CECL allowance management where every qualitative factor arrives with its support already written
Every qualitative adjustment arrives with its support already written.
See one case, screen by screen ↓Software that runs the quarterly allowance for credit losses close around your CECL model: the dated steps from loan-tape cut to committee, the qualitative factors and their written support, the change bridge, the committee pack and the evidence binder. CECL Allowance Management adds six agents that do the checking, drafting and assembling; the Allowance Manager decides.
How to choose AI solutions for lending →
Updated
Every quarter the allowance team has to defend the number to the committee, the external auditor and the examiners. The model does the arithmetic. What eats the close is everything around it: nine segments of trend data to pull, 81 qualitative factor cells to re-prove, a quarter-over-quarter bridge, a committee memo and an evidence binder.
So the analyst rolls last quarter’s memo forward. A February office survey survives into October while three new reappraisals say something else. A national-economy factor stays on after the model starts carrying national unemployment and GDP. A concentration factor is held flat while the ratio climbs. Those are the gaps examiners circle. In a Federal Reserve review of community bank exams, most CECL concerns were inadequate documentation of the allowance process or unsupported qualitative factors.
Estimated split. The totals are the ones the close screen shows: Q2 2026 took 19 working days by hand; the Q3 plan with the agents is 11.
Six agents test every factor cell and write its support; the Allowance Manager decides and the committee approves.
Ashgrove Plains Bank, $3,618.4M in loans, Q3 2026. It is Monday 12 October, day 8 of a 12-step close, and the allowance committee meets Thursday at 09:00. Here is what happens next, screen by screen, in the working solution.
Dana lands on the allowance walk: $44,812K at 30 June, then net charge-offs, volume and mix, the economic forecast, credit quality and individually evaluated loans, each a bar, ending at a draft of $46,843K. One bar is dashed. Jamal rolled all 81 June factor cells forward unchanged — $9,028K on September balances, none of it tested against this quarter’s data.
“What the draft cannot show: a June source that new data contradicts · a factor the model already counts · a factor held flat while its trend moved.”
She clicks Individually evaluated. The rail shows the arithmetic to the dollar: Prairie Gate Office Center, loan balance $8,421,800, as-is appraisal from Kessler Valuation Group on 9 September at $7,915,000, less 8% cost to sell, net collateral value $7,281,800 — a reserve of $1,140,000. The office reappraisals and all seven individually evaluated loans open from the same rail.
Dana presses one button. The data reconciler confirms the tape ties to the ledger. The forecast reader lists what model v3.2 already carries. The trend analyst refreshes nine segments. The Q-factor support writer checks every cell and drafts the support; the change bridge builder prices it. Six changes come back: one stale source, one factor counted twice, one against its trend, one new product and two local trends.
“Six agents test every cell and draft the reasons. They propose; only you change a factor.”
The dashed bar fills to +$666K and splits into its six changes. The largest is CRE office collateral values, 15 → 35 bps, +$572K, tagged stale support. Its drafted paragraph cites three office reappraisals 18%, 24% and 31% below their prior values and the local vacancy rate — each number with a numbered source.
“Office vacancy in the Lincoln–Omaha market reached 19.8% at 30 September, from 15.1% a year earlier.”
The office factor’s page shows the reappraisals — Prairie Gate Office Center −31%, Saltillo Road Offices −24%, Haymarket Exchange Building −18% — the written support, and four checks. Source current fails: the Q2 support is dated February 2026, two quarters old. Not already in the model passes: the model uses a national CRE price index only. Dana can keep June’s 15 bps, change the basis points, or accept 35.
The June memo kept an 8 bps economic factor on commercial and industrial loans for “softening national conditions”. Since model v3.2 the forecast already includes national unemployment and real GDP for that segment, and Nebraska unemployment is 2.9%, the same as in June. The proposal: 8 → 0 bps on $702.6M, −$562K.
“The policy statement says adjustments should not be made for information already considered in the estimate.”
Dana accepts the six changes. The bridge from 30 June to 30 September ties to the segment schedule with an unexplained difference of $0: allowance $47,509K, 1.31% of loans, a change of +$2,697K and a provision of $3,911K. Ratios sit beside it — 218% of nonaccrual loans, 9.8 years of net charge-offs covered, a qualitative share of 20.4%, within the 15–25% the committee set.
Sensitivity shows the allowance at the 80 / 20 / 0 scenario weights, the severe scenario alone at $61,434K, and what each proposed change does. Moving every qualitative factor together by 50% moves the allowance by $4,847K either way — the answer to the committee’s first question, ready before it is asked.
The committee and binder assembler drafts the memo to the allowance committee from the decided factors, the bridge and the sensitivity: the recommendation, what moved, the six changes and why the C&I factor comes out — every number cited. Pack contents are ready: 7 sections, 64 cited items. Dana submits it; on Thursday Tom Whitaker, Ellen Marquez, Rachel Brandt and Owen Sato approve it 4 votes to 0.
The assembler never marks anything approved — the committee’s vote is recorded by name and time.
Everything behind the Q3 allowance is indexed to the memo: 64 items from 31 sources, from the tape tie-out to the audit trail of every factor decision, and 81 of 81 factor cells with support dated this quarter. Hollis & Crane’s open requests are each answered by an item — support for the office collateral factor with item 3.2. The auditor sees it read-only, after the committee approves.
The same quarter, seen by the five people who carry it — what their close looked like, and what it looks like now.
Six agents test every factor cell and write its support; the Allowance Manager decides and the committee approves.
