Messy Inputs Limit AI Bookkeeping
Andy Su, co-founder of InDinero, on tech-enabled bookkeeping's 14-year evolution
The bottleneck in accounting is not model intelligence, it is messy inputs and missing business context. A bookkeeping system can learn that a recurring check is probably rent, but it still breaks on edge cases like unlabeled checks, mixed Amazon purchases, invoices that arrive after the service period, or contracts that create accruals before cash moves. That is why bookkeeping remains a human in the loop workflow, with automation helping on the repetitive parts rather than replacing judgment end to end.
-
Most tech enabled bookkeepers still run on QuickBooks as the ledger of record, then add their own customer workflow layer on top. The software pulls data from Stripe, Gusto, cards, and banks, but humans still reconcile records, chase missing context, and approve categorization when the raw feed is ambiguous.
-
The real friction is not just classifying card charges. Accounting also needs non cash events like accruals, deferred revenue, lease obligations, and contract terms that never appear cleanly in bank data. That work often means reading invoices and contracts, then turning them into journal entries, which is where rules alone still fall short.
-
Owning monthly books creates the right to sell adjacent services because the provider already has the records, the customer relationship, and the trust. In practice, the first add ons are usually tax, R&D tax credits, and FP&A, because they reuse the same underlying financial data and solve urgent money questions for the customer.
The next phase is not full autopilot bookkeeping, it is narrower automation applied to specific workflows with clean feedback loops. The winners will be the firms that use AI to read documents, suggest entries, and ask for missing context earlier, then turn bookkeeping into a broader finance control point for tax, planning, and compliance.