Digits AI Categorization Explained: How It Actually Works

Digits AI categorization is the part of the platform everyone notices first: transactions show up already sorted into accounts, with no rules engine to configure and no CSV to hand-tag. For a Marketing and Advertising Agency running a dozen ad platforms, a handful of 1099 contractors, and retainer deposits that don't map cleanly to a single invoice, that first impression is also where most of the misplaced trust comes from. The AI is doing real work, but it's narrower work than it looks like from the dashboard. Here's what's actually happening under the categorized feed.

What Digits AI categorization is actually doing

The engine is matching each new transaction against patterns it has already seen: the same vendor, the same amount range, the same account, the same memo text, recurring on a similar schedule. When a charge from a recognized vendor lands in a familiar amount band and a category has been confirmed for that pairing before, the system assigns it automatically and attaches a confidence score to the call. High scores, generally north of 95%, post without any flag. Mid-range scores, roughly 80-95%, post but sit in an exception queue for review. Anything lower gets held entirely until a person assigns it. That three-tier split is the actual product: not categorization itself, but a triage system that decides which transactions are safe to trust unattended.

Where it gets agency transactions right, fast

Recurring, single-purpose charges are where the AI earns its keep. A monthly SaaS subscription, a biweekly payroll run, a fixed-fee software renewal, a rent payment: these hit the same vendor, same amount, same cadence every time, so confidence climbs quickly after two or three correctly confirmed instances. For a 15-person agency, that's often 50-60% of total transaction volume by count, even if it's a smaller share of total dollars. Clearing that volume automatically is the real time savings, and it's what the platform is built to do well.

Where it breaks on agency-specific money movement

Three transaction types common to agencies consistently land in the low-confidence tier, for the same underlying reason: the dollar amount alone doesn't tell you what the transaction is. Ad platform charges from Meta, Google, or TikTok vary every billing cycle because spend varies, so the amount-matching signal the AI relies on for recurring charges never stabilizes. Whether that charge should hit a pass-through liability account or a cost-of-delivery expense account depends on which client it's for and how your contract treats media spend. We cover that split in detail in ad spend pass-through accounting, and no AI categorization engine can infer a contract term from a bank memo that just says "GOOGLE ADS." Retainer deposits have the opposite problem: the amount is often identical every month, but whether it's earned revenue, a deposit against future work, or deferred revenue waiting on delivery depends on the service period, not the transaction itself, a distinction we walk through in retainer revenue recognition. And 1099 contractor payouts routinely get miscategorized as regular operating expenses instead of being tracked against the $600 reporting threshold, because the AI has no visibility into cumulative payments to that vendor across the year unless someone has tagged the vendor as a 1099 contractor from the first payment.

The correction loop is where the real training happens

Every manual recategorization feeds back into the model, but it's weighted toward the specific vendor and account pairing you just corrected, not toward agency transactions in general. Fix one Google Ads charge for one client and the system gets better at categorizing that client's Google Ads charges. It does not generalize that correction to a different client's Facebook Ads charges, because the pattern it learned was narrow by design, which limits false confidence. The practical upshot: the first one to two months on any new client or ad account involve more manual correction than the system's reputation suggests, and skipping that correction period doesn't skip the work, it just moves the error further downstream to whoever reviews the financials at month-end.

What this means for your month-end close

The exception queue, not the auto-posted feed, is where a month-end close actually happens. A reviewer who only looks at total transaction count and sees 92% auto-categorized can walk away thinking the books are 92% done. In practice, the 8% sitting in the queue is disproportionately the ad spend, retainer, and contractor transactions that determine whether gross margin is calculated correctly, so that queue needs a human who understands the agency's specific contracts and chart of accounts, not just someone clearing a backlog. We cover the full sequence in our month-end close process for agencies.

Should you trust it without a bookkeeper reviewing it?

For a single-owner business with no contractors and one or two recurring client charges, running the AI with light oversight is reasonable; the error surface is small. For any agency running ad spend pass-through, a contractor bench, and more than a handful of retainer clients, the AI is a filter that clears the easy 50-60% of volume so a bookkeeper's time goes toward the transactions that actually require judgment, not a replacement for that judgment. We go deeper on where Digits fits an agency's specific workflow in Digits for marketing agencies and cover the platform more broadly in what is Digits accounting software.