The CECL Allowance Is Mostly Aging Buckets. AI Can Run That Part.

Of the roughly 67,800 live accounting and finance postings in our database this week, 80 name CECL or "current expected credit loss" in the description, and 44 of those also mention AI, machine learning or automation somewhere. That's a tiny slice of the market, and the people it describes are paid like it matters: among the 26 controller, senior accountant and accounting manager postings that name CECL, 22 also ask for a CPA, and the 17 that show pay have a median midpoint of $140,000.

I read that as a skill gap with money attached, and the part of CECL that most companies dread is the part AI is best at. The judgment calls are a handful of decisions you make once a year, and the allowance calculation itself is arithmetic you re-run every quarter whether you like it or not.

Where does the judgment in a CECL allowance actually live?

CECL (ASC 326, from ASU 2016-13) asks you to book lifetime expected credit losses on day one instead of waiting for a loss to become probable, and private companies have been living with that since 2023. The decisions that matter cluster in four places, and all four belong to you.

You decide how to pool receivables, which means whether your enterprise customers and your small-business customers share a loss rate or get separate ones. You decide how much history to look back on and whether a bad 2020 or a freak 2022 belongs in it. You decide whether you have a reasonable and supportable forecast worth adjusting for, or whether current conditions are the best information you've got. And you decide what to carve out for individual assessment, like the customer who just missed two payments and got pissed off when you called about it.

A model can inform every one of those, and none of them is a decision it should make for you. They're the parts an auditor will push on, so they're the parts where your name goes on the memo.

How did ASU 2025-05 change the receivables side?

For short-term trade receivables it made the forecasting step a lot smaller. FASB's ASU 2025-05, as Grant Thornton summarizes it, adds a practical expedient that lets you assume conditions at the balance sheet date stay the same for the remaining life of current accounts receivable and current contract assets arising under ASC 606. It's available to all entities, public companies included, and it's effective for fiscal years beginning after December 15, 2025, so for a calendar-year company 2026 is the first year it applies. Companies that aren't public business entities get an extra election to consider collection activity after the balance sheet date.

Read that with some care, because it covers current receivables and current contract assets and nothing broader. Long-dated notes, non-current contract assets and anything with a real forecast horizon still need the full treatment, and the expedient is a policy election you'll want documented before fieldwork rather than during it. But for a company whose credit exposure is mostly 30 and 60 day invoices, a chunk of the allowance just became a computation: aging buckets, historical loss rates by bucket, apply, book.

Which parts of the CECL allowance should you hand to AI?

The computation, the roll-forward and the first draft of the memo. Start with the aging. Pull the receivables subledger at period end, bucket it by days past due and by the pools you've already defined, and let the AI check that the buckets foot to the GL control account. Foot-and-tie is tedious as hell, and it's the step where a mystery variance is most likely to show up and ruin your afternoon.

Next is the loss-rate history. Give it eight or twelve quarters of write-offs and recoveries matched back to the original invoice dates, and have it produce loss rates by bucket and pool, showing its work in a table you can audit. This is the step that eats a week when a person does it in Excel, and a good AI workflow turns it into an hour of review.

Then the allowance roll-forward: beginning balance, provision, write-offs, recoveries, ending balance, tied to the GL with every break flagged. That's the same pattern we walked through in rolling reconciliations and the AI-assisted close, just pointed at a reserve account instead of cash. If you'd rather start from the collections side, our accounts receivable AI playbook covers the upstream work that feeds your aging.

Last, the memo skeleton. Policy, pools, methodology, inputs, results, and a variance explanation against last quarter, drafted from the numbers you just produced so you edit rather than write from scratch.

Where does AI get CECL wrong?

Mostly in the places where the facts live outside the spreadsheet. A model doesn't know that your biggest customer's CFO resigned last week, that a disputed invoice is a pricing argument and has nothing to do with credit risk, or that the "recovery" in your data was a one-time settlement you'd never see again. It will also produce a confident loss rate from thin data, and a loss rate built on 14 data points is bullshit no matter how nicely the output is formatted.

So I'd keep a hard line. The AI computes and drafts, and a person owns the pool definitions, the individual assessments and every overlay, with the inputs and outputs saved as evidence. That's also what keeps you in good shape when the auditors ask, and we covered the control side in letting AI into the close without blowing up your SOX audit. If you're shopping for tools that do this natively instead of building it yourself, the AR and close platforms in our accounting software directory are a decent place to see who has real credit-loss support and who has a slide about it.

What I'd do

Take your last eight quarters of aging and write-offs and rebuild the allowance in a spreadsheet next to the one you booked, using AI to do the bucketing and the loss-rate math, then compare the two. If they land within a rounding error you now have a repeatable monthly process and a cross-check you can show your auditors. If they're far apart, you just found out something about your reserve that was worth finding out in October rather than in February.

Either way, do it on one pool before you do it on all of them, and keep the prompts you used. Building the first version badly beats waiting for the perfect tool, because the damn allowance is due every quarter regardless, and the 80 employers asking for CECL experience are mostly asking for the person who already tried.

Frequently asked questions

What is a CECL allowance?

It's the reserve a company books for expected credit losses over the life of its financial assets, based on ASC 326 (ASU 2016-13). Instead of waiting for a loss to become probable, you estimate lifetime expected losses up front and update the estimate every period.

Does CECL apply to trade receivables?

Yes. Trade receivables and contract assets from ASC 606 fall within the scope, which is why non-financial companies carry a CECL allowance too. ASU 2025-05 added a practical expedient for current receivables and current contract assets.

When does the ASU 2025-05 practical expedient take effect?

For fiscal years beginning after December 15, 2025, including interim periods within those years, with early adoption permitted if the financial statements haven't been issued yet.

Can AI calculate the CECL allowance?

It can do the aging, loss-rate math, roll-forward and memo draft well. Pool definitions, individual assessments and forecast overlays stay with an accountant, and the inputs and outputs should be saved as audit evidence.

Do accounting jobs ask for CECL experience?

Some do. As of October 2026, about 80 of roughly 67,800 live postings in our database name it, and among controller, senior accountant and accounting manager roles that name it, 22 of 26 also want a CPA.