Digits

Not All AI Is Created Equal and Your Next Software Decision Depends on Knowing the Difference

Earmark Team · April 17, 2026 ·

When Jeff Seibert ran consumer product at Twitter, he asked the finance team for his budget to throw a team event. They said they’d get back to him in 45 days. So he just ran the event without them.

That gap between real-time data and 30-to-90-day delayed financial reports was frustrating, and it eventually led Jeff to build Digits, a new general ledger designed from scratch for the machine learning age. After raising $100 million pre-launch, testing 2,000 monthly closes, and getting 80% of clients closed in under an hour, Digits launched in March 2025. Now, just over a year later, hundreds of accounting firms are onboarding thousands of clients onto the platform.

Jeff launched Twitter’s algorithmic timeline in 2016, and it was one of the first global deployments of machine learning. Now, the AI revolution Jeff helped launch is flooding the accounting profession with claims that are hard to verify. Every accounting software company seems to include AI in its marketing copy, promising everything from “fully automated bookkeeping” to capabilities that don’t add up under scrutiny.

In a recent Earmark webinar, host Blake Oliver and Rob Hamilton, Head of GTM at Digits, pulled back the curtain on how AI in accounting actually works. He was joined by Megan Reid, Product Specialist & Firm Enablement at Digits, who fielded questions throughout the session.

Every AI claim in accounting software isn’t real. But accountants who understand the four core model types (plus one common lie) will make smarter investments, automate the right parts of their workflow, and position their firms for a shift Rob sees coming by the end of 2026.

The AI hype problem (and one question to cut through it)

Before making any technology decision, you need a filter for separating real capabilities from marketing fluff. Rob offered a simple one that cuts through the noise.

He showed screenshots from multiple accounting software companies making bold AI claims. One promised “fully automated bookkeeping.” Another asked, “Do you do AI bookkeeping or do you use a dedicated team of experts?” The positioning has gotten so confusing that firms can’t tell what’s real anymore.

The confusion isn’t new. About five years ago, tech investor Naval Ravikant tweeted, “In most pitch decks, AI stands for Anonymous Indians.” For a long time, that was literally true. Services like Botkeeper rose and fell using offshore labor dressed up as automation. Today, “AI actually means we just bolted on and sent all of your data to ChatGPT,” Rob explained.

Here’s your filter: “AI is the same thing as machine learning,” Rob stated. “If someone is talking to you about AI and they’re not referring to machine learning as the underlying premise, it’s just BS.”

But this filter only works if you understand what machine learning actually is.

Traditional software is straightforward. You write code that tells the computer exactly what to do. It’s tedious to build, but rock solid once it works. Machine learning flips this completely. You feed the system thousands or millions of examples, and the model learns the patterns itself. As Jeff explained in a clip Rob played, “You give the computer the goal state—I want this outcome—and then the computer itself is learning how to do it.”

These models are neural networks. Thousands of hidden layers mimic how neurons connect, based on Google’s 2017 “transformer” research paper (the “T” in GPT). It’s a massive matrix multiplication problem where the system figures out how variables relate to each other.

But machine learning isn’t one thing. Different model types have different strengths and uses in accounting. Understanding these distinctions helps you avoid buying the wrong software and shows you exactly where AI can save time and where vendors are overselling.

The model types that matter (and one that doesn’t)

Rob walked through five categories that get lumped under “AI,” but understanding the differences is what separates informed decisions from expensive mistakes.

Generative models

Large language models (LLMs) are the ones you hear about most, ChatGPT being the prime example. GPT stands for “Generative Pre-Trained Transformer,” and these models generate the most likely continuation of whatever prompt you give them. Rob showed a useful application: turning bullet-point close notes into polished client emails. His advice is to write a “job description” for the AI once. Tell it who it is, give context, specify output format, add examples. Then just paste in different client notes as needed.

But generative models have serious limits. They’re “super eager” and always want to complete prompts, making them prone to hallucinations, or making things up that sound real. They’re bad at math because they generate text rather than calculate numbers. And they’re trained on the internet, not your specific clients. “The ways that it is hallucinating is stuff that maybe even is hard for humans to catch sometimes,” Rob warned. Always review the output.

