A chief risk officer evaluates credit-risk policy against capital requirements at two levels: the
macro view, where the accumulation of every customer's income and commitments shows how much of the book
would tip into stress, and the micro view, a single customer's income and expenditure behaviours. Sheaf sits
over the AI making these decisions, so a recommendation that contradicts a customer's affordability is held before
it reaches them.
Illustrative view for Pollen Street Capital · representative lending book, sample data · stress bands and behaviours per the Castlight affordability model
2,204
Customers in the book
3.5%
In the red (Stressed) at current policy
£1.10m
Capital at risk (expected credit loss)
9
Unsafe AI credit actions held (30d)
88/100
Decision coherence health
Macro · stress-test the book · every customer's income & commitments, accumulated
Apply an increase to all expenditure — loans, mortgages, groceries+0%The CRO's lever: simulate a rate rise or cost-of-living shock, and watch customers migrate toward the red as their headroom disappears. Click the Stressed band to drop into a customer.
Customers in the red
77
3.5% of the book · Stressed band
Capital at risk
£1.10m
expected credit loss across the book
Policy headroom
Within appetite
capital requirement covered at this shock
Micro · a customer in the red · the income & expenditure behind the score
Customer ref M‑4471
"Running to stand still." Income is steady on paper but increasingly topped up by borrowing; almost every pound is pre-allocated before the month begins.
● Stressed · over-committed
726
Bureau score "looks fine"
41
Affordability score
Where the £2,180 monthly income goes
Rent£950
Existing loans & credit£430
Utilities, council tax, phone£280
Groceries & transport£380
Gambling£150
Committed + essential£2,150
Left over to absorb any shock£30
Behaviour flags · what the bureau score misses
!
Replacing income with loans. £600 of last month's "income" was a short-term loan drawdown, not earnings.
!
Gambling = 22% of discretionary spend (£150/mo), trending up over the last quarter.
!
3 bank charges in the last 90 days; overdraft used every month before payday.
!
One debt-collection arrangement active, and rising cash withdrawals, a classic early-distress signal.
90-day cashflow "heartbeat" · balance dips below zero near every month-end
90 days agotoday
Where Sheaf sits · the AI decision, gated
Held action caught before it reached the customer
An AI credit agent recommended a £6,000 debt-consolidation top-up for M‑4471.
Smooth and confident, and built on a contradiction no single step caught: two agents in the chain used two
different incomes.
Affordability agent
Income £2,780 · approve, comfortably affordable
these never compared notes
Open-banking read
True income £2,180 · £600 was itself a loan
Sheaf measured across the agents, coherence dropped below threshold at that exact seam, and the
The Gate parked the approval and routed it to a human specialist under Consumer Duty
(FG21/1). The customer was never sent an offer that would have deepened the stress. Every held and released
decision carries the evidence of why, ready for the FCA, the board, and diligence at exit.
Why a CRO runs the book on Sheaf
Macro and micro, one view
Policy you can stress in seconds.
The whole book aggregated against capital requirements, and one click into any customer's real income and expenditure. Set appetite where the capital actually sits, not where the bureau score guesses.
Measure before you lend
Affordability the score can't see.
Income replaced by loans, gambling creep, bank charges, thinning headroom. The behaviours that put a customer in the red are surfaced before an AI acts on a number that looks fine.
Deploy AI, prove it's safe
Speed with an evidence trail.
Contradictions are caught and gated, and every decision carries proof. The lender ships AI across the book faster with the risk measured, and the same record answers the regulator.
Illustrative. The lending book, customer M‑4471, scores and held-action counts are representative sample data
to show the shape of the view, not a real customer. Stress bands (Secure → Stressed), affordability
scoring, and the behaviour flags follow the Castlight affordability model. Populated with a lender's own
open-banking and credit data, this becomes the live credit-risk dashboard. · sheaf.one · the coherence layer for agentic AI