Author: Lincoln Wang | Founder of MindsLeap | Global Partner at Founders Space | Founder of Founders AI Club
Late-Night Task Lists for Blