See what a map looks like in your sector
Every AI-enabled decision is a chain of stages. Here is how AI-FPM shows where a small failure could travel, and where a control should interrupt it, in eight illustrative scenarios.
Credit decisioning
Bureau data pull
The credit agency's record on each applicant is fetched.
AIncome data pull
Income comes from a third-party provider. The feed goes 30 days stale. Values still look valid, so nothing alerts.
BRecord matching
Bureau and income records are linked to one person. A name or address variation links the wrong person's income. Every application still matches.
CRatio building
Raw figures become model inputs such as debt to income. Built on stale or wrong income, the ratio comes out understated.
Scoring model refresh
The model is retrained on recent lending outcomes.
Application scoring
Each applicant gets a score from the inputs it is given.
Score banding
Scores become the bands underwriters see.
DDecision cut-offs
Rules turn a band into approve, refer or decline. After a lending policy change they are not revisited.
- ACAmplifiesThe stale income makes the ratio wrong on every affected application.
- BCAmplifiesThe wrong person's income adds to the same error.
- CDMasksThe score looks comfortably safe, so nobody questions the old cut-off.
A loan is approved that should have been referred, and every step passed its own check.
A data-freshness monitor on the income feed, with a circuit breaker that pauses automatic approvals and routes to manual review when freshness is breached.
Claims triage
AClaim intake
The claim form, photos and documents arrive. Incomplete or oddly formatted documents are accepted anyway.
BDocument extraction
Amount, date and cause are read into fields. A supplier invoice layout changes and the amount is misread without an error.
Fraud model refresh
The fraud model is retrained on past claims and investigation outcomes.
Fraud scoring
Each claim gets a fraud score from the extracted fields.
Score and reasons
The score and its main reasons are shown to the claims handler.
CTriage routing
Thresholds send a claim to fast payout, handler review or investigation. They are tuned to last year's claim mix.
- ABAmplifiesIncomplete documents make a misread more likely.
- BCMasksA misread amount makes a risky claim look routine, so nobody sees the thresholds are out of date.
A claim that should be investigated is paid fast, and every step passed its own check.
A monitor that compares extracted amounts with the source documents, with a circuit breaker that routes mismatches to a handler instead of fast payout.
Referral prioritisation
AReferral intake
GP referrals arrive as letters and forms. One practice's format is read as blank fields.
BSymptom coding
Free text becomes standard clinical codes. Uncommon phrasing is coded as a milder condition.
Risk model refresh
The model is retrained on past referrals and outcomes.
Urgency scoring
Each referral gets an urgency score from its codes.
Worklist priority
Scores are shown as priority bands to the booking team.
CClinician review
Clinicians review the referrals flagged high priority. Under-scored cases never reach them.
- ABAmplifiesBlank fields leave less text to code, so a milder code is more likely.
- BCMasksA milder code lowers the score, so the referral never reaches a clinician.
An urgent patient waits, and every step passed its own check.
A monitor on the share of referrals with blank fields, with a circuit breaker that sends them to manual triage.
Dynamic pricing
ACompetitor price feed
Competitor prices arrive through the day. One feed freezes and keeps showing the last price.
Stock feed
Stock levels are loaded from the warehouse.
BDemand signal preparation
Sales, promotions and seasonality are combined. An unflagged promotion is read as real demand growth.
Demand forecast refresh
The forecast is retrained weekly.
Price optimiser
The optimiser recommends a price for each product.
Price publishing
Prices are pushed to the website and stores.
CMargin review
The team reads margin and sales to judge pricing. A fall in margin is put down to the market.
- ACAmplifiesA frozen competitor price pushes prices lower than the market needs.
- BCMasksThe false demand growth explains the sales, so the margin fall looks like the market.
Margin erodes for weeks before anyone sees that the pricing is the cause, and every step passed its own check.
A monitor on how recently each competitor feed updated, with a circuit breaker that holds prices when a feed is stale.
Benefits eligibility triage
AApplication intake
Claimants submit details and evidence online. Blank optional fields are treated as no change.
BRecords matching
The application is matched to existing departmental records. Two people with the same name and date of birth are linked.
Eligibility model refresh
The model is retrained on past decisions.
Eligibility scoring
Each case gets a score.
Case list
Cases are ordered by score for caseworkers.
CCase prioritisation
Caseworkers work down the list. Low-scoring cases are rarely opened.
- ABAmplifiesMissing details make a wrong match more likely.
- BCMasksThe wrong record lowers the score, so the case sits where nobody looks.
A claim is refused on someone else's record and nobody sees it, and every step passed its own check.
A monitor on match confidence, with a circuit breaker that sends uncertain matches to a caseworker.
CV screening
AApplication intake
CVs and forms arrive from job boards and email. A job board changes its export format and fields shift.
BCV parsing
Roles, dates and skills are read into fields. Non-standard layouts parse with gaps, so a career break reads as missing experience.
Ranking model refresh
The model is retrained on past hires.
Candidate ranking
Each candidate gets a position.
Shortlist
The top candidates are shown to the hiring manager.
CHiring decision
Managers choose from the shortlist. It is a fixed size, so candidates just below the line are never seen.
- ABAmplifiesShifted fields give the parser less to read.
- BCMasksParsing gaps push capable candidates down, and the fixed shortlist hides that they were cut.
A capable candidate is never seen, and every step passed its own check.
A monitor on the share of CVs with empty parsed fields, with a circuit breaker that routes them for manual review.
Event safety briefing
Research request
Analysts ask an AI assistant to research a group of visiting supporters ahead of an event.
Source gathering
Public reports and news items are collected for the assistant to use.
Assistant updates
The provider updates the assistant from time to time.
ASummary writing
The assistant summarises the group's history. It includes incidents that never happened, cited to sources that look real.
BBriefing note
The summary goes into a formal note with its citations. The citations are not opened, so the note reads as sourced.
CSafety decision
A panel decides whether the group should attend. Invented incidents back a view the panel already held.
- ABMasksReal-looking citations hide that the incidents were invented.
- BCLegitimisesThe sourced-looking note gives a view the panel already held the look of evidence.
A group is turned away on evidence that never existed, and the note looked sourced all the way through.
A rule that every citation in a briefing is opened and confirmed by a person, with a circuit breaker that stops a note leaving the team while any source is unconfirmed.
Complaint handling agent
Complaint intake
Complaints arrive by email, chat and call transcripts.
ACase file assembly
The agent gathers account history and earlier contacts. It pulls in an unrelated old dispute and treats it as part of this case.
Agent instructions
The agent's goals, tools and rules are set and refreshed.
BAgent actions
The agent checks records, drafts replies and books follow-ups. It keeps working the old dispute and loses sight of the customer's actual complaint.
Case status
The team dashboard shows each case as in progress, with the agent's latest activity.
COverdue review
Team leads review cases past their deadline. Cases showing recent activity are not escalated.
- ABDelaysThe tangent keeps the agent busy on the wrong problem, so the real fix slips later and later.
- BCMasksConstant agent activity makes the case look healthy, so it is never escalated.
A customer waits weeks for an answer while the case shows as in progress, and every step passed its own check.
A monitor that compares what the agent is working on with the original complaint, with a circuit breaker that hands the case to a person when it drifts.
What Every Example Shares
A chain, not a component
Each risk is traced across stages, so you see how an upstream issue becomes a downstream outcome.
A control with an owner
Every failure mode points to the monitor, control or breaker that interrupts it.
A view for the board
The same map rolls up into a plain-language picture of risk, decisions and actions.
Map your own AI-enabled decision
Start with one pipeline that matters and see where its risk really sits.