Black Ore’s Tax Autopilot now generally available

AI financial services platform Black Ore announced the general availability of its Tax Autopilot solution, which is claimed to be capable of executing even complex tax workflows from start to finish with little to no need for human intervention. 

Processing Content

Overall, Tax Autopilot is said to be able to execute the full lifecycle of a complex tax return autonomously with no need for human intervention from Black Ore experts. To do so, it can accept a variety of document types, including W-2s, 1099s, K-1s, brokerage statements and bank records in any format. It can then classify those documents and extract their data automatically, even resolving discrepancies between them. Using this data, it then applies federal and state tax code logic to prepare a complete return and produce detailed supporting workpapers for professional review and final signoff. 

Other features include the ability to auto-generate client data request lists, Excel-based tie-out packages, AI-powered anomaly and error detection, centralized document storage and indexing, evidence-linked review and preparation steps, real-time validation and evidence tracking, exportable audit packages, and the ability to connect with major tax software platforms so the platform operates directly within the relevant workflow. 

Eyal Shinar, co-founder and CEO of Black Ore, in an email, emphasized the software’s autonomy. He said some products that are marketed as AI actually run on a hybrid model where the technology handles some parts and the vendor’s own staff handles the rest. 

“Tax Autopilot does not work that way,” he added. “The firm uploads documents, the platform runs end to end — ingestion, extraction, return preparation, workpaper generation, software integration — and the firm’s CPA reviews and signs off. No Black Ore employee touches customer returns. No client data passes through a third-party preparer.”

Asked about the degree of the AI’s independence, he said that for returns ranging from straightforward to moderately complex, it is 100% touchless. The CPA can queue 10 or even 100 returns at once, walk away to get coffee, and come back to finished work products ready for review in just minutes. 

For more complex tasks, he said it’s about 98% touchless. A human CPA is needed to attest to the return before it goes to the client and the IRS; and to address edge cases that need more attention, such as a partnership ambiguity that cannot be resolved from documents alone, or a transaction where the right tax treatment depends on context the system does not have. In such cases, the CPA answers and the system continues, just as a human CPA would after a clarifying question. 

“Outside of those edge cases, the platform does not need someone in the chair,” he said. 

Under the hood operates a proprietary AI layer trained on what Shinar described as “how the tax code actually works, including even the parts that are not truly documented,” which acts as the main tax model. This is then combined with another proprietary AI and orchestration model that uses deterministic, not probabilistic, tools much in the way a human accountant would. In this respect, he said, Black Ore heads off potential hallucinations to give people confidence in the outputs. 

“The system operates like an accountant: When it hits something it cannot accurately reason through and requires further context — like a specific question only the client can answer — unlike general AI, it doesn’t hallucinate. It pauses and asks the CPA. This is what allows our AI to take all the needed actions (document fetching/organization, data extraction, calculations, etc.), and generate outputs that are auditable, accurate and trustworthy enough for production use at scale across the largest CPA firms in the country,” he said. 

The product launch comes after a two-year early access period, during which time Black Ore onboarded 75 firms, which includes about 40% of the top 20. The company was able to wait as long as it did because it had been able to raise $60 million upfront (see previous story), which allowed Black Ore to skip the minimal viable product phase and take the time to build the platform it wanted from the beginning. 

That also gave Black Ore time to incorporate feedback from its early users. The platform was, and still is, designed to compress the functions of admin, junior staff, senior staff and reviewers into a single agentic workflow. But as time went on, the company added features and capacities aligned with how firms operate today, such as tie-outs generated for seniors, downloadable binders and reviewer workflows for managers. 

It also expanded options for firms to apply their own judgment where they wanted, as there were concerns from some leaders who, due to workload compression, prepared returns with “a higher level of abstraction, due to workload compression” and so were worried the precision of the outputs would create “noticeable delta versus prior years’ returns,” enough to create real audit risk in the transition year. 

Overall, Shinar described a process where, rather than demand the firms adapt to the technology, Black Ore instead adapted the technology to the firms. 

“Ultimately our AI can produce a perfect output, but if the process isn’t somewhat natural to how an accountant works today, change management becomes difficult. Consistent feedback across multiple tax seasons helped us get there,” he said. 

While AI has changed a lot over the last two years, Black Ore did not feel the need to dramatically reorient itself around new advances, prioritizing instead control over its own stack while continuously orchestrating between the different models in different parts of the architecture and workflow. This decision was vindicated, said Shinar, during the brief firing of OpenAI CEO Sam Altman, which drove home the importance of owning one’s own AI models and orchestration layers and avoiding being locked into someone else’s. While the work took longer than if Black Ore had just licensed a third party model, Shinar felt the wait was worth it. 

“What has not changed in two years: General-purpose models still do not have a tax ontology. We invested years and tens of millions of dollars to build one. Every advance in the broader AI ecosystem makes that investment more valuable, not less,” he said. “Having said that, the way we work on the engineering side has evolved as we need less and less human engineers. I believe we would soon need less and less of other functions, but the goal is to keep the top 1% of people with the best tools and various AI models that are available.”