Box

In-depth analysis and frontier reasoning on challenging document analysis with Grok 4.5

We dropped a 30-page office lease into Box AI and asked @grok 4.5 one simple question: what would it actually cost us to walk away after three years? The lease is written to make you think the answer is easy: six months' rent (about $267K) stated right in the termination section. But the real cost is scattered in pieces across the whole document. The termination clause quietly points to a defined term that, when you chase it down, pulls in leftover build-out costs and broker fees from two other sections. The "free rent" you were given at signing is really taken back by a separate clause if you leave early. And there's a cleanup-and-restore obligation tucked into an exhibit at the very back. You only get the true number if you find all five and add them up — across 33 sections and 10 attachments. Grok 4.5 nailed it: ~$830K, about 3x the number the lease wants you to see, and it showed exactly where each piece came from. Then it kept going, pointing out the costs that have no cap, the auto-renewal trap that kicks in if you miss a deadline, and which terms to push back on before signing. That's the difference between skimming a contract and truly reading one and it's the kind of careful, connect-the-dots work over long documents that real business runs on. And this is the power of the latest reasoning models right now on working with unstructured enterprise data.

Full conversation: youtube.com/watch?v=-slXAX…

"The data that we've been working with as people now becomes extremely useful in a world of agents. You're having this rebirth of the file system." @levie on Navigators with @pk_iv on why unstructured data, the 90% of enterprise data that has always been the hardest to automate, is now the biggest unlock for AI agents. And why the real AI conversation is a context conversation.

Register today → boxworks.box.com/series/boxwork…

The leaders shaping the future of AI aren't waiting to see what happens next. They're at BoxWorks 2026. Come to learn: 🔹 The Content + AI strategies that are actually working at scale 🔹 How to navigate emerging AI models, frameworks & governance 🔹 How to make your content infrastructure AI-ready — now Walk in with questions. Walk out with a playbook.

The reason I have an unhealthy obsession with AI right now is because I've spent my entire professional life on essentially one problem: how do you increase the value of content in the enterprise. How do you secure it, how do you collaborate on it, how do you govern it, and how to integrate it across all your applications. But there's been one glaring issue that we've dealt with since the founding of Box. We could never really process information at scale in any real automated way. There have been many attempts at this problem (often in the search space), but nothing that really fundamentally transformed what you can do with enterprise knowledge. For years the primary kind of data that we could query, analyze, and process with computers was structured data. This meant anything you could shove into a database you could understand with computers - your CRM, ERP, product analytics, HR, and other data. But all of the unstructured data that powers our daily knowledge work - marketing assets, contracts, financial documents, medical research, engineering documentation - was only valuable when a human was operating on it. There was just simply no real way to apply automation at scale to any of this data, which meant all knowledge work was largely rate limited by our ability to process information ourselves, often manually. AI models have obviously dramatically changed this reality. And the past couple weeks perfectly highlight this incredible progress. GPT-5.6, Fable 5, Grok 4.5, Muse Spark 1.1, and a leading array of open weights models are all showing incredible advancements on working with unstructured data. The inherent broad intelligence, reasoning, math, and coding skills in these models, combined with deep domain expertise trained into them across finance, legal, healthcare, life sciences, and other critical fields, means that we're able to completely change what we can do with this unstructured data at scale. What this unlocks is the ability to ask insanely complex questions of your data that were never before possible, and let agents just run on for minutes or hours across these data sets to accelerate knowledge work. And it's not just about automating the work that we already do. While this is highly valuable, it wouldn't be particularly transformative. What's exciting is that you can now throw compute at unstructured data problems that wouldn't have been possible before. Analyze every risk on my contracts, do due diligence more deeply on a prospective investment or acquisition, look through all past client interactions in an industry to find best practices to replicate, comb through life sciences research or clinical trial data for new insights, and on and on. So that's why we're insanely excited about what AI Agents can now do with content on Box.

GPT-5.6 Sol is a breakthrough in complex reasoning and data analysis. Here, it analyzes hundreds of pages across a lending deal, reconciles terms across agreements, financials, diligence, collateral, and risk materials, flags issues, and saves a source-cited report to Box.

Sign up for free dev account here: developer.box.com/guides/tutoria…

Embedding a document viewer in your app sounds simple. Keeping it secure is where things get complicated. This demo shows how to build a secure in-app document review experience using Box View. Users can preview, annotate, and collaborate on files directly inside your application, with Box permissions and access controls enforced throughout. No file downloads. No data leaving your governed environment. Watch here.👇

If I could only follow one person on X, it would be @levie, CEO of Box. We sat down to speak about why the REAL economy hasn't felt AI yet, the "model overhang" between capability and adoption, and what founders should actually build now that software is cheap. 00:00 - Cold open: agents use data like people do 01:53 - Is Redwood City the new AI epicenter? 05:18 - 90% of enterprise data is unstructured and was never automatable 07:46 - The model overhang: capability is outpacing adoption 10:35 - Why coding got the fastest AI takeoff 16:24 - Inside Box's agent-first rebuild 19:15 - Headless SaaS: agents will outnumber humans 100:1 25:41 - Model routing + why the applied layer wins, even bitter-lesson-pilled 27:45 - GTM is the new moat Checkout Navigators, our new pod on Youtube, Spotify, and Apple Podcasts!