Agents

These are LLMs with help. You give them a job description, a task, and tools, like computer programs they can use to generate reports, list accounts, or run calculations. The agent makes a plan, uses its tools, checks if the task is done, and loops until complete. Rob showed Digits’ agent answering “I want to hire 20 software engineers next year. Can I afford to?” with a data-backed response.

Guardrails are critical. Microsoft’s early agent “started asking people on dates in the chat,” Rob noted. “You don’t want your accounting agent dispensing dating advice.” Agents work well for updating schedules, running quality checks, and answering analytical questions, but they’re slow and need careful boundaries.

Predictive models

These got Rob visibly excited, and for good reason. These models take an input and predict an output from known options. When the model sees a $5 Starbucks charge, it considers the client’s location, history, and chart of accounts. For a local client, it’s meals and entertainment. For a New York client in California, it’s travel. A $157 Starbucks charge is probably an event, regardless.

What makes predictive models perfect for transaction categorization is they can’t hallucinate; they only choose from existing options. They’re deterministic (same input, same output), include confidence scores, and run fast and cheap once trained.

Digits built a “layer cake” of predictive models:

Client-level (learns each business)
Firm-level (encodes your best practices)
Global (trained on 180 million transactions worth nearly $1 trillion)
An LLM fallback for completely new transactions

The result was over 97% accuracy, compared to standalone LLMs that plateau below 80%, which is about the same as outsourced bookkeepers.

Document extraction models

These combine OCR with layout-aware language models that understand document structure. Previous tools used Amazon’s Mechanical Turk, which relied on humans manually extracting data and took hours. Modern extraction models work in seconds. Digits’ bank reconciliation automatically pulls PDF statements, matches transactions to the exact spot in the PDF, and generates audit reports.

Data analysis

This model is where Rob pulled the rug out. Financial reporting and analysis is “actually just math. It’s not ML.” Computers have done statistical analysis for decades. Could you build an agent to do it? Sure, but it would be slow, expensive, and probably wrong. “If anyone says their AI does reporting and statistical analysis, please ask them what they’re talking about.”

Here’s how the right models map to your month-end close:

Book transactions: Predictive models (with LLM fallback)
Reconcile statements: Extraction models plus matching algorithms
Update schedules: Agents
Review and correct: Agents with quality checklists
Analyze and report: Statistical analysis plus agents for questions

“Shoehorning an LLM in to solve a problem and just sending a bunch of information is fundamentally incorrect,” Rob emphasized. Each step needs the right model. No single AI approach handles everything.

The 2026 prediction

Understanding model types is just the foundation. The urgency comes from how fast everything is converging.

“Across a large client set in different industry types, it’s highly likely that the month-end close process is looking to be completely automated by the end of 2026,” Rob predicts. Even his “95% automated” hedge probably sounds aggressive. But his logic follows directly from the technology.

If predictive models hit 97% accuracy on transactions, extraction models automate reconciliation in seconds, agents handle schedules and quality control, and statistical analysis covers reporting, then manual work drops to a fraction. Rob’s goal is to see accountants doing “1/20th of the work you’re doing today.”

He acknowledged limits. Construction firms with complex job costing might not hit that threshold. But for firms serving professional services, cash-basis businesses, and straightforward accrual clients, the automation curve is steep.

An AI-native firm will focus on value instead of tedium. Deeper industry expertise. Stronger client relationships. Higher margins. You’re not reviewing every transaction, you’re supervising the system and handling exceptions. Those hours saved on reconciliation become advisory time clients actually value.

But this is also a competitive necessity. “AI won’t replace you. Someone who’s good at using AI is going to,” Rob said, quoting a common warning. And he was direct about the stakes. Firms that don’t adapt will face “a cascading effect on business models” as early adopters pull ahead.

For overwhelmed firms, which Rob acknowledged includes most firms, he offered practical starting points:

Map your processes first. If you use workflow tools like Karbon or Keeper, you’ve probably documented your steps. If not, start there. You can’t identify where AI fits until you know what you’re actually doing.
Start small and low-stakes. Don’t tackle your biggest challenge first. Try drafting emails, testing categorization, or visualizing data. Build your intuition gradually.
Get hands-on with new tools. Rob mentioned being impressed by Claude Opus, which could build HTML dashboards from his data (something he couldn’t do as a non-engineer). The specific tool doesn’t matter; hands-on experience builds judgment.
Know your business before choosing where to start. As Rob put it, “You need to know the details of your business to know where you can start and where the right places to poke and prod are.”

