The feature that enticed us to finally pay for AI
Not Agents…not benchmarks…Excel
Recently, our company committed to a paid Team subscription to Claude. Prior to that, many of the company's employees were using different AI tools, Ollama and LM Studio for the tinkerers, ChatGPT for the curious, Claude Code for Developers, and one person even trying to use Microsoft Copilot (it hasn't gone well). Ad hoc models, ad hoc tool use, low adoption rate, no signs of productivity or quality improvements.
The tipping point for our decision wasn't recent SWE bench scores, or how Opus 4.5 performed on Humanities Last Exam, nor the emergence of ever-improving agentic capabilities in Claude code or Cowork, it was meaningful and useful integration with good ol' Excel.
I realize this is anticlimactic. In a world where AI companies are racing to achieve AGI and researchers are publishing papers about world models, what actually mattered to our business was the ability of our non-coding employees to immediately see value and be able to recognize the use cases and opportunities for improving the way they work using the exact software, skills, and vernacular that they currently use in their work. Accessibility and relevance are what unlocked widespread AI adoption in our small company.
The importance of Excel in industry cannot be overstated. I regularly assert that it's the most successful business application ever created. Excel is the universal adapter. It doesn't matter what industry you're in, what role you hold, or what size your company is. Everyone has Excel. Everyone knows Excel. Everyone has workflows living in Excel that they've built themselves, understand intimately, and know could be better.
I know someone who used to work at SAP—a company that exists to help organizations run their businesses with enterprise software. Inside SAP, the actual operation runs on Excel. Roadmaps, financials, reporting, projects: all Excel. The company that sells you software to replace spreadsheets runs on spreadsheets.
For developers, Claude Code and agentic coding tools provide a similar 'unlocking', meeting coders where they already work and speaking their language. Excel does the same thing for everyone else. And the number of users actively using Excel in their day-to-day jobs far outpaces the number of coders (I have no actual data to back up that claim, but I'm sure AI would be happy to fabricate some data to support it).
No translation required
As an AI/NLP enthusiast/alarmist, I'm astounded at the pace of development and innovation in the past three years, but in terms of applications of the technology, the industry is awash with content that highlights use cases that simply don't apply to the masses. There are countless examples of code-writing tutorials, and the vast majority of influencers (I can't believe I just used that word) regale us with demonstrations that never seem to extend much beyond "summarize my YouTube videos", "create 10 catchy video ideas and make thumbnails for each" or "plan my vacation" followed by "This changes everything!" and a thumbnail of them pointing, mouth agape, at the icon for whatever it was that just changed everything.
For all the talk about AI transforming knowledge work, most knowledge workers have been stuck on the sidelines, watching demos and reading articles and thinking "that's cool, but I'm not sure how to apply it to my job." The barrier was a gap between what AI could do and what a specific person actually needed it to do.
And that last part is crucial: People know what they need help with, but the solution needed to be more obvious.
Most people hear "AI" and think: "sounds powerful, but what would I even ask it to do beyond 'tell me a joke?'" Excel users hear "AI in Excel" and immediately think: "Finally, instant pivot tables, instant charts that don't have the X and Y axes flipped, debugging that formula that worked yesterday but doesn't work today, automating that monthly report that takes four hours of copying and pasting, slicers, analysis, and more." The list writes itself because Excel users have been compiling it in their heads for years every time they hit a wall or burned time on something tedious.
The best AI interface is one where users already know what to ask for.
Creating Projects in Claude
One of the challenges for non-coders creating software (ugh, fine…vibe coding) and applications using AI is that the vocabulary, skills, and methods for creating secure scalable software are still very much required in order to build something commercially viable. If you don't know what libraries are (not the place where you get books), or build systems, version control, databases, front-end vs. back-end, and a LOT more—you really aren't heading down a path to success. There are also multiple levels to that vocabulary and skills that really matter when moving from MVP/demo-ware to production software.
By integrating with Excel, a much larger audience of would-be AI users can leverage vocabulary they already possess, structures (scaffolding in AI-speak) they themselves built, and software they're fluent with to automate existing processes, improve their workflows, and create features and functions that were previously just out of reach or too time-consuming to justify.
Learning how to create projects in Claude is still a skill set unto itself. Between Project instructions, Memories, Skills, Artifacts, Capabilities, and now Cowork, as well as the fundamentals of how to effectively talk to/work with LLMs to maximize results and minimize token consumption, there is a fair amount to learn, but when you already have the project goal in mind and you understand the operation of the software you'll be interacting with (Excel, not the LLM), the learning curve for Project creation, improvement, and optimization are easily attainable.
What's next
This is a take-off point for AI. Widespread adoption with current technology and staff skills is there for organizations paying attention.
I wrote previously about how Gartner's Hype Cycle is dead, killed by AI. The argument was that the pace of innovation has been so relentless that markets never get the chance to descend into the Trough of Disillusionment. That there's always something new arriving before disappointment sets in. But reaching the Plateau of Productivity, where technology becomes genuinely useful for the majority of people still requires mass adoption. Benchmarks, demos, influencers, etc. don't get you there, but meeting people where they already work, with tools they already understand, and solving problems they've already identified; that gets you there.
The hints for what's coming are clearly visible in the current Claude capabilities, with read-only access to Outlook and Teams, but if that changes to 'create an account for Claude and let it send and receive', we'll be at the front of the line to test and implement it. This will allow us to leverage our current projects and scale in a way we can't at present. Humans will remain very much in the loop for everything we do, but our ability to scale will (continue to) change materially.
The future of AI isn't only about what models can do. It's about what people can do with them (for now). In this moment, the answer is: a lot more than last year, especially if they know their way around a pivot table.