.@OpenAI just announced ChatGPT Work, and you can use these new features with the Box plugin today. The Box plugin, powered by the Box MCP server, is available in the new ChatGPT Plugin directory. That means that all of your workflows in ChatGPT Work - across Finance, Legal or Ops - can be grounded in your governed enterprise content in Box. In this demo, the agent reads a deal workspace, drafts the credit memo from the source documents, and saves it back to Box, where classification and approval routing take over.

GPT-5.6 is now out. We've been evaluating the model family on the Box AI Complex Work eval, which tests the model with the Box AI Agent on a variety of extremely hard tasks using enterprise document sets. Sol is a big step up from GPT-5.5, especially on complex data-oriented tasks that require deep reasoning and analysis, and the gains concentrate exactly where enterprise work is hardest. Here are a few examples that we saw across our tests: * Financial Services (76% vs 71%): On a multi-year projection, Sol anchored to the correct opening balance sheet date rather than assuming a clean January 1 start, then carried revenue, earnings, and interest through to the right figures year over year where one early wrong assumption compounds through every downstream cell. * Healthcare (58% vs 46%): On a critical-care case review, Sol identified the correct diagnosis and intervention and avoided the dangerous misstep of ordering imaging before the time-critical procedure, a trap GPT-5.5 walked into. * Public Sector (74% vs 63%): Handed a class's raw gradebook and a new grading directive, Sol mapped each assignment to the right weight bucket, treating homework as zero-weight practice per the directive. It recomputed every student's grade to within a tenth of a percent, where GPT-5.5 drifted partway through. * Life Sciences (60% vs 51%): Across four separate compound datasets, Sol intersected the ranked target lists exactly (case-sensitive, no shortcuts) to find the biological targets common to all four, catching the shared targets GPT-5.5 missed. Sol reasons from the source definitions and checks the documents rather than taking them at face value and it's most reliable exactly where the numbers drive real decisions. This will be huge for enterprise agents using unstructured enterprise data. GPT-5.6 will be available to customers shortly within the Box AI Studio for building custom agents with.

GPT-5.6 from @OpenAI is a three-tier model family designed to match the right capability to the right workflow. We evaluated models Sol, Terra, and Luna on enterprise document tasks: → Sol overall: 64.4% vs GPT 5.5's 63.3% → Healthcare: +12.7pp → Public Sector: +11.0pp → Life Sciences: +9.0pp → Data Analysis: +7.2pp Sol gets the hard, high-value work right. Terra and Luna bring near-flagship quality at 16-19% faster speeds for high-volume production workflows. Read the full evaluation here. 👇 blog.box.com/how-gpt-56-han…

Making every app and agent part of the conversation. 🤝 25+ partners, @Atlassian, @Box, @Docusign, @LucidSoftware and counting, have built MCP-native apps that connect agents directly into Slack channels and Slackbot.

We evaluated Muse Spark 1.1 from @AIatMeta Superintelligence Lab on Box’s Complex Work Eval. It was competitive with top-tier models, especially on structured data analysis and report drafting. The takeaway: models matter, but governed content is what makes enterprise AI useful at scale. Read more.👇 blog.box.com/muse-spark-11-…

Grok 4.5 reviewed a full Credit & Security Agreement stored in Box — the kind of dense, multi-section facility document that typically requires significant counsel time. @Grok 4.5 used Box MCP to access the file securely, extract key terms across the agreement, identify potential conflicts with existing debt covenants, and compile a summary of items for counsel to review, and finally saved the memo back to the same folder. As frontier models keep leveling up, they are unlocking more opportunities for companies to automate and unlock their enterprise content. Check-out the generated report here: app.box.com/s/zfwfud9ojwbv…

And learn more from the announcement: blogs.nvidia.com/blog/nemotron-…

Box Agent uses @LangChain’s Deep Agents harness to bring specialized agents into the enterprise content platform.w @NVIDIA + LangChain’s work with Nemotron 3 Ultra reinforces where AI is headed: open, tunable agent systems embedded where work happens. Box makes them enterprise-ready. Learn more about using Box and LangChain together in our upcoming webinar. events.box.com/box-webinars/b…