A software engineer at a mid-sized logistics firm recently discovered something unsettling about the invoice processing tool his team had built. Run the same PDF through their AI-powered extraction pipeline on Monday, and it pulls $1,247.83 from line item three. Run it again on Tuesday—same document, same code, same model—and suddenly it's $1,247.38. The difference might be pennies, but when you're processing thousands of invoices a week, those pennies compound into a reconciliation nightmare.
Welcome to the consistency problem that nobody wants to talk about during AI earnings calls.
This particular flavor of friction—the gap between what probabilistic language models can do and what business workflows need them to do—has cracked open a market opportunity for Interfaze, a five-person outfit listed in Y Combinator's Spring 2026 batch. While the rest of the industry chases ever-larger foundation models, founders Yoeven D Khemlani and Harsha Vardhan Khurdula are placing a different bet entirely. For unglamorous tasks like optical character recognition, web scraping, and pulling structured data from messy documents, they argue, determinism trumps general intelligence every time.
The pitch sounds almost quaint in an era of agents that can supposedly write code and book flights. But talk to developers building production systems, and you'll hear a different story—one where the probabilistic nature of large language models isn't a feature but a bug.
The Temperature Problem
Here's the technical rub. Modern LLMs generate text through probabilistic sampling: parameters like temperature and top-k that give models their creative range also bake in run-to-run variability. Even when you lock down the output format with structured modes and constrained decoding—guarantees that you'll get valid JSON back—the actual values inside that JSON can still drift.
Research from early 2026 examining consistency-accuracy correlation in entity extraction put numbers to what developers had been grumbling about: hard-prompted LLMs still exhibit what researchers called "value-grounding errors" despite format guarantees. You get your nicely shaped JSON object back, sure. But invoice total might be wrong.
Microsoft Research has been exploring this problem with something they called verified-speculation decoding, published as LLM-42. The approach uses verify-rollback loops to enforce determinism—essentially checking the model's work and forcing it to try again if answers diverge. Clever engineering, though it comes with latency costs.
Another research group took a more pointed stance. In a January preprint, they described what they termed the "plausibility trap"—using expensive probabilistic engines for simple deterministic tasks like OCR. Their benchmarking found 6.5× latency penalties and what they delicately called "sycophancy risks," the tendency of models to generate plausible-sounding but incorrect outputs. Why use a sledgehammer, they asked, when the task calls for a scalpel?
That question resonates beyond academic circles. Gartner projects AI agent software spending will reach $206.5 billion this year, climbing to $376.3 billion in 2027—staggering numbers that reflect genuine enterprise adoption. But a McKinsey survey fielded late last year paints a more textured picture. While 23% of organizations reported scaling an agentic AI system somewhere in their operations, the deployment bottlenecks often centered on reliability and consistency rather than raw capability. The tech works, in other words, until it doesn't.
Stacking the Deck Differently
Interfaze's architecture, laid out in a February arXiv paper by Khurdula and co-founders Vineet Agarwal and Khemlani, takes a heterogeneous approach. They stack specialized deep neural networks and small language models for what they call perception tasks—OCR, automatic speech recognition, the unglamorous plumbing of data extraction. A context-construction layer handles web crawling and parsing across pages, code repositories, and PDFs. An action layer provides tools: browsing capabilities, retrieval, code execution sandboxes, a headless browser.
Then comes the trick. A thin controller routes these specialized tools and forwards distilled context to whatever LLM the user wants—OpenAI, Anthropic, doesn't matter. The general-purpose model handles reasoning and final response, but it never touches the raw OCR or scraping work. That gets delegated to components built for deterministic outputs.
The team published benchmark numbers in February: MMLU-Pro at 83.6%, MMLU at 91.4%, GPQA-Diamond hitting 81.3%. The company reports 98–99% structured-output accuracy and sub-five-second response times for specialized tasks on their website, though these are vendor-reported figures without independent verification. Take them with appropriate skepticism, as you would any startup's self-benchmarking.
In an April blog post titled "CNN + VLM > VLM"—a headline that probably made their marketing team wince—the founders argued for what they call model-level fusion. The idea: embed a convolutional neural network OCR head directly into the decoder alongside a general vision-language model encoder. You get first-class bounding boxes and confidence scores plus structured outputs in one pass, rather than the multi-step tool-calling sequences that add latency and create handoff points where things break.
By mid-April, the company noted customers were using Interfaze as a deterministic component inside broader workflows, essentially outsourcing the reliability-critical pieces while keeping their existing LLM providers for everything else.
