“It’s rare these days for an AI company to actually want to demo the product,” Blake Oliver, CPA, said at the start of this Earmark Expo. “Maybe it’s because there is not that much there in a lot of cases.”
Savant Labs showed its product anyway. After Shaan Mistry (VP Solution Consulting, Savant) set the scene, Barry Burns (Co-Founder and VP Customer Success and Community, Savant) built an automated month-end trial balance live while Blake and David Leary asked questions. The session made a useful point for anyone being told to “use AI” in their close. As Barry put it, AI can’t do your close today. The payoff comes from using an LLM to build and document a deterministic, auditable workflow, so humans and controls own the answer instead of the model.
Below, we cover why LLMs fall short for regulated work, how the demo turns AI output into a fixed process, which controls make it auditable, and what questions remain.
The problem: AI that can’t show its work
Shaan said the core issue is LLMs aren’t deterministic. “Ask it a question today, and it gives you an answer. Ask it tomorrow, and it may give you a different answer.” There’s also no audit trail, so the results can’t withstand regulatory scrutiny.
Barry pointed out that a forecast only needs to be “directionally correct,” so it’s fine to have Claude help you build one. The close and tax returns are different, because each has one right answer that you have to prove. “You can’t use AI for that,” Barry said, at least not today, “because you can’t show your work.” You also can’t show a record of humans making decisions and approvals along the way.
David shared a familiar headache. When something lands in QuickBooks through an API, a bad connection can make it look like he did it, even when it came from Blake or Claude. Shaan put the stakes plainly: accounting, finance and tax are heavily regulated, and they “can’t afford to have AI making things up.”
Let AI write the process, not the numbers
Savant points the AI at the process, not the result. Barry, playing a senior accountant at a fictional company, opened Claude and asked Savant’s connector, Savvy, to help automate his month-end trial balance. He supplied August and September NetSuite exports, a chart of accounts and his adjusting entries.
Savvy asked clarifying questions before jumping into classifying transactions. Should only approved entries post, or proposed ones too? Was the default flux threshold of 10% or $10,000 acceptable? Then it produced three things:
- Process documentation that identifies the preparer, reviewer, escalation owner and workflow owner. Barry noted that the document works even outside Savant, in Excel or any existing process.
- A proposed work paper with tabs that match the process documentation.
- A canvas workflow in which every step is a deterministic data process. “The same inputs go in. The same outputs come out every time.”
The workflow targets a specific pain point: restating the trial balance “four or five times in the same month,” with each restatement taking “five hours to readjust.” NetSuite exports add their own chore, because you have to calculate balances manually. Blake joked, “You have never had a problem with reporting until you’ve seen a NetSuite export.”
“Savant isn’t really doing my accounting AI for me; it’s helping me build a deterministic process,” David said, summing it up.
Controls an auditor would recognize
What makes this usable for accountants is governance. When Barry edited a variance rule on the canvas, nothing changed in production. Every change goes through maker-checker approval. The submitter adds notes, someone with the right access approves the change, and only the published version runs. “No one can come in here and just cowboy it,” Barry said.
The run history records each execution, the version it used and the data at every step. Barry pointed to the classic Excel nightmare, where, after 30 changes to a work paper, someone asks, “What was change number 15? That’s so hard in a manual Excel-based process.”
Other controls include:
- Automatic process documentation updates: Publishing a new workflow version updates the documentation, so it doesn’t drift out of sync.
- Logged review decisions: Review items are sent to reviewers through email or instant messaging apps like Microsoft Teams. Reviewers’ decisions feed into the next run, and the run history captures them.
- SOX activity log: It records who accessed, modified, submitted, approved and ran each flow.
Blake described the overall effect as, “You’ve taken the probabilistic technology and made it deterministic.”
Reality check: adoption and lock-in
Barry, putting on his “sales hat,” described controllers who are “getting pressured to leverage AI” while their teams have been “doing things the same way in Excel for 20 years.”
David raised a fair concern that past the first prompt, the canvas involves “a lot of clicking.” Barry pointed to the Savant Academy courses and live onboarding, as well as Savvy, Savant’s AI copilot, which helps users make changes through plain-language prompts. He said some users sign up on Monday and have their first workflow in production by Friday. This was a vendor demo, so take the claims with a grain of salt. The session didn’t test them.
David also asked what happens if a firm switches from Claude to Copilot. Barry said the workflows are LLM- and ERP-agnostic and would stay intact.
If you can’t show your work, you can’t sign it
The case Savant made is that the defensible use of AI in the close today is building governed, repeatable processes, not generating answers. You can’t stand behind a number you can’t trace. A deterministic workflow with version approvals and run history gives you something to defend in an audit.
The demo covered a trial balance, but Barry listed reconciliations, accruals, flux analysis and apportionment as other candidates. Watch the full on-demand webinar to see the live build, and judge the workflow and audit trail for yourself.