We started with one CPA's impossible box of paper.
Accord AI didn't start as a product idea. It started as a favor for a CPA, then got rebuilt twice because the people asking kept raising the stakes.
- Before
- Built and shipped several companies
- Now
- Runs every engagement's quality bar personally
- james@accordai.net
A CPA asked whether the messy books could be automated.
James Dougherty had built companies before. What he hadn't done was sit with an accountant's onboarding problem: every new client arriving with years of messy historical books, and a tax deadline that didn't care. Before anyone could give advice, before the real conversation could even start, someone had to reconstruct the financials by hand.
The ask was narrow: could this be automated well enough to produce a trustworthy first set of financials? That question turned out to have a much longer answer than anyone expected.
A forensic client changed what "accurate" had to mean.
The first version shipped as software: a SaaS tool for cleaning up historical books. It worked. Then a CPA client brought in a forensic matter, and the tolerance for error collapsed. In bookkeeping, a misread figure is a correction. In a forensic engagement, it is a question on the stand.
We chose to be a service, not another subscription.
The obvious move was to sell the forensic engine as software. James decided not to. Firms are tired of another login, and nobody wants a self-serve upload box for evidence. More than that, human review is not a nice-to-have here. When the output has to survive scrutiny, a person has to sign off.
So Accord AI is a technology-enabled service. We own the engine, the review infrastructure, and the analysis suite. A US-based team reviews the deliverable, and samples the extraction using stratified sampling under AICPA standards. You get the throughput of automation, with a named person accountable for the ledger.
Finding the fact pattern became the obsession.
Somewhere along the way, the technical problem stopped being the interesting part. What holds our attention now is the moment a case starts to take shape in the numbers: the payee that shows up in three accounts, the round-number cadence that should not exist. Chasing that is why the quality bar sits where it does, and why it is not negotiable.
Three things we will not bend on.
A human signs off
Low-confidence fields always go to a named reviewer. A US-based team also samples the extraction under AICPA standards and reviews the deliverable. We keep a record of who made each decision or correction.
Numbers come from tools, not models
Detection and reconciliation are deterministic functions. Run them twice on the same evidence and you get the same result. That is what defensibility requires.
The engagement stays consultative
We scope around the case, not around seats. You talk to the people doing the work, not to a support queue.
Bring us your hardest case.
We'll process your pilot at a discount and you'll get back a structured, source-linked database.
Don't have a tough case right now? Send us a sample of one you are working on, or one you worked on in the past.