■ CRITICAL RISK ■ Finance
Image-based check clearing killed this role, and the decline of checks themselves is salting the earth. AI's only involvement was teaching computers to read handwriting — after that, banks stopped needing rooms of people to encode checks at midnight.
“Check processing is digital now. Your machine is in a museum.”
Our AI replacement risk score — how we score jobs
In every sizable bank's back office through the 20th century, proof machine operators processed the day's checks: keying each check's dollar amount onto the machine, which printed it in magnetic MICR ink along the bottom edge, sorted checks into pockets by destination bank, and totaled each batch to prove — hence the name — that the amounts balanced against the deposit tickets. Speed was everything; skilled operators keyed thousands of items per hour on night shifts timed to clearing deadlines, and an out-of-balance batch meant hunting a transposed digit through a stack of paper at 2 a.m.
Two technologies ended it in sequence. Optical and magnetic character recognition automated the reading — machines that could decipher the courtesy amount, including handwriting, removed the keying. Then the Check 21 legal framework (2004 in the US) made a check's image legally equivalent to the paper, so checks stopped traveling physically at all: branches and even phones capture images, software reads and routes them, and the proof department's pockets, cannonball runs to the Fed, and midnight balancing evaporated. On top of that, check volume itself has been collapsing for two decades as cards and electronic payments took over.
Nothing meaningful resists, because the artifact and the workflow both vanished. The residual human work is image-quality exception review — squinting at the checks the recognition software rejects — and fraud inspection, both folded into small item-processing teams a fraction of proof departments' former size. As checks continue their long fade toward extinction, even that dwindles. This is a 97 of the completed kind: the museum reference in our one-liner isn't a joke, it's a literal description of where the machines are.
Automatability: our editorial assessment of current and near-term AI capability
Effectively concluded. Character-recognition equipment shrank proof departments through the 1980s-90s, and image-based clearing after Check 21 in 2004 eliminated the physical workflow entirely — within a few years, most banks had dissolved their proof operations. The vestigial exception-review work continues to shrink alongside check volume, which falls every year. This is a completed automation story that happened to a large, skilled workforce almost entirely out of public view.
They were the throughput engine of check processing: keying every check's amount into a proof machine that encoded it in magnetic ink, sorted it toward its paying bank, and totaled batches to prove deposits balanced. Banks ran whole night-shift departments of them, since the day's checks had to clear by morning. It was high-speed, high-accuracy piecework at the heart of the payment system.
Recognition technology plus a law. Machines learned to read check amounts, removing the keying; then Check 21 (2004) made check images legally equivalent to paper, so checks stopped moving physically and cleared as data. Proof departments dissolved within a few years of that change. Falling check usage since has erased most of what remained.
Small teams, yes — reviewing images the recognition software rejects, inspecting suspicious items, and handling the shrinking paper remnant. It's a fraction of the former workforce and contracts annually with check volume. The growth in bank back offices moved to fraud analytics, payment operations, and compliance, which is where former item-processing staff largely went.
Watch for the legal switch, not just the technical one. The recognition technology existed for years; the workforce collapse came when regulation made the digital artifact legally sufficient. Any role built on physically handling documents — checks, titles, wet signatures, original records — should track the statute book as closely as the software releases.