Box

Most HR compliance problems aren't document problems, they're infrastructure problems. The documents exist, but what's missing is a content layer that connects them to the questions being asked of them. Bolt had 170,000 employee documents sitting in Box, mostly unaudited and filed under inconsistent naming conventions. Manually reviewing them would have taken approximately two years, on top of the team's existing day-to-day workload. A compliance question as straightforward as how many remote workers Bolt had in a given country could take weeks to answer, because the answer was buried somewhere across those 170,000 files. What Bolt built was a governed content pipeline that routes and renames files automatically, tags metadata on ingestion, and surfaces everything in a queryable dashboard. The same compliance question now takes minutes, and the team that once faced a two-year backlog is now focused on oversight and quality review rather than manual processing. Link to the full story in the comments.

We evaluated Claude Haiku 5.5 from Anthropic on our Complex Work Eval, an agentic benchmark that hands a model a realistic bundle of enterprise source material and asks for the finished deliverable: retrieve the right files, reason across spreadsheets, PDFs and decks, and write the output a knowledge worker would hand over. Claude Haiku 5.5 scored 60% against 49% for Haiku 4.5, an 11-point gain, and leads in all four modes of analytical work. What stood out is where the gains concentrated. The largest margin, 19 points, came on drafting reports from source data, the mode that most directly tests multi-step numerical derivation. Three examples from the eval: On an equipment overhaul cost report, the workbook carried a summary row alongside the individual line items, so adding up the column naively double-counts the entire total. Claude Haiku 5.5 recognized the duplication and excluded it; the previous generation included it and separately mis-added a set of figures it had transcribed correctly. 88% vs 49%. On a set of projected financial statements for a company running on an overdraft, interest expense depends on the overdraft balance, which depends on the tax payment, which depends on profit after interest, a circular dependency that has to be iterated to convergence. Claude Haiku 5.5 resolved it and produced internally consistent statements. 75% vs 33%. On a diagnostic product's per-site performance data, the previous generation transcribed the site figures correctly but computed the dispersion statistics on the wrong scale, off by an order of magnitude, and every threshold judgment downstream inherited the error. Claude Haiku 5.5's figures reconcile against the source data. 81% vs 37%. It also reached those answers with about 27% fewer tokens, 36% fewer tool calls, and in roughly half the latency, with deliverables of essentially the same length. Higher accuracy and lower execution overhead usually trade off against each other on this eval; here they moved together.

Last week, OpenAI launched dots, and we asked one to get us ready for a closing call on a $420 million acquisition. The dot uses its browser to find and open the latest closing-readiness brief and equity amendment in Box, then gives us a short briefing for the call. The funding gap is covered, but a required customer consent is still missing. It highlights what needs follow-up and what to raise with counsel. Connect Box to ChatGPT to give your dot the business context to get work done.

Everyone is building AI agents. Almost nobody has solved the content problem. BoxWorks 2026 is where that gets answered. San Francisco, November 5 to 6, and I'm excited about this one. Here is why I'm going. - 96% of organizations say AI agents need access to company-specific content. Only 36% have connected agents to trusted content across many use cases. That gap is where most enterprise AI projects get stuck right now. - Box sits right in the middle of it. Your business lives in your files, your contracts, your policies, your research. Agents only become useful when they can work with that content safely. The keynote lineup shows how big this conversation has become. - Aaron Levie, Co-Founder and CEO of Box, will share where enterprise content and AI are heading. - Jensen Huang, Founder and CEO of NVIDIA, will break down how the world's leading organizations are winning with AI. - Lip-Bu Tan, CEO of Intel, will explore how to maximize the power of AI agents. Three leaders across software, chips and compute on one stage. That does not happen often. Beyond the keynotes, I'm looking forward to: 1. Digging deeper into AI-powered security and governance, since only 34% of organizations have formal standards for how agents access company data 2. Learning how automated workflows can turn file-heavy manual processes into something faster, with humans still reviewing the exceptions 3. Exploring how open standards like MCP let teams connect any model or agent to their content 4. Hands-on demos and real conversations with people putting AI into practice Why this matters: every company wants agents. Very few have figured out how to give them the right context for success without risking sensitive data. The answer to that question will decide who moves fast and who stays stuck in pilots. I'll be there with The Ravit Show, talking to founders, CIOs and practitioners about what is actually working. If you are going, message me. I would love to meet in person. If you have not registered yet, use code RavitShow50 for 50% off your registration: https://lnkd.in/g25XwJYR See you in San Francisco. #boxworks2026 #box #boxpartner #enterpriseai #aiagents #contentmanagement #aigovernance #sanfrancisco #theravitshow

Enterprise search has always had a signal problem. The internet gets billions of searches a day to learn from. Your company's internal search gets a fraction of that. But the bigger problem is that keyword search was never designed to understand what you meant. It was designed to match what you typed. Those are two very different things. The shift to semantic search changes that relationship entirely. The same technology that allows AI to understand language can now convert a question into meaning, not just words, and find content that matches the intent behind the search rather than the literal string of characters. Our CTO Ben Kus says we're entering a golden era of enterprise search, and that the most interesting development isn't semantic search itself. It's what happens when you put an agent on top of it. Instead of an employee searching, refining, and searching again, an agent handles the retrieval entirely. You ask a question once and get the answer. For enterprises where the right answer is already somewhere in your content, the infrastructure to find it is finally catching up.

Extracting critical business data at scale is only useful if teams can trust the results. Confidence levels and bounding boxes in Box Extract help teams quickly identify which extracted values may need review and verify them directly against the source content, reducing the effort required to validate extraction results. See it in action with a clinical trial protocol use case. 👇

Blocking downloads isn't enough anymore. AI agents don't need to download a file to read it. If a connected integration can preview a document, it can comprehend the content and reproduce what's inside. From a data loss prevention perspective, an LLM reading a file is functionally the same as downloading it, and most enterprise security policies weren't built with that distinction in mind. The controls that kept content safe from human misuse were designed around a click, a download, and an audit log. Agents don't click. They call APIs, read text representations, and answer questions on behalf of a user in seconds. Governing what an agent can read is now as important as governing what it can download. The diagram walks through the full enforcement path.

Customers are facing a critical transition with AI, and a powerful model alone won't transform their business. They need a unified, governed content management platform, integrations with their existing systems, custom applications and workflows, and expert guidance through the change. BoxWorks brings Box and our partners together to deliver all of it. On November 4, we're kicking things off with the Box Partner Summit, a dedicated day designed to deepen our partnerships, accelerate joint go-to-market impact, and look at where Content + AI is headed. Partners will walk away with roadmap insights, executive relationships, and the tools to help your customers turn their content into a real competitive advantage with AI. The day will kick off with our partner keynote and customer impact showcase, followed by our partner awards, and a reception to close out the day. Then on November 5 and 6, our partners take over the expo hall. This is where you'll see the ecosystem in action and experience firsthand how Box and our partners are driving the future of enterprise AI together. To our partners, thank you for building this with us. To everyone else, we can't wait to see you there. Register now: https://lnkd.in/giYCJkuK

Box is now a supported integration in HubSpot's Breeze Assistant. Drop a signed MSA in Box, ask Breeze for the renewal deadline and key terms, and it creates the follow-up task in HubSpot automatically, all without switching tabs or lifting a finger. Your content and your CRM, finally working together. Watch our demo.

Most teams are still manually sorting files that could route themselves. A question from our community said it best: "Can we automatically tag metadata and sort files into the right client folder without a third-party tool?" The answer is yes, using a simple Extract → Branch → Act pattern that most modern workflow platforms already support. What's a process your team is still doing by hand that you suspect could be automated?