Finance teams come to us for one of three reasons. Month-end is taking nine days. AP keeps absorbing every new hire. The CFO is quitting. Sometimes all three. The fix is rarely a new tool — it is finally giving the existing tools a common nervous system.
The bottleneck
The most common shape we see: four pieces of finance software, each bought to solve a real problem, none of them speaking to each other. The team is not paying invoices. They are reconciling between the AP tools so the books will close. Three of the five AP people are doing copy-paste work that nobody told them was their job.
A useful gut check: if your AP team is bigger than two people on under $20M in revenue, the bottleneck is almost certainly connective tissue, not headcount. Adding a sixth person does not fix it. It absorbs the missing tissue into a salary line.
What we actually fix
- Invoice classification. Fine-tuned on three years of your historical GL data. 96% accuracy in the buckets that matter, with the model's reasoning attached for audit.
- Reconciliation. One integration layer reads from accounting, bill-pay, and expense; writes back to the system of record. No new SaaS. No GL migration.
- Vendor follow-up. Drafted emails for missing PO context, tone-matched to your existing communication. The AP person reviews and sends.
- Month-end packet. Variance audit, exception flags, and a board-ready close summary generated automatically — checked by your controller, not written by them.
A real engagement
A $5M DTC company with five AP people, nine-day close, and a CFO threatening to leave. Two weeks of paid diagnostic. Six weeks of build. We kept the accounting system, kept the bill-pay rails, kept the expense tool. We added one integration service (~600 lines of code) and one fine-tuned classifier. AP went from five people to two. Close went from nine days to two. Annual unlock: $340k. The CFO stayed.
The least-glamorous answer is almost always the right one when the bottleneck is connective tissue.
That engagement is the field note titled How a $5M DTC company actually runs its books. The diagnostic call, what was broken, what we kept, what AI got to touch.
Where AI shows up
In two places, both narrow. One: classifying invoices into the right GL accounts — the 4% the model is not confident about route to a human with the model's reasoning attached. Two: drafting vendor follow-up emails when an invoice is missing PO context, with the tone and escalation history baked in. That is the AI footprint. It is small on purpose.
We measure every AI step against a deterministic baseline. If the model is not measurably better than a rule, we ship the rule and move on.
Where AI does not
Approval routing. Fraud checks. Anything that touches a payment leaving your bank account. Those stay deterministic, audited, and human-owned. We do not build AI into the path of money going out the door. Anyone who tells you they do is either lying or about to be sued.
Engagement shape
- Two-week paid diagnostic. Operator interviews, workflow shadowing, the real bottleneck map. Credit applies to any build.
- Build sprint, 3–4 weeks. Senior operators in the room, code in your repo, KPI signed off in writing.
- Two rounds of post-launch tuning. Then your team owns it. We hand over the code, the eval harness, and the runbook. No retainer required.
If your AP function feels like it absorbs every new hire, book a diagnostic. The first call is 30 minutes and it costs nothing.