The “wait and see” window is closing. Firms that develop AI literacy now by asking questions about models, data handling, and use cases, will be ready for the rest of 2026 and beyond.

Your next move: Better questions, smaller steps, faster action

Let’s turn Rob and Megan’s insights into actionable takeaways:

Not all AI is equal. Four real model types plus one fake (data analysis) get lumped together. When vendors pitch “AI-powered reporting,” you now know to dig deeper.
Each close step needs a different model. Predictive for transactions. Extraction for reconciliation. Agents for schedules. Statistics for reporting. Anyone claiming one solution does everything deserves scrutiny.
Predictive models beat LLMs for categorization. Layered architectures that learn your clients and firm patterns dramatically outperform chatbots. Bigger isn’t always better.
Ask vendors the hard questions. What model type? Where does data go? Are you training your own models or sending financial data to third parties? This is due diligence.
The tipping point is closer than you think. Whether Rob’s 2026 prediction proves exactly right or directionally right, the trajectory is clear. Understanding these distinctions now positions you to take action.

For those interested in going deeper, Rob mentioned resources like the AI Native Accounting Foundation and the AI-Native Accounting podcast hosted by Kacee Johnson, where industry leaders discuss the latest developments.

The accounting profession is at a real inflection point. Smart firm leaders will develop the literacy to ask smart questions, experiment in the right places, and redirect time from tedium to advisory work clients value.

Rob noted this might be one of the few professions with such clear AI use cases, putting accountants at the forefront of innovation. That’s an opportunity to shape how technology serves the profession, not the other way around.

Watch the on-demand webinar for complete details, including live demonstrations, security architecture specifics, and audience Q&A covering nonprofits, inventory clients, and platform migrations. The future of accounting is being written now. Make sure you’re part of the conversation.

Four Years of Cleanup in Four Hours: Inside the AI Ledger That Learns From Your Work

Earmark Team · January 15, 2026 ·

What if you could stop programming bank rules forever? No more tweaking text strings, adding exceptions, or debugging why “COSTCO WHSE #1234” won’t match your Costco rule. During a recent Earmark Expo webinar, accounting software company Digits demonstrated exactly how that future works, and they achieve 96% automatic categorization accuracy without a single bank rule.

Host David Leary has been watching Digits since before ChatGPT existed. “I remember seeing a pitch deck about Digits, and it was being emailed around on backchannels in the accounting industry,” he recalled. “This pitch deck was super ambitious. At the time, the back channel hallway talk was like, ‘Great, here comes another bank feeds accounting app.’”

Now, years later, that ambitious vision is reality. Rob Hamilton from Digits’ partnerships team showed David and his co-host Blake Oliver what the company calls the world’s first “agentic general ledger,” software built from the ground up with machine learning at its core.

Why Digits Took Six Years to Build

Before diving into the technology, Rob shared the origin story. Digits founder Jeff Seibert sold his previous companies to Box and Twitter. At both companies, he noticed a stark contrast: product and engineering teams had real-time dashboards showing exactly who was on their website and what buttons they clicked. But when he wanted to check if he had a budget for a team event, finance told him to wait 45 days for the books to close.

“As a founder of a company, you’re like, ‘This is crazy. I’m just going to do this event without your approval,’” Rob explained. When Jeff left Twitter, he wanted to use machine learning for good, and accounting emerged as the perfect candidate.

The result took six years to build. “Turns out that it takes a while to build a general ledger from the ground up in the machine learning era,” Rob admitted. But that ground-up approach makes all the difference.

The Three-Layer Intelligence That Replaces Bank Rules

Traditional accounting software makes you act like a programmer. You write rules, define patterns, and hope the software follows instructions. Anyone who’s debugged bank rules knows the frustration.

Digits flips this completely. Instead of you teaching the software through rules, the system learns from your work at three levels.

First, it learns from each specific company. When you connect QuickBooks to Digits, it imports your historical data and trains on how you categorize that company’s transactions. “We actually train on a company level,” Rob explained. “When transactions start coming in, it actually leverages the work you’ve already done within that individual company.”