The technical claim aligns with research gaining traction this year. A March paper presented at EACL's industry track—dryly titled "OCR or Not?"—benchmarked large-scale document extraction and questioned whether multimodal LLM-only pipelines could match OCR-plus-MLLM hybrids. Multiple studies suggested hybrid approaches often outperformed pure-LLM methods on accuracy, determinism, and cost. Particularly, the researchers noted, when downstream workflows need bounding boxes and confidence scores that LLMs struggle to provide reliably.
Incumbents Aren't Sleeping
Interfaze enters a market where established players are also rethinking pure-LLM strategies, which should tell you something about the broader shift.
Hyperscience positioned its Spring release around what it called a move "From IDP to Intelligent Inference," introducing model-agnostic orchestration flows that route document workloads to whichever model fits best. Rossum announced Aurora, a proprietary LLM built specifically for transactional documents, claiming 92.6% accuracy after training on just 20 documents in a case study with Adyen. Nanonets released OCR-3 in April, explicitly targeting what they called "the agentic stack."
The intelligent document processing market sits at an estimated $3.17 billion this year, projected to reach $7.18 billion by 2031, according to Mordor Intelligence's February update. The broader OCR market hit $15.8 billion last year and is forecast to reach $48.1 billion by 2034, per IMARC's April report. These aren't startup-scale numbers—this is infrastructure.
Enterprise adoption signals are mixed but directional. That McKinsey survey from late last year found agentic AI moving from experiments to scaled deployment across roughly 500 organizations. Yet Forrester's predictions for this year, published last October, framed 2026 as an AI "reckoning" marked by ROI scrutiny and stronger demand for auditable, deterministic outputs. Gartner's January forecast tagged the year as entering the "Trough of Disillusionment," with enterprises favoring AI sold by trusted incumbents over moonshot promises.
That backdrop might actually favor specialized approaches. Perhaps counterintuitively.
The Intelligent Document Processing benchmark released by Businesswaretech in January and third-party leaderboards surfacing in recent months point to a market increasingly willing to evaluate task-specific performance rather than general capability alone. Good at everything means mediocre at the thing I actually need—a sentiment you hear more frequently in enterprise IT conversations these days.
Beyond Documents
The determinism question extends past invoice processing into web automation and browser agents, where the stakes can be higher. OpenAI's computer-use API documentation and Anthropic's computer-use research preview—reportedly available on macOS as of March—signal frontier lab interest. Infrastructure providers like Browserbase released Stagehand v3.6.3 on March 31, with action caching and semantic trees designed to improve reproducibility. Playwright's Model Context Protocol integrations emerged around the same timeframe.
Notice the pattern. General-purpose models handle reasoning and planning, but specialized components handle perception and action with guarantees. Microsoft Research's verified speculation work, grammar-constrained decoding libraries like XGrammar presented at MLSys last year, dynamic model routers from IBM and Microsoft—all point toward inference architectures that route tasks to appropriate engines rather than forcing everything through one model.
Interfaze positions itself as OpenAI-compatible, a practical decision that lets developers swap in their API for structured tasks while keeping existing LLM providers for reasoning. The April updates added GUI detection for computer-use workflows, faster reasoning, and scraping improvements. The company released what it calls a Structured Output Benchmark on April 28, scoring models on value accuracy, JSON pass rate, type safety, and other metrics—though as a vendor-run benchmark, independent reproduction would strengthen the claims considerably.
Regulatory context adds a layer of urgency that isn't always visible in technical discussions. The EU AI Act's general applicability date lands August 2, with enforcement powers for general-purpose AI also kicking in. California's CPPA launched the DROP data-deletion platform in January, with data broker compliance required by August 1. Deterministic, auditable outputs become more defensible under governance regimes that demand transparency and explainability. Legal departments care about this stuff, even if engineering teams don't yet.
The Wedge Question
The question for developers building automation workflows isn't whether LLMs are useful—that debate is settled. It's whether every task needs one.
Interfaze's bet is that document extraction, web scraping, and classification belong to specialized models that guarantee consistency, with LLMs reserved for the reasoning layer that needs flexibility. Whether that architecture becomes standard or remains a niche tool depends on how much the industry ultimately values reliability over generality.
The answer probably won't come from research papers or benchmark leaderboards. It'll come from production deployments over the next year, from developers wrestling with those Tuesday invoice discrepancies and deciding whether probabilistic outputs are acceptable trade-offs or deal-breakers. From CFOs asking why the AI-powered accounts payable system keeps flagging phantom discrepancies. From compliance officers demanding audit trails that don't include "the model hallucinated this field" as an explanation.
The boring stuff, in other words. Which is often where the real money gets made.
