ReviewDesk guide · By the Ingenious AI team · Published 14 November 2026 · Reviewed 14 November 2026
How do you let someone correct wrong data after an adverse automated decision?
Generate a free adverse-outcome letter — no sign-up Build a free pre-use notice See pricing
The short answer. Say what can be corrected, accept proof through one clear route, pause the argument while you check, re-assess on the corrected data where that is feasible, and record the whole sequence against the original decision. Most corrections are small — a namesake's record, a stale balance, a transposed date — and a fast, gracious correction route converts your most justified complainants into people whose problem you solved. ReviewDesk logs corrections against the same docket as the decision and its review.
What is data correction after an adverse decision?
Data correction is the affected person's remedy for factual error in the information an automated decision consumed: Colorado's framework contemplates correcting incorrect personal data where feasible, alongside the explanation and human-review rights. It differs from review — correction repairs the input and can trigger a fresh assessment; review re-examines the decision as it was made. A complete adverse-outcome letter offers both routes and explains each in a sentence.
Why is correction the highest-value route you operate?
Because its hit rate dwarfs every other remedy's. Sample your overturned reviews and a pattern emerges: the majority turn on facts, not judgement — records merged across namesakes, payments posted after the data snapshot, qualifications the parser failed to read, availability recorded from an outdated form. Each was a correct application of the criteria to incorrect information. A correction route catches those cases at their cheapest point, before positions harden, and it improves the dataset for every future decision that would have consumed the same wrong record.
It also changes the emotional register of the entire exchange. A person told 'here is how to fix your record' is being treated as a participant in accuracy; a person told only 'you may appeal' is being treated as an opponent. The letter that leads with correction — as in the worked pattern in the 30-day explanation guide — draws a measurably different class of response: documents supplied, specifics given, problems solvable.
What should the correction route accept?
Anything that bears on the factors the decision actually used — which is why the explanation letter must name those factors specifically enough for the person to aim. Proof can be documents (statements, certificates, records), corrected values with a source, or a demonstration that a record belongs to someone else. Set the evidential bar at plausibility-plus-verification: accept the submission, verify it against an independent source where one exists, and never demand a standard of proof higher than the standard the original data met when the tool consumed it uncritically.
Publish what happens next, in the letter itself: where the correction goes, who checks it, the timeframe you work to, and the consequence — a fresh assessment on corrected data where feasible, with its outcome communicated and recorded. An unpublished process is experienced as a black hole, and black holes generate duplicate submissions and escalations. The inbox pattern in the inbox guide treats corrections as first-class docket items for exactly this reason.
How does re-assessment work once data is corrected?
The corrected record replaces the wrong input, and the assessment — automated or human — runs again on the repaired facts. Two disciplines matter. First, scope: the re-assessment uses the same criteria as the original; correction is not a side door to a more sympathetic standard. Second, linkage: the new outcome is a new record referencing the original decision and its correction, appended to the chain rather than overwritten — the append-only rule from the hash-chain guide. The person receives the new outcome with a short note identifying what changed and why.
Where the corrected data does not change the outcome, say so plainly and show the working: the corrected value, the criterion it was tested against, the result. A correction process that only ever announces unchanged outcomes without working will be assumed rigged; one that shows its arithmetic earns belief even when the answer stays no. The remaining route — human review of the decision itself — stays open and the letter's final paragraph should say so.
How do corrections interact with the human-review right?
They are parallel doors into the same room, and people will use both, sometimes in one message. Handle them as one case: the correction checked first because it is factual and fast, the review then conducted on the corrected record if the person still wants the decision re-examined. Never bounce a person between routes with form letters — the docket holds both threads, and one officer owns the case to its conclusion. ReviewDesk's inbox supports exactly this shape: request types differ, the case is one.
Record which remedy resolved the case. Over a quarter, the ratio of correction-resolved to review-resolved cases is diagnostic gold: a high correction share indicts your data pipeline; a high review share indicts your criteria or their communication. The annual risk assessment wants that ratio with its interpretation attached — see the risk-assessment guide. Programmes that cannot produce it are guessing about their own failure modes.