Second, it learns from your entire firm. When a new vendor appears—say, a coffee shop that just opened—Digits checks if any other client in your firm has seen that vendor. Your work for one client helps all your clients.

Third, it taps into global intelligence. For truly novel transactions, Digits uses its global model trained on every transaction the platform has ever processed.

The payoff is significant. “For September, we’re at a 96% rate of transactions getting booked into Digits that then were subsequently not touched by a human afterwards,” Rob revealed. That’s not just categorized; that’s categorized correctly enough that accountants didn’t change them.

“You’re not editing any rules,” Rob points out, contrasting Digits with traditional systems. “You don’t have to add an extra appendage to pull out the specific Costco transaction. We learn from your behaviors directly inside the product.”

How the System Handles the Other 4%

No AI system is perfect. What matters is how it handles uncertainty. When Digits encounters a transaction it’s unsure about, it doesn’t guess silently. It flags the uncertainty and shows its reasoning.

During the demo, Rob showed a US Patent and Trademark Office transaction where Digits displayed, “I have this as taxes, but I actually think it could be legal.” The system even suggested adding “intangibles” as a new account category for companies still building their chart of accounts.

The learning happens instantly. “We’ve built our architecture to be uniquely quick in the training,” Rob emphasized. “The second we see a similar transaction, it’ll effectively be perfect based on your prior action.”

Quality control is proactive rather than reactive. Each month, Digits flags all new vendors so accountants can verify they’re categorized correctly. It also highlights vendors booked to multiple categories, like Apple transactions split between fixed assets and software subscriptions.

When accountants don’t know what something is, they can ask clients directly within the platform. Questions attach to specific transactions, clients get email notifications, and responses flow back to the same transaction. The AI suggests categorization based on the client’s answer, though accountants confirm before applying.

David appreciated the unified workflow. “Now I don’t have to have five browser tabs open where one browser tab is the report, the transaction is in a new browser tab, and I make the edit and refresh the report in the other browser tab.”

Reconciliation in Minutes, Not Hours

Bank reconciliation should be simple, but when something doesn’t match, like $15 missing from Stripe, the detective work begins. Digits transforms this process entirely.

Statements enter the system three ways. Banks like Wells Fargo send them automatically via API. For others, accountants drag and drop PDFs directly onto the platform. Every Digits account also gets a single email address that accepts any document type, including statements, bills, or receipts, The AI routes them appropriately.

“So one email for all the transactions in a client’s company file,” David noted. “You don’t have special HR email and AP email where you send it to the wrong box and it creates a mess.”

The reconciliation interface shows the bank statement PDF alongside ledger transactions. As you hover over transactions, green boxes highlight the matching line on the statement. David’s reaction captured what every accountant will recognize: “I used to do this with a highlighter and my fingers. I had to find it on both.”

But the real magic is proactive problem detection. Digits identifies specific issues and offers one-click fixes for things like:

Uncleared transactions that should move to next month
Statement items missing from the ledger
Date discrepancies between records and statements

Each issue comes with a resolution button. The system does the detective work; accountants just confirm the fix.

“We had an accountant come in the other day. He was like, ‘I did four years of cleanup in four hours’ because he just linked the bank accounts, dragged all the statements in, and the AI did everything,” Rob says.

Beyond Bookkeeping: Reports Clients Actually Read

With traditional financial reports, only 15% of business owners even open those black-and-white PDF attachments. Digits studied this and found that when firms use visual reporting tools, over 70% of clients actually open and interact with the financials.

The reporting system works like “Google Docs for your finances,” as Rob described it. Accountants can add commentary directly on line items, tag clients with questions, and create visual dashboards that tell the story of the business.

The platform includes built-in bill pay ($0.50 for ACH, $2 for checks) and invoicing. The system automatically recognizes and routes dragged-in documents. Bills queue for payment, receipts match to transactions, and statements trigger reconciliation.

Behind the scenes, AI agents continuously research every vendor, building what Rob called “a dossier” with logos, phone numbers, and descriptions. “This is what your team does when they don’t know what a transaction is. They Google it and find the information.”

What This Means for Your Practice

The shift from rule-based to AI-native software fundamentally changes the accountant’s daily work. Instead of programming rules, you review AI suggestions. Instead of hunting for reconciliation errors, you confirm one-click fixes. Instead of sending reports that get ignored, you create interactive dashboards that clients actually use.

