections are acted on and the decision looked at again; review is by a person who can change the outcome. Do not promise outcomes — promise the process, which is what the duties actually offer. And keep the kit's footer intact: the explanation is a templates-and-records document, not legal advice, and it does not judge whether the decision was fair. That boundary, stated on the document, is part of its credibility.
Drafting workflow: capture at decision time, send on a cadence
Explanations fail operationally more often than rhetorically: reasons not captured, drafting batched monthly, approval stuck with one person. The workable pattern mirrors the clock guide: reasons and data summary entered when the decision is made; drafts generated from them per person; a named sender and a deputy; sent dates logged against decision dates. The kit automates the generation and the logging precisely because the drafting inputs decay within days.
Before the workflow goes live, write three explanations by hand for three real (internal, anonymised) past decisions and read them cold. If a colleague outside the team cannot say what happened, why, and what to do next after one read, the template needs work — not the reader. Then generate the production version in the free generator and compare structures. Colorado's timing rules are in Colorado's 30-Day Explanation Clock.
- State what happened first: the decision, its date, and the tool's role at its true weight
- Write principal reasons the person could check or correct — ordered by weight, no category labels
- Summarise the data honestly by category and source, completely for the categories that carried weight
- Print both next steps as actions — correction and human review — each with its named contact
- Capture reasons and data summaries at decision time, log sent dates, and sample-read explanations weekly for vagueness
Sources, as carried in the kit's own data: the duty summaries in app/rules.py and the checker data in app/checker_data.py, verified against primary or authoritative sources on October 28, 2026. Run the free duties checker for your jurisdiction, decision type, and the tool's role, and compare the kit's plans for the record, builder, and review queue. More from Ingenious AI is on the hub.
Related guides
- Adverse-Outcome Explanations in Plain Language
- Colorado's 30-Day Explanation Clock: How It Runs and How Not to Miss It
- Colorado AI Act Automated Decision Notices: What to Send
- ADMT Deadline Tracker: Which State Duties Land When
See all ADMT guides for the full set.
Frequently asked questions
What must an adverse-outcome explanation contain?
What happened (including the tool's role), the principal reasons for the decision, a summary of the data the tool used, and what the person can do next — correct their data and ask for human review — with named contacts for both.
How specific must the reasons be?
Specific enough that the person can check or correct them. 'A rent-to-income ratio below the stated threshold' is a reason; 'insufficient profile' is a label.
When is the explanation due?
In Colorado, within 30 days of the decision. California's access answer (logic, outcome, purpose) is given on request with no fixed day count in the kit's data; Connecticut outcome questions follow its profiling rules.
Who writes the explanation — the business or the tool vendor?
The business. The duty sits with the deployer; the kit generates the document from the business's own profile and the facts it enters for that person.
Should the explanation say whether the decision was fair?
No. The kit's documents state that the kit records what was said and when and does not judge whether a decision was fair. The explanation states reasons, data, and next steps.