Your books run on judgment that was never written down. An invoice lands, and booking it right depends on the purchase order, the contract behind it, what was already accrued, and what an accountant with twenty years in the company happens to know. When volumes grow or that accountant leaves, the close gets slower and the audit gets harder.
The software industry calls this invoice automation and sells extraction. Extract the fields and route the approval. The invoice is just the paper that shows up. The real work is the gap between what happened in the business and what the ERP needs recorded. Today a person closes that gap, one document at a time.
Motive Force has invested in Talky. It is the first system I have seen built to own that translation. Talky learns how your company does its accounting, then posts entries in your ERP the way your team would, with the reasoning attached to every entry. When it meets something new, it asks the person who knows, and the answer goes into what the company knows, so it asks less every week.
Accounting is the hardest place to trust an AI
I spent the first half of October in conference rooms on both coasts, at The AI Conference in San Francisco and AI Engineer in New York. In San Francisco, Ion Stoica opened his keynote with an agent that built the fastest key-value store ever measured by that benchmark. It stored nothing. In New York, the Langfuse team opened their evals workshop with a show of hands. Almost everyone in the room had shipped an agent to production. Asked who had solved evals for those agents, fewer than five hands went up.
The same workshop walked through a support agent that told a customer it had issued a refund. The refund tool was never called. No errors, no latency spike, the dashboards stayed green, and the customer found out weeks later when the money never arrived. Swap the refund for a journal entry and you have books that are wrong and look fine.
At the same conference I heard from Field Guide, which builds AI for audit firms. Audit already has a test for this: an experienced auditor who never touched the work should be able to follow the evidence to the decision. Their speaker applied it to agents, and his warning holds for any system near a ledger: a defense of “the model said so” will not survive an inspection. Producing correct output got cheap. Establishing that the output is acceptable to the people who carry the consequences became the cost. I wrote about this in Correct Isn’t Admissible. Accounting has worked under that rule for five centuries. Double-entry spread because merchants needed to prove their books were right, and the approvals and audits grew on top of it. A system that posts into a ledger meets that bar on day one or it stays a demo.
Talky puts an audit trail under every entry
Ramp told the same conference they spent four to six months building evals before launching their bookkeeping agent. That is what the trust layer costs when you build it yourself. Talky builds it as the product. It starts by collecting what your company already has: the systems and documents, and what people know but never wrote down. It works out how your accounting runs and turns that into learnings. Then it posts entries, and every entry keeps its trail: the learning that filled each field, the source that learning came from, and the person who validated it. Export the trail and hand it to your auditor.
Talky’s published numbers say 86 to 93 percent of documents post with nobody touching them, across more than 1.8 million purchase orders and over €2 billion in spend. What I underwrote is the trail under those numbers. Two years from now, when an inspector pulls one entry and asks why it was booked that way, your team opens the trail and reads the answer off it. That question is where I spent my diligence.
Talky earns the right to post, one entry type at a time
Talky does not ask for the keys to the ledger on day one. Your team starts by reviewing everything it posts. The progression runs from recommending to approving to executing to governing by exception. At the beginning, your people do the accounting and Talky helps. At the end, Talky does the accounting and your people govern the exceptions. The teams furthest along took reviews they had insisted on at go-live and removed them, because the system had earned it, one entry type at a time.
Jorge and Maxim built Talky by sitting beside accountants for weeks before they wrote a line of the product, and it shows in how the system treats the people around it. When Talky does not know, it asks the person who does, and it keeps that person’s name on the answer.
Talky connects to the systems where accounting already lives, SAP and Oracle among them, and deploys with large enterprises and the BPOs that serve them. If you run a finance organization and want to watch your accounting explain itself, book a demo, or write to me and I will make the introduction.
Welcome to the portfolio, Jorge, Maximilian, and the Talky team. I could not be more excited to partner with you.
More essays: Frontline workers software sold for $3.6B (my first check), Investing in Healthcare breakthrough, What makes 5% of AI agents work in production, AI Admissibility in Enterprise Workflow, Mapping work, not markets