The compound effect is striking. Every correction teaches the system, improving accuracy for that client, your entire firm, and eventually all Digits users. Time savings stack up, allowing firms to shift toward advisory work.

Digits offers a partner program with volume discounts. The standard price is $100 per month per client for full features, with special pricing for tax write-up work. Accounting firms get their own firm account free when joining the partner program.

Rob emphasized that construction and other complex industries might see slightly lower accuracy rates than the 96% average, but the system continuously learns and improves. Features like sales tax support and project tracking are coming soon, while departments and locations tracking are already available.

For firms evaluating new software, the question has shifted from “What rules do I need to create?” to “How well does this system learn?” The four-years-in-four-hours cleanup example shows what’s possible when AI handles the tedious work.

Watch the complete Earmark Expo webinar to see the full demonstration, including reconciliation workflows, client communication tools, and the visual reporting system that gets clients actually engaging with their financials. Whether you’re ready to switch or just want to understand where accounting technology is heading, this demo shows what accounting looks like when bank rules become obsolete.

The Accounting Platform That Achieves 96.5% Automation Reveals How They Did It

Earmark Team · December 22, 2025 ·

“No one’s going to be outcompeted by the AI itself. You are going to be outcompeted by firms that really adopt this aggressively,” warns Jeff Seibert, whose company just hit 96.5% accuracy in automated bookkeeping—something that seemed impossible just a few years ago.

In this milestone 100th episode of the Earmark Podcast, Blake Oliver sits down with Jeff Seibert, co-founder and CEO of Digits, to explore how AI is fundamentally changing the architecture of accounting software. Seibert brings fresh eyes to accounting—he previously led consumer product at Twitter and built Crashlytics (now running on six billion smartphones). His frustration was simple: Why could product teams access real-time analytics while business owners waited weeks for black-and-white spreadsheets?

Founded in 2018, Digits set out to reimagine accounting in the age of machine learning. While traditional software treats transactions as meaningless text in rigid databases, Digits achieves near-perfect automation by treating financial data as interconnected objects that learn from patterns across millions of transactions.

The 30-Year-Old Problem Holding Back Accounting

As Seibert sees it, the fundamental issue facing bookkeeping automation is that every major accounting platform—QuickBooks, Xero, and even NetSuite—runs on relational databases designed 20-30 years ago. These systems treat transactions as simple text entries with no understanding of what they mean.

“QuickBooks is just going to see an Uber transaction as “U-b-e-r”. It just sees text,” Seibert explains. “It doesn’t understand the data, it doesn’t know what Uber actually is.”

This limitation explains why Intuit, with all its resources, has yet to deliver meaningful automation. The answer is architectural. Each QuickBooks company exists in its own isolated database, preventing the software from learning patterns across businesses. The constraints run so deep that QuickBooks still can’t handle having a vendor and customer with the same name—it appears they chose “name” as the primary database key decades ago.

Digits takes a completely different approach using what’s called a vector graph data model. Everything becomes an object—Uber is an object, your expense categories are objects, your bank accounts are objects. Transactions become connections between these objects, creating a web of financial relationships the AI can understand.

This mirrors how large language models (LLMs) work, converting transactions into vector embeddings, essentially plotting them in multi-dimensional space where similar items cluster together. When trained on 170 million transactions representing nearly $1 trillion in business activity, patterns emerge that would be obvious to humans but invisible to traditional software.

“When you have that scale of data and you see how everyone has booked Uber before, you start to see patterns,” Seibert notes. “The model starts learning. If it sees Lyft in your accounting for this client, it then knows how to book Uber.”

How AI Agents Actually Work (Hint: Like Clever Interns)

The accounting world is buzzing about “AI agents,” but what are they really? Seibert explains, “An agent is simply an LLM that you run in a loop. You give it a task, it attempts the task, you ask if it completed it. If not, it continues until it’s done.”

Think of them as clever interns who never get tired. Digits has been running these agents in production since January 2024, primarily for researching unfamiliar transactions.

The system uses three layers of intelligence. First, it checks if this specific client has seen this transaction before. If yes, it books the transaction exactly the same way. Second, if the transaction is new to this client but familiar to the platform, it uses its global model trained across all users. Third, for completely novel transactions, the agent literally Googles them.

