Khosla Ventures is making a bet that feels both familiar and risky: it’s putting $10 million behind Ian Crosby, the founder of Bench—an accounting startup that, according to reporting, ultimately imploded—and asking him to do something even more ambitious this time. Crosby’s new company, Synthetic, is building what it describes as a fully autonomous AI bookkeeping service for other startups. In plain terms, the pitch is not just “AI helps with bookkeeping,” but “AI runs the bookkeeping.”
That distinction matters, because bookkeeping is one of those business functions that looks straightforward until you try to automate it end-to-end. It’s not only about categorizing transactions. It’s about reconciling messy reality: incomplete data, inconsistent vendor descriptions, timing mismatches between bank feeds and invoices, refunds that arrive weeks later, chargebacks, sales tax edge cases, and the constant need to produce outputs that accountants and auditors can trust. A system that merely suggests categories can be forgiven for being imperfect. A system that claims autonomy has to be right often enough—and explain itself clearly enough—that finance teams (or founders acting as their own finance teams) don’t lose confidence.
Synthetic’s core promise is that it will handle the recurring work of bookkeeping with minimal manual intervention. For early-stage companies, that’s a compelling value proposition. Startups don’t just need “accounting.” They need clean books that support decisions: runway calculations, burn-rate tracking, fundraising narratives, and the ability to answer basic questions quickly—What did we spend on? What did we sell? Are we growing profitably or just growing revenue? When bookkeeping is delayed or inaccurate, those answers become guesswork.
But the deeper question is whether autonomy is achievable in a domain where correctness is non-negotiable. Khosla’s investment suggests the firm believes Crosby has a path to make it work—not by pretending bookkeeping is simple, but by designing the system around the realities of how companies actually run.
The founder story adds another layer. Crosby previously led Bench, which many people in the startup ecosystem associate with the broader wave of “outsourced accounting” and “software-assisted finance operations.” Bench’s trajectory, as described in reports, ended badly. That kind of outcome can be interpreted in two opposite ways. One interpretation is that the market is too hard for automation and too expensive to serve at scale. The other is that the first attempt revealed exactly where the bottlenecks are—data quality, operational coverage, unit economics, or the gap between what customers expect and what systems can reliably deliver.
Synthetic appears to be built around the second interpretation: that the problem isn’t that bookkeeping can’t be improved; it’s that the earlier approach didn’t fully solve the operational and reliability challenges required to make automation feel like a product rather than a prototype.
So what does “fully autonomous” mean in practice? The most important part is not the model itself—it’s the workflow. Bookkeeping is a chain of steps, and autonomy means the system can move through that chain without constant human correction. That includes ingestion (pulling transactions from banks and payment processors), normalization (turning raw strings and amounts into structured events), classification (mapping events to the right accounts), reconciliation (ensuring the ledger matches source-of-truth statements), and then producing outputs that align with accounting standards and the expectations of downstream users.
In a typical startup, the inputs are rarely tidy. A single vendor might appear under multiple names. A payment processor might split charges and fees across different lines. Subscriptions might change tiers mid-cycle. Expenses might be reimbursed later. Even the “simple” act of categorizing a transaction can require context: was that expense truly an operating cost, or is it a capitalizable item? Is that refund a reversal of revenue, or is it a separate event? Does that charge include taxes that should be tracked separately? Autonomy means Synthetic must either infer that context reliably or detect when it can’t and then handle the exception in a way that doesn’t break the ledger.
This is where AI bookkeeping becomes less about pattern recognition and more about systems engineering. A model can learn to categorize transactions, but it can also hallucinate or overconfidently misclassify. The difference between a helpful assistant and an autonomous operator is the presence of guardrails: validation checks, reconciliation logic, anomaly detection, and escalation paths. If Synthetic is truly autonomous, it likely needs to know when it’s uncertain and what to do next—whether that means requesting missing information, reprocessing with additional data, or flagging items for review in a way that still preserves the “mostly hands-off” experience.
There’s also the compliance dimension. Bookkeeping isn’t just internal bookkeeping; it’s often tied to tax filings, payroll coordination, and audit readiness. Even if Synthetic is not directly filing taxes, it has to produce records that won’t create downstream chaos. That means the system must be consistent in how it handles categories, dates, and adjustments. It also means it must maintain an audit trail: what it did, why it did it, and how it corrected errors. Autonomy without traceability is a recipe for silent failure—especially in finance, where the cost of being wrong is often discovered months later.
