Our Favorite AI Workflows of 2025

Dec 23 2025

Lan Xuezhao, Chang Xu, John Mannes, Wilson Kyi, Rachel Wong, Jennifer Lee, Ingrid Chen and Lisa Chen

With the year wrapping up, we asked the Basis Set team to share some of the AI workflows that had the biggest impact on how they worked in 2025. Of course we build a lot of our own tools in house, but we wanted to highlight some of our favorite off-the-shelf tools. We hope this offers a glimpse into how we use AI day to day, where we see it creating real value in everyday tasks, and most importantly, sparks a few ideas you’ll want to run with yourself!

Automating LinkedIn list building: Simular

Warm introductions are one of the highest-leverage ways we can support founders, but it’s often more challenging to figure out who to reach out to, than to actually reach out. Finding the correct roles at the right companies typically means hours of searching, filtering and cross-checking across LinkedIn, time better spent on judgment, prioritization and crafting introductions that actually land.

We use Simular to automate the heavy lifting of list-building. By guiding browser-level workflows, Simular helps us pull targeted LinkedIn profiles for customers, hiring, or investor outreach, which we then review and refine manually. This human-in-the-loop approach increases both the quality and quantity of introductions, enabling us to hit SLAs when timing matters most.

Video creation: OpenArt + Canva

Video has become one of the fastest ways to communicate ideas, but historically it’s been expensive and slow to produce. By pairing generative image models with lightweight editing tools, we can iterate quickly on content and quickly adjust our framing in real-time based on market feedback.

We use OpenArt to generate custom backdrops trained on our visual corpus, then bring those assets into Canva to add motion, text, music and structure. This workflow lets us produce polished videos without heavy production overhead.

AI-voice workflows: Flow & Willow

Voice is emerging as a bridge between human intent and AI execution. By speaking directly to AI tools, we’re able to move beyond the constraints of keyboards and small screens and make progress on complex work in moments that would otherwise be lost.

In practice, tools like Flow and Willow let us handle follow-ups and run complex research threads, while moving between meetings or commuting. Instead of waiting to be back at a desk, voice allows us to trigger multi-step reasoning and actions that don’t fit neatly into mobile interfaces, turning in-between time into meaningful forward motion.

OCR document extraction: Claude Code

A surprising amount of important information still lives in PDFs, scans and screenshots, formats that are easy to store but hard to work with. Yet robust automation still starts with getting access to clean data that is easy to manipulate.

We use Claude Code to vibe code OCR-based extraction pipelines to turn those static files into searchable, structured text we can actually analyze, summarize and reuse. The payoff is speed and accuracy: less manual copy/paste and fewer missed details, especially when we’re moving through messy operational docs or supporting our companies in their own document automation workflows.

Meeting-to-insights pipeline: Quill

Making AI systems useful depends on getting the right data into the right formats quickly. We use Quill to produce structured outputs, then rely on Quill's automation to consistently format that data into predefined templates, ensuring it’s clean, predictable and ready for downstream use.

The formatted content is automatically exported as markdown and organized into folders for ingestion via Claude Code, allowing our internal models to incorporate new information faster. This workflow shortens the path from raw data into something that’s queryable and meaningfully improves the intelligence of the systems we’re building.

AI legal workflows: ChatGPT

AI has become a useful second set of eyes in our legal workflows, helping us read faster, question assumptions and flag where language doesn’t quite do what it intends. It doesn’t draft or decide for us, but it meaningfully shortens the path to informed judgment, especially when speed and clarity are critical.

In practice, that looks like using AI to simplify dense clauses, sanity-check standard provisions, explore alternative financing structures and refine tone in sensitive communications. The goal isn’t to outsource legal thinking, but to spend less time parsing boilerplate and more time on the decisions that actually shape outcomes.

Agentic browsing: Simular Browser, Perplexity Comet & OpenAI Atlas

Agentic browsing has changed how we research. Instead of manually hopping between tabs, AI agents can explore, summarize, cross-reference and report back with context intact.

Tools like Simular Browser and Perplexity Comet made browsing feel less like searching and more like delegating. And with it, we were able to meaningfully expand how much ground we could cover in a single sitting.

Animated images: OpenArt + Reve

As it becomes easier to produce high-quality content by pairing human judgment with the speed and memory of large language models, we’ve been able to ship and experiment more. That momentum has opened up new ways to bring our work to life visually, including the use of animated GIFs.

With our own OpenArt model trained on our full visual corpus, we can quickly generate the right image for each context, then use Reve to add motion and bring it even more to life.

Podcast creation: NotebookLM + Spotify

Long-form research and technical content often struggles to travel, even when the ideas matter. Audio gives ideas a second life, but only if content can be shipped quickly. In AI, waiting weeks to produce content means the conversation has already moved on.

We use Notebook LM to turn dense materials, like NeurIPS research, into podcast-style audio quickly. By collapsing production timelines to hours, we’re able to share new ideas while they’re still shaping the conversation, translating important work into an engaging format without losing momentum or substance. Check out our NeurIPS 2025 podcast here!