What AI Actually Does at Gelt (and Who Is Allison?)

What is the AI in our platform actually doing, and where does the human CPA start? To answer it plainly, we sat down with three people who live it every day.

In this conversation:

• Tal Binder, CEO and Founder
• Rachel Richards, CPA, Head of Tax Product
• Fraser Robinson, Head of AI, R&D

Why can't AI just do taxes end to end?

Tal: Two things. The first is liability and accountability. If you want a licensed professional to sign your return, someone is taking responsibility for what gets reported. Any professional who respects themselves reviews whatever the AI produces. Unless the AI companies are willing to sign the return and carry that liability themselves, the human stays in the loop. Self-signed returns also correlate with higher audit exposure, so that accountability matters.

The second is data quality. The output of AI is only as good as the input. What we see again and again is that the information people hand a system is not at the level AI needs to produce a fully automated, correct return or projection. AI does not have clean, complete control over all of a taxpayer's information, and it does not take on the accountability for the result. Those two gaps are why it cannot solve tax end to end today.

So where does AI begin and end at Gelt?

Fraser: AI begins at the data. AI is fundamentally a transformation of data, whether that is a document we are processing or the billions of tokens the model was trained on. Anywhere there is data in a process, AI at least wants to have a look in.

Rachel: And from the client's side, AI enhances the experience. We are not replacing the CPA relationship. AI helps our CPAs get better answers faster, consolidate context, and put information together in ways that are not realistic to do by hand. It helps the team work faster and smarter, and deliver better results consistently and at scale.

Give us a concrete example where AI caught something or saved real time.

Tal: Reasonable compensation. To land on a reasonable salary for an S-corp owner, you need a defensible study behind it, and it is a spectrum, not one number. The tax team used to spend a lot of time getting there. Now it takes seconds.

Rachel: That is a good one. When a return or projection comes in, either prepared by our team or provided by the client at onboarding, we use AI to read it and then reuse that information throughout the process. We extract the numbers to build estimates, planning, and safe harbor calculations, and to surface insights about the return. The reasonable salary study is the clearest example. Finding the basis to support a recommended salary used to take an exorbitant amount of time. With AI, it takes a couple of minutes. It saves time and unlocks work that is not accessible to everyone without the technology.

[cta_block]

You mentioned an internal assistant. What is it?

Rachel: Allison, our internal AI assistant. One challenge a CPA faces with a full book of clients is context. If it has been a few weeks or months since your last conversation, all of that history matters for planning and for being proactive. Allison goes through meeting transcripts, email, and platform data, and consolidates it so the team has the context at their fingertips right when they need it. It helps make sure nothing falls through the cracks.

Fraser: The other benefit is the second set of eyes. A person usually wants someone to check their work. You can ask Allison the same question you are trying to answer, and it is like a colleague saying, "I agree with this," or, "Did you see this email?" It catches the things a human naturally misses.

Will clients ever talk to Allison directly?

Tal: Not anytime soon, and it comes from listening to people. When a client asks a question, they want a CPA to answer. If AI answers and a CPA still has to review and approve it, that is an extra cycle for no gain. More importantly, any recommendation a licensed professional makes carries weight. We cannot hand that weight to an AI and let a client treat its output as formal advice when it is not. So this stays with our CPAs.

Where do you want the product to go next?

Rachel: Real-time tax projections and insights. Calculations are a real limitation today, and honestly that capability is not available to people anywhere right now. Taking everything we know about a client, what they have done and what they are considering, and turning it into insights they can act on to optimize their taxes for the current year. That is what I am most looking forward to.

Fraser: We are building a data refinery, a first-class, evidence-backed system that goes well beyond throwing documents at an assistant and asking questions. Right now we are deep in the document processing and pipeline part of it. Once it is running, I expect an explosion of tools we can build on top, because this is the foundation everything else sits on.

Tal: Two areas matter most to me. Compliance, meaning the filing and management work, where taking 90% of the lift off our professionals would be a huge time saver. And calculations. A statistical model is bad at math, so even with perfect context, a calculation it produces can be flat wrong. Giving AI the ability to run accurate calculations unlocks real scenario modeling for clients. Beyond that, we see AI as the main tool for operational efficiency, not as a chatbot clients talk to all day while professionals rubber-stamp the output.

Clients hand you their most sensitive data. How does that shape what you automate?

Tal: This is a big internal discussion. Connecting sensitive client data directly to an external AI model can expose that information to attacks, and that risk is real. So today we do not share clients' private information with an external model automatically without controlling the flow end to end. We are constantly working on how to do it safely, and if we cannot get comfortable that a client's information is protected, we will not do it. Cost is the other factor. AI priced purely on tokens gets expensive fast, and there is a point of diminishing returns where the cost outweighs the value unless you pass it to clients, which we do not want to do.

Fraser: Information security is the first thing we discuss when we evaluate any new solution. How you interact with LLMs and still build secure systems is uncharted territory for the whole industry right now. On top of that there is hallucination. It is not only about not exposing data, it is about making sure the system does the right thing with the data it does have. That is front and center in everything we build.

Can AI turn a good CPA into a great one?

Tal: It is specific to each firm and each tech stack. We invest heavily on two sides: giving our CPAs the tools, and training them to use those tools properly so they get the most out of them. A great tool used inconsistently is worth nothing. Both sides have to be there.

Rachel: I would go further. You can make a good CPA great. Look at what the great ones notice, the questions they ask, what they do to deliver exceptional service. There are patterns in that, and you can turn those patterns into tools that help a good CPA elevate their skills and deliver a level of service they might not reach on their own.

Fraser: I land in the middle. AI can make a good CPA great, but it cannot make a bad CPA good. It is a force multiplier. Good becomes great. Bad just becomes bad with fingers in more pies.

Tal: Which is exactly why good people matter. Even with perfect technology, without the right people to operate it, clients will not be happy. It starts with good people, and it ends with AI.

Frequently Asked Questions

Can AI do your taxes without a CPA?
Not reliably, for two reasons. A licensed professional who signs a return takes legal responsibility for what gets reported, and AI companies are not willing to carry that liability. And AI output is only as good as its input: most taxpayers' information is not clean or complete enough for a fully automated, correct return. Self-signed returns also correlate with higher audit exposure.

What is Allison, Gelt's AI assistant?
Allison is Gelt's internal AI assistant. It consolidates meeting transcripts, email, and platform data so CPAs have full client context at their fingertips, and acts as a second set of eyes that catches things a human naturally misses. Clients do not interact with Allison directly; every recommendation still comes from a licensed CPA.

How does Gelt use AI in tax planning?
AI reads returns and projections as they come in, extracts the numbers to build estimates, planning, and safe harbor calculations, and surfaces insights about the return. The clearest example is the reasonable compensation study for S-corp owners, which used to take hours of research and now takes minutes.

Does Gelt share client data with external AI models?
No. Gelt does not automatically share clients' private information with an external AI model without controlling the flow end to end. Information security is the first consideration when evaluating any new AI solution, and if the team cannot get comfortable that client data is protected, it does not ship.

Want a tax team that pairs licensed CPAs with technology like this? Schedule a call with Gelt.