011 AI For Operators

Osheen Mahajan, Chief of Staff at General Agents, shares how a CoS embedded in an AI lab automates workflow gaps, vibe-codes internal tools, and helps train agents that act.

In this episode, Osheen and I cover:

• What it’s like working inside an applied research lab, where she supports teams building computer-use agents that aim to automate digital work

• Applying a biz ops mindset to ML ops: mapping workflows, automating repetitive steps, removing daily annoyances so researchers can focus on high-impact work - similarities to more traditional companies with some important departures.

• Using Lovable to build internal tools as a non-technical generalist: accelerating time-to-completion with a usable UI, clear processes, integrations with tools used across the team (e.g. Notion) allows her to focus on driving adoption with the team, rather than requesting technical resources.

• Her commitment to context, not code: how she’s doubled down on the core ‘people’ skillsets of a Chief of Staff (integrating teams, providing context) vs. learning to become deeply technical, and how this works to her advantage, even at a highly-technical company like an AI research lab.

• The key to being credible in a highly technical team: push to be included in product and engineering meetings to ensure you have the context and know the workflows so you can spot gaps and solve problems

• The full playbook: learn AI tools, embed with the technical team, prototype lightweight apps, and combine knowledge-through-proximity with people skills to become a true ‘Chief of AI’.

What It Does

Julius is your AI data analyst - create charts or do more in-depth analysis.

Why Ops Leaders Should Care

If part of your/your team’s job is analyzing data or preparing presentations or memos about analysis you’ve done, automating a lot of it is a great way to save time and improve the quality of your outputs, especially if you’re part of a small team that’s asked to do a lot.

Key Features (Pros & Cons)

Pros

• User-friendly: upload data or connect your data and ask questions in plain English to get analyses and charts

• Handles a variety of data formats and inputs, including PDFs, Google Sheets, Excel, CSV, JSON, images of tables, and more

• Direct connections to your databases, including Postgres, Snowflake, and others

• Transparency - if you’re technical, being able to view the code that’s generating the analysis (and then being able to tweak it), can be helpful to help you get exactly what you want - and verify that it’s accurate

• Collaborative features: shared workspaces and annotations allow teams to work together

Cons

• Not as effective with big, messy data sets (according to reviews), despite supporting uploads of files up to 32 GB

• Restrictive free trial limits free users to 15 messages with AI per month - if you’re using it with any regularity, you’ll need to upgrade (understandable, but could be limiting). Additionally, because pricing is essentially tiered based on usage, if you’re a power user, it can get pricey quickly.

An Operator’s Perspective

It was very simple to onboard to Julius - the free trial is available without inputting a credit card. Connecting data sources was straightforward, too: it took just a few clicks to connect a Google Sheet or upload a document.

Once connected, beginning the analysis is straightforward, too. I queried the data with natural language and saw it get to work, spinning up Python while simultaneously explaining the steps it was taking in bullet points below. Finally, it created a chart to show the output along with a written explanation.

To ensure this kind of analysis was board ready, I’d definitely want to scrub the data first and sanity check the results, but as a first cut, this was pretty useful.

Other Options

Bottom Line

If you’re in a Chief of Staff or Biz Ops role and want better charts or to speed up your data analysis work, but don’t want to spend all day in spreadsheets, write too much code, or spend a lot of money, Julius could be the solution for you.

Practical

• 25 proven tactics to accelerate AI adoption at your company: A few themes include making sure the CEO is leading from the front, incentivizing and rewarding adoption (e.g. in performance reviews and all hands), and reducing the barriers to trying new tools.

• How Zapier measures its employees AI fluency: A role-by-role breakdown from Zapier CEO Wade Foster.

• The exploding cost of tokens: the subscription pricing prisoner’s dilemma, and why enterprise deployments may be the only solution for AI companies to stop setting giant piles of cash on fire.

• How to build an internal AI tool that everyone will love: A demo and tactics from the Chief Engineering Officer of Amplitude.

Perspectives

• Some interesting thoughts from @fleetingbits on X on the the GPT-5 launch and why the launch may have been more about the competitive dynamics between OpenAI and Anthropic than blowing consumers’ minds

•  Zvi Moskovitz on GPT-5: personalities, costs, and is this the best that OpenAI can do or just the best that they can safely release?

• Nabeel Qureshi on GPT-5

• The hidden penalty of using AI at work: Research from HBR shows how using AI can actually make certain populations appear less competent, and offers tips on how to correct this.

A Few Links on Vibe Coding

• How Mailchimp got 40% more efficient: but had to deal with governance challenges and where vibe coding falls short, along the way.

• Vibe code is legacy code: a warning by analogy.

• More vibe code, more tech debt: findings from a big GitClear study.

• Lovable’s slimmed-down approach to product management: perspective from a growth leader at the vibe coding company

• A detailed vibe coding workflow: from an experienced engineer.

Thanks for reading,

Tom Guthrie

If you liked this edition of AI for Operators, share it with another operator you know!