Where does wrong data usually come from?
Four sources recur. Third-party data suppliers merging identities or selling stale snapshots. Parsers extracting values from documents and missing endorsements, suffixes and second pages. Applicants' own older submissions persisting after circumstances change. And internal joins — the customer record matched to the wrong account by a fuzzy key nobody audited. Naming the source matters because each has a different fix: supplier challenge, parser rule, refresh prompt, join audit. A correction log categorised by source becomes a data-quality programme without anyone having to found one.
Feed corrections back to the source, not just the case. If a supplier's file produced nine namesake errors this quarter, that is a contract conversation with exhibits attached. If your own parser missed the same endorsement format eleven times, that is a rule fix that prevents the twelfth. TheInbox that keeps corrections as docket rows with source tags makes both conversations evidence-based rather than anecdotal — one more return on the audit-trail investment described in the records guide.
What should the person experience, end to end?
A route they can find from the letter in one step. An acknowledgement with a reference and a timeframe. A check conducted at the speed the original data was trusted — fast. And an outcome that shows its working either way. Total elapsed time for a documented namesake error should be days, not weeks, because the verification is usually a single lookup. Where verification genuinely takes longer — third-party records, manual file retrieval — the status link keeps the person informed rather than wondering, the pattern in the status-tracking guide.
Close the case in writing, from the record: what was corrected, what was re-run, what resulted, what remains open. That closing note, filed on the docket, is the artefact the person keeps — and the one they will quote if they ever describe your process to anyone else. Correction done graciously is the single most quotable thing an adverse-decision process can produce.
Your five actions for data correction after an adverse decision
- Name the decision's actual factors in the letter specifically enough for the person to aim a correction at them.
- Accept corrections through one route, acknowledge with a reference and a timeframe, and log them on the original docket.
- Verify at the standard the original data was trusted — fast — then re-assess on the same criteria and append the new record.
- Show the working when the outcome does not change, and keep the human-review door explicitly open.
- Tag every correction by source and feed the pattern back to the supplier, parser or join that produced the error.
Worked example: a namesake error, corrected in three days
A fictional tenant-screening case: an applicant — Daniel Reyes — is declined principally on an eviction filing surfaced by the screening file. His correction submission is one page: the filing's case record, showing a different middle name and a birth year nine years apart — a namesake in the same city. The officer verifies against the court record the same day, repairs the file, and the re-assessment runs on the corrected data: the remaining factors — income at 3.4 times rent, two years of on-time payments — clear the criteria. Approval issued on day three, appended to the original decision's record with the correction referenced.
The supplier receives the case as a data-quality exhibit, and the screening workflow gains a merged-record check for same-name files. One person's three-day correction improved the pipeline for every Daniel Reyes who follows — the compounding return that makes the correction route the best-run part of any adverse-decision programme.
Frequently asked questions
What data can a person correct?
The personal data the decision consumed — especially the factors named in the explanation letter. Corrections can be documents, corrected values with a source, or proof that a record belongs to someone else, such as a namesake with different identifiers.
How is correction different from human review?
Correction repairs the input and can lead to a fresh assessment on accurate data. Human review re-examines the decision as made. The routes are parallel; ReviewDesk holds both on one docket so a person can use either or both without being bounced between processes.
What if the correction does not change the outcome?
Say so, and show the working — the corrected value tested against the same criterion, with the result. Then confirm the human-review route remains open. Shown arithmetic earns belief; a bare 'no change' does not.
How fast should corrections be handled?
At roughly the speed the original data was trusted — days for documented errors like namesake merges. Acknowledge immediately with a reference and timeframe, and keep the person informed through longer verifications.
Is this guide legal advice?
No. ReviewDesk is templates and workflow tooling, not legal advice. The correction path is described as the regimes frame it in our research; consult counsel for your situation.
Keep reading: How do you write principal reasons for an adverse decision that a person can actually use? · What must you tell job applicants when a hiring algorithm screens them? · How do ADMT notice and review duties apply to automated credit decisions? · What must landlords and screening services disclose when screening is automated? · How do notice and human-review duties apply to automated insurance underwriting and claims? · 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.