AI’s next target: the back office

3 trends in startups/tech/venture

⚙️ AI’s next target: the back office

• MIT found that 95% of corporate AI pilots fail to achieve a positive ROI. Why? Because corporations kept buying sales and marketing automation tools rather than focusing on the real automation opportunity: the back-office.

• My prediction? The more front-office automation efforts flop, the more companies will start trying AI for the back-office. (So, if you’re a founder or investor in one of these back office automation startups that’s feeling a lack of love, hang tight—it will soon be your time to shine.)

🤖 AI is actually creating jobs for some young workers

• Stanford researchers analyzed millions of payroll records and found that since late 2022, AI has led to a 13% decline in employment among young workers (ages 22-25) in fields like software development and customer service.

• However, in roles where AI augments rather than replaces, young workers are actually gaining ground. AI can substitute codified knowledge from formal education, but it struggles to replicate the tacit, experience-based skills of older workers.

💰 Silicon Valley has the highest venture capital efficiency

• According to CB Insights, the Bay Area has the highest venture capital efficiency in the U.S. market, with a Multiple on Invested Capital (MOIC) of 10.07 (which means that, for every $1 invested, $10.07 in value is generated).

• While Boston (MOIC of 7.89) and Seattle (MOIC of 7.48) show commendable results, especially given that they have less capital to work with, the Bay Area is consistently #1 when it comes to VC efficiency.

• The Bay Area’s success is likely due to 3 advantages: (1) top universities that attract talent, (2) leading companies that train talent, and (3) great investors that fund new ventures.

2 theses on what’s next

🤖 The next AI race: running efficient “evals”

• As is now widely known, many “AI startups” are nothing more than a foundational model (like ChatGPT, Claude, or Gemini) “fine tuned” for a specific use case (e.g. AI for doctors). These startups can be incredibly successful, but the key questions are: 1) how good is the startup’s fine-tuned model compared to everyone else’s, and 2) can it become / remain the best?

• Two factions have formed within the builder community: One side believes that it’s crucial to do “evals” (to “evaluate” the quality of one’s own output), so you can iterate and improve quickly. The other side believes that evals are a waste of time, because even the best companies like Anthropic, OpenAI and Google run limited evals in-house on a regular basis.

• My view? In the long run, of course evals are important. The question we should be asking isn’t whether to do evals—but rather how to do evals most efficiently. The startups that win in this AI race will be the ones that select, fine-tune, and test models most efficiently.

🤔 The best founders know what success looks like

• Founders obsessed with simply being the best often have worse outcomes than those who focus on getting better at getting better. The key is to know what to get better at. The first step? To define what success even looks like so they know where to focus their energy.

• One of my favorite questions to ask founders is, "What does success look like for you?" Founders’ answers to this single question can shed light on how they think and how they’d most like to be helped.

1 tool I love

📊 Transforming unstructured data into plain English

• 80% of a company’s data is unstructured and messy (think: customer call recordings or customer transactions).

• Storytell helps leaders at companies like Paramount and T-Mobile "talk to their data" and combine the insights with external benchmarks to better understand how their business is performing.