“What would you do as an accountant? You would probably Google it,” Seibert explains. “What do our agents do? They literally Google it, research the transaction, build a dossier about it.”

As a result, only 4-5% of transactions now require human review, compared to the 20% that typically slip through even well-maintained rule-based systems. Notably, the system maintains strict confidence thresholds. Any transaction it is unsure about gets flagged for human review. It never guesses when uncertain.

The upcoming reconciliation feature shows how sophisticated these agents have become. The system pulls statements directly from banks or extracts them from PDFs, then matches transactions with pixel-level precision. “You can literally click on the transaction and see it on the statement and vice versa,” Seibert says. This builds trust with accountants who need to see exactly where the numbers come from.

What This Means for Your Firm’s Future

As of August, Digits hit 96.5% accuracy, up from 93.5% in spring. Each percentage point represents thousands of transactions that no longer need human touch. But it begs the question: how do you price services when the work happens automatically?

“If you’re charging purely per hour right now, then automation may make that challenging,” Seibert acknowledges. But forward-thinking firms are already adapting. They’re moving to fixed-fee models for routine work like monthly closes, which become increasingly profitable as automation reduces time investment. Many use a hybrid approach, charging fixed fees for the close, and hourly rates for advisory work.

At a flat $100 per month (with volume discounts for accounting partners), Digits offers predictable pricing that contrasts sharply with QuickBooks’ constant increases. The platform even offers specialized SKUs for ledger-only or reporting-only clients, accommodating diverse practice needs.

The staffing implications are real but not apocalyptic. Junior bookkeeping roles focused on data entry will diminish. But Seibert points out this could make the profession more attractive: “You don’t want to just sit there doing data entry all day long. You want to learn how to advise businesses.”

Seibert recommends firms start small when implementing automated bookkeeping. “Pick one client in your firm and see what you can achieve,” Seibert challenges. Choose a simple, digital-native business like consultants, SaaS companies, or agencies with predictable electronic expenses. Build confidence, then expand to complex cases.

Building Trust Through Transparency

With financial data flowing through AI systems, security is crucial. Digits addresses this with architecture developed at Seibert’s previous companies, where they handled crash data from billions of smartphones.

Everything stays within Digits’ systems; they don’t send raw data to OpenAI or other third parties. All data is encrypted at rest using per-object envelope encryption, where each object has its own encryption key. Even if breached, stealing one key wouldn’t compromise the system.

The platform is SOC 2 Type 2 certified, with complete audit trails showing who changed what and when. You can even grant granular access, like giving your marketing manager visibility into only marketing expenses. “They can see marketing, all the transactions booked to marketing, and nothing else,” Seibert explains.

Importantly, when AI does the work, you can trace exactly what happened. Click on any transaction to see the activity log. This solves the common problem of clients making changes in QuickBooks without anyone knowing.

The Competitive Reality Check

Seibert’s warning deserves repeating: “No one’s going to be outcompeted by the AI itself. You are going to be outcompeted by firms that really adopt this aggressively.”

This isn’t hypothetical. Firms using advanced automation already serve more clients with similar-size teams, offer competitive pricing while maintaining margins, and provide real-time insights that clients increasingly expect.

You don’t have to become a tech expert. Set aside time each month after the close to try new tools. Watch YouTube videos about AI agents (though Oliver warns to avoid the hype channels). Most importantly, maintain healthy skepticism. As Seibert notes about AI doing math, “If it’s not 100% correct, what’s the point?”

Remember, AI agents are like clever interns. They’re eager, overconfident, and need supervision. They excel at tedious, repetitive tasks but need human judgment for nuanced decisions. The goal isn’t to replace accountants but to eliminate the work accountants wish they didn’t have to do.

Taking the First Step

Thoughtfully evaluate how these innovations can augment your practice. Start with one simple client. See what 96.5% automation actually feels like. Build confidence, then expand gradually.

Listen to the full episode to hear Seibert’s complete vision and practical guidance on everything from selecting pilot clients to restructuring pricing models. The tools to eliminate tedium while amplifying expertise aren’t coming; they’re here, proven, and improving rapidly. How quickly and thoughtfully can you integrate it?