0 result
Behind the build: what general AI gets wrong about equity, and how we fixed it
We've been watching a macro trend play out across the industry for the past couple of years. Web traffic is shrinking. Not just ours, but across the board. More and more people are going to AI tools first and search engines second. When we noticed fewer people using our calculators, we started investigating.
What we found was pretty consistent: people started using ChatGPT for the equity questions they'd normally bring to us. "Should I exercise my incentive stock options (ISOs) before the end of the year?" "What's my alternative minimum tax (AMT) exposure if I exercise 10,000 shares?"’
Conversational AI is fast and always available. But when we looked closer at the answers people were getting, we knew we had an opportunity.
General-purpose AI tools are designed to be helpful and confident, even when they lack the context and capabilities required to be accurate. Ask ChatGPT to calculate your AMT liability, and it'll give you an answer. Ask again with slightly different phrasing, and you might get a completely different number. That's not a bug. It's how this technology works. They're prediction engines, not calculation engines. As one analyst put it in a recent New York Times piece on AI and taxes, “these models give you what is roughly the right answer, and that's not what you want.”
The data backs this up. TaxCalcBench, an independent benchmark that tests leading AI models on tax calculation accuracy, found that even the best-performing model produced strictly correct returns only about 63% of the time. And consistency is its own problem: ask the same model the same question multiple times, and you'll get different answers. For general tax filing, that's a known limitation worth flagging. For equity decisions with six-figure implications, it's a different problem entirely.
Equity math makes this harder, not just because general AI tools lack your specific grant details. The problem runs deeper than missing data. Accurate equity planning requires working across several genuinely difficult layers at once: your equity itself, which is a dense data model with plenty of company-specific variation, especially around vesting schedules and grant types. US taxes, which layer federal and state systems, and then add the subtleties of AMT and ISO qualifying dispositions on top. The near-total absence of public information about private companies, which means employees are trying to build intuition around an exit value they can barely see. And then, underneath all of that, the ability to model it out across multiple scenarios and variants simultaneously. A model that approximates any one of these layers can introduce errors that compound through the rest.
We'd already spent years building the systems that could handle this. So we decided to use them.
As a product team, we were in an interesting position. We already had the hard parts.
Secfi has a proprietary tax calculation engine, not a heuristic, not a prompt. It knows how ISO exercises interact with AMT. It can model the difference between exercising today versus at exit, across multiple scenarios, using your actual grant data. We'd built this infrastructure to power our tools, our advisors, and our platform. Our advisors rely on it for clients at 90%+ of US unicorns.
The calculation layer existed. The data layer existed. What we didn't have was a conversational interface that connected them in a way people could actually use. So, that's what we set out to build.
We chose Claude, Anthropic's AI model, as our foundation. But the model itself was only the starting point. The real work was building Secfi's specialized logic, data connections, and calculation engines around it so that the answers it gives are grounded in real math, not pattern-matching.Hooking a language model to an equity planning engine sounds straightforward. In practice, the model sits on top of dozens of Secfi's data sources, like a user's holdings, their tax situation, real-time market data, equity modeling, and Secfi's proprietary data and knowledge base. It has to decide which to query, in what order, and how to combine the results into a clear answer. It needs to pull the right context without being asked. And when information is missing, it needs to be transparent about it rather than filling gaps with something plausible-sounding. "Trust the AI" isn't a good enough answer when someone is deciding whether to exercise $500,000 in options.
That last point drove a lot of our product decisions. Every number Maeve surfaces from our calculation APIs is cited inline, so you can see exactly what was calculated by Secfi and what came from the AI model's own estimations, which are inherently less reliable. It's the difference between an answer and a verifiable answer. We wanted you to be able to follow the math, not just read the conclusion.
Our team said: "The hardest part was getting Maeve to lean on Secfi's verified data while still benefiting from the broader capabilities of a generalist AI model. The citation system required some out-of-the-box thinking, which is the piece the engineering team is most proud of."
When you ask Maeve a question, it's interacting with services and tools that a general AI tool simply doesn't have access to: an equity service, a tax service, a calculation engine, and live market insights.
Your equity and tax data. Grant details, vesting schedules, exercise history, and tax filing status. Maeve works with your actual numbers, not an incomplete picture. Most people who turn to ChatGPT describe their situation as best they can, and the model works with whatever fragments it's given, confidently filling in the rest. Maeve doesn't do that. It works from your data, or it tells you what's missing.
Secfi's tax engine, the same calculation APIs our advisors use, AMT modeling, exercise scenario comparisons, and capital gains projections. Real math, not pattern-matching. Live market signals, 409A valuations, preferred pricing, fund marks, and secondary market data. For market data, Maeve pulls from Pitchbook and Caplight. Caplight's valuation estimates are particularly useful here because they're built from public comps and secondary market trading activity, which means employees get a grounded sense of what their equity might actually be worth, not just what was set at the last 409A.
Because whether it makes sense to exercise often depends on what your company is actually worth right now, not what it was worth at your last grant.
We didn't build Maeve in isolation. We built it as part of a shift we've been watching happen across the AI space.
For a while, the story around AI was about breadth: how many domains could these tools be useful in? That question has been answered. The more interesting question now is where AI can be genuinely reliable, not just helpful, in domains where the stakes are high enough that approximation isn't good enough. You're already seeing it: Harvey for legal work, Copilot for code. Tools built on top of domain-specific data and logic, not just a general model with a different interface.
Equity planning belongs in that category. The decisions people are making with their stock options are some of the most financially significant of their careers. They deserve AI that was actually built for it.
Maeve is live now and free to use. It's built for the questions that actually keep you up at night: whether to exercise before a tender offer closes, whether to accept a secondary bid, whether to take more options or more cash in a new offer, whether a down round changes your calculus entirely. The kinds of decisions where getting the math wrong has real consequences.
We built it because the infrastructure to do this well already existed. It just needed a front door.
Your equity is likely one of your most valuable assets. It deserves AI that actually does the math.