What are the most common back-office automation patterns?

Five recur everywhere: intake-extract-update (documents in, systems updated), reconcile-and-flag (two sources compared, mismatches surfaced), chase-and-collect (missing items pursued politely and persistently), watch-and-escalate (deadlines and thresholds monitored), and assemble-and-send (recurring outputs built from live data).

Which back-office automation pattern should come first?

Whichever one your volume says: score the queues by hours consumed, structure, and stability. In practice intake-extract-update wins most often because document handling eats the most salaried reading, but the scoring beats the folklore.

Which patterns need AI and which need plain automation?

The split runs through reading. Watch-and-escalate and assemble-and-send are usually pure rules: cheap deterministic automation. Intake-extract-update and reconcile-and-flag need reading, so their middle step is a model with evals. Chase-and-collect is rules plus drafting, a hybrid.

Five back-office automation patterns

The short answer

Most back-office automation reduces to five recurring patterns: intake-extract-update, reconcile-and-flag, chase-and-collect, watch-and-escalate, and assemble-and-send. Two are usually pure rules and cheap; two need reading and therefore a model with evals; one is a hybrid. Naming the pattern before buying anything prevents the classic mistake: paying agent prices for cron-job work, or scripting a task that needed a reader.

Back-office processes feel endlessly particular to the people running them, and structurally they repeat like wallpaper. Five shapes cover most of what finance, HR, and operations admin do between decisions. Name the shape, and the tooling question mostly answers itself.

The five patterns

–Intake-extract-update. Something arrives (invoice, form, order, application), facts get pulled out, systems of record get updated. The volume king: the reading-heavy middle in its purest form.
–Reconcile-and-flag. Two sources that should agree get compared (contract versus invoice, ledger versus statement, system versus spreadsheet), and disagreements surface with the discrepancy named.
–Chase-and-collect. Missing items pursued: unsigned documents, unpaid invoices, unfiled expenses, with polite persistence, context in every nudge, and escalation when silence wins.
–Watch-and-escalate. Deadlines, renewals, thresholds, and anomalies monitored, with owners alerted before lapses instead of after.
–Assemble-and-send. The recurring outputs (weekly reports, board packs, client statements) built from live data and delivered on schedule.

The rules-versus-reading split

Run each pattern through the hierarchy of fixes and the tooling sorts cleanly. Watch-and-escalate and assemble-and-send are usually deterministic end to end: dates, thresholds, and templates, which means scripts and scheduled jobs, cheap and exact. Intake-extract and reconcile need reading in the middle: the invoice that is almost the standard format, the near-duplicate record, so their core step is a model with evals and a human on the flagged cases. Chase-and-collect is the hybrid: rules decide who to chase and when; drafting the chase that a human would actually send is reading's cousin, writing. Paying agent prices for the deterministic patterns is the most common overbuy in this category; scripting the reading patterns is the most common underbuy, and the weekly-breaking bot is its signature.

Name the pattern and the tooling answers itself: rules for the rules, readers for the reading.

How patterns compose

Real processes chain the shapes: supplier onboarding is chase-and-collect feeding intake-extract-update feeding watch-and-escalate; month-end is reconcile-and-flag feeding assemble-and-send. The chains are why owned tools beat tool sprawl here: five patterns wired inside one small system share data, logs, and boundaries, where five subscriptions share only an invoice pile. It is also why the back office migrates well from the spreadsheets currently holding the chains together: the sheet was the composition layer all along, just an unowned one.

Picking the first one

Not by pattern elegance: by the queue's numbers. The scoring method (volume times structure times stability) ranks the five as they occur in your operation, and the honest ROI arithmetic keeps the winner honest after it ships. In most operations the answer lands on intake-extract-update because document reading eats the most salaried hours, but the point of scoring is that yours gets to be different. The patterns are the vocabulary. The queue is the sentence.