Ties the loan tape to the general ledger by segment and maps new product codes before anything is calculated.
Reads the forecast vintage and local indicators, and records which drivers the model already carries, so no qualitative factor repeats them.
Grades, past dues, nonaccrual, growth, concentrations and new products by segment, quarter over quarter — every figure dated and cited.
Checks each of the 81 factor cells for stale sources, double counting and direction, and drafts the written support with citations.
Builds the quarter-over-quarter bridge, the ratio analysis and the sensitivities to scenario weights and to all factors at ±25% and ±50%.
Writes the committee memo and minutes draft from the decisions, and indexes the evidence binder for the auditor.
Accepts each change; the committee approves. The agents propose; a named person decides.
Dana and her team ask in plain words — or tell it what to draft. Every answer cites its source.
Why does the office collateral factor go to 35 basis points?
Three office properties were reappraised this quarter at 18%, 24% and 31% below their prior values, and Lincoln–Omaha office vacancy reached 19.8%. June’s 15 basis points rested on a February survey that predates all three. The model’s CRE index is national, so this is not counted twice. At 35 bps on $286.0M the factor adds $572K.
Which factors could the auditor say are counted twice?
One: the 8 bps economic factor on commercial and industrial loans. Since model v3.2 the forecast carries national unemployment and real GDP for that segment, and the policy statement says not to adjust for information already in the estimate. The proposal removes it (−$562K). The other 80 cells were checked against the model’s driver list.
What moved the allowance from $44.8M to $47.5M?
From $44,812K at 30 June: net charge-offs −1,214, volume and mix +571, economic forecast +1,048, credit quality +486, Prairie Gate individually evaluated +1,140, qualitative factors +666 — $47,509K at 30 September, 1.31% of loans against 1.26% in June. The largest single item is the Prairie Gate Office Center reserve.
Draft the memo to the allowance committee
Drafted. It recommends an allowance of $47,509K, 1.31% of loans, and a provision of $3,911K. It lists the six factor changes with their reasons, explains why the C&I economic factor comes out, and gives the sensitivity at 70 / 30 / 0 weights. Every number in it opens its source. It is in the committee pack, not yet sent.
13 screens from the working solution, on its sample data. Pick one to see it large.
The rules every factor is checked against are settings. Changes are versioned, need the Controller’s approval and are on the audit trail.
| Setting | Default | Choose from |
|---|---|---|
| A source goes stale after | One quarter | One quarter · Two quarters · Six months |
| No factor may repeat a model driver | On | On or off · checked against the model’s drivers: national unemployment, real GDP, CRE price index |
| Direction check | On | On or off · a factor held or lowered while its own trend worsens needs a written reason |
| Cap on one factor | 50 bps | Set in steps of 5 bps |
| Cap on a segment | 120 bps | 100 · 120 · 150 bps |
| CRE concentration reference | 300% of capital | 250% · 300% · 350% of capital |
| Reasonable and supportable period | 4 quarters, 4-quarter reversion | 4 quarters, 4-quarter reversion · 8 quarters, 4-quarter reversion · 4 quarters, immediate reversion |
| Individual evaluation above | $500,000 | $250,000 · $500,000 · $1,000,000 |
Every figure is labelled: a target the solution is built to, an estimate, a typical published result, or a proven one.
“demo” = seen in the working solution, on its sample bank data · “target” = the design goal, measured in the live solution · “estimated” = our estimate · “typical” = published findings (Federal Reserve, CECL at Community Banks, Community Banking Connections, 2025 — most supervisory concerns were inadequate documentation of the allowance process or unsupported qualitative factors). The Interagency Policy Statement on Allowances for Credit Losses (2023) is named as context. People, the bank, the auditor and the properties named on this page are fictional — characters and sample data in the working solution.
Software that runs the quarterly allowance for credit losses close around your CECL model: the dated steps from loan-tape cut to committee, the qualitative factors and their written support, the change bridge, the committee pack and the evidence binder. CECL Allowance Management adds six agents that do the checking, drafting and assembling; the Allowance Manager decides.
No. It works around the model you already run. It reads the model’s results by segment and its driver list, and builds the qualitative factors, bridge, sensitivity and pack on top of them.
The Q-factor support writer checks every cell — nine segments by the nine interagency factors — against this quarter’s data, last quarter’s memo and the model’s inputs. It flags stale sources, factors that repeat a model driver and factors moving against their trend, proposes a change only where the data supports it, and drafts the support with a citation on every number.
It checks every factor cell for it. The economic forecast reader records which drivers the model already carries, and a setting forbids any factor from repeating one. In the working solution it removed an 8 bps C&I economic factor that repeated national unemployment and real GDP.
No. The agents propose; only the Allowance Manager accepts, changes or rejects a factor change. The committee approves the allowance and the Controller books the provision. Agents never post or change a journal entry, change a risk grade or mark anything approved.
A read-only evidence binder and the memo, from the day the pack is final: items indexed to the memo, the audit trail of every factor decision, and each open request answered by the item that supports it.
Yes. When a source goes stale, the double-count and direction checks, the cap per factor and per segment, the CRE concentration reference, the reasonable and supportable period and the individual evaluation threshold are all settings. Changes are versioned and need the Controller’s approval.
The Agentic Solution Engine builds and deploys it from your requirements — your allowance policy, model output, segment map and last quarter’s memo — and it goes live once every quality gate has passed. We will walk you through it on your own last close first.
We’ll run CECL Allowance Management on your last quarter’s model results and qualitative factor memo.