Khosla’s involvement signals that the firm sees a credible path to solving these issues. Khosla tends to invest in companies that can build durable technology advantages, not just services. A $10 million check suggests confidence that Synthetic can become more than a “managed AI” offering. The goal, presumably, is to create a repeatable system that can onboard a startup, ingest its financial data, keep the books current, and reduce the need for ongoing human labor.
That’s a big deal for startups, because finance headcount is expensive and hard to justify early. Many companies start with spreadsheets and ad hoc tools, then graduate to outsourced bookkeeping, and eventually hire a finance leader once they have enough complexity. If Synthetic works as promised, it could compress that timeline. Instead of paying for a human-led process that scales slowly, startups could pay for a system that scales automatically—at least in theory.
But scaling is where the real test begins. Bookkeeping is high-volume and detail-sensitive. The unit economics depend on how much human time is required per customer, even if the product is “autonomous.” If autonomy is achieved by adding more and more exceptions that require manual intervention, the cost structure may not improve. If autonomy is achieved by building robust workflows that handle the majority of cases correctly, then the economics can improve dramatically.
A unique angle in this story is how Synthetic positions autonomy as a product experience rather than a technical demo. Many AI startups can generate plausible outputs in controlled settings. The challenge is turning that into a reliable service that behaves well under the messy conditions of real businesses. That includes handling onboarding variability: some startups have clean chart-of-accounts setups; others don’t. Some have consistent naming conventions; others have years of inconsistent vendor descriptions. Some use accounting software that integrates smoothly with bank feeds; others rely on exports and manual uploads. An autonomous system has to adapt to these differences without requiring a finance expert to configure everything.
If Synthetic succeeds, it could also change how startups think about finance operations. Instead of treating bookkeeping as a periodic chore—something you do monthly or quarterly—it becomes a continuous background process. That would allow founders to see trends sooner, catch anomalies earlier, and respond to cash flow changes with more confidence. It could also improve fundraising readiness. Investors often ask for financial statements and metrics that require clean categorization and reconciliation. If Synthetic keeps the ledger accurate continuously, the “fundraising scramble” becomes less chaotic.
There’s also a strategic implication for the broader fintech and accounting ecosystem. Traditional accounting services have long relied on a combination of expertise and labor. AI threatens the labor component, but it doesn’t eliminate the need for accountability. If Synthetic can deliver trustworthy books, it may force incumbents to rethink their value proposition: not “we do bookkeeping,” but “we provide oversight, compliance assurance, and exception handling on top of automated systems.” In other words, humans may shift from doing the work to supervising the system.
That shift is not trivial. It requires new operational models, new pricing structures, and new ways to measure performance. It also requires trust. Startups will want to know: if the system makes a mistake, how is it detected? How is it corrected? Who is responsible? What happens when the system encounters a scenario it can’t interpret? Autonomy is only valuable if it’s paired with accountability.
Crosby’s past experience with Bench may influence how Synthetic approaches these questions. Bench’s model, as widely discussed in the market, involved a blend of software and human support. If Bench struggled, it may have been due to the difficulty of maintaining quality at scale, the cost of service delivery, or the mismatch between customer expectations and what the service could reliably deliver. Synthetic’s autonomy claim suggests an attempt to reduce reliance on human labor while still meeting quality requirements.
Still, the market will judge Synthetic on outcomes, not promises. The most telling indicators will be accuracy rates, reconciliation success, time-to-close for monthly books, and how often the system needs human intervention. Another indicator will be customer retention: if startups adopt Synthetic and then later revert to manual processes or human-led bookkeeping, that’s a sign the autonomy wasn’t real. If startups stay and expand usage, it suggests the system is becoming embedded in their operations.
There’s also the question of integration. Bookkeeping doesn’t live in isolation. Startups use payment processors, expense management tools, payroll providers, invoicing systems, and sometimes inventory platforms. The more Synthetic can integrate seamlessly with the tools startups already use, the more likely it can maintain autonomy. Integration reduces friction and improves data completeness, which in turn improves classification accuracy and reconciliation reliability.
Synthetic’s focus on startups is particularly interesting because startups are both simpler and more chaotic than mature businesses. They often have fewer transactions than large enterprises, which can make automation easier. But they also have rapid changes: new products, new pricing models, shifting revenue streams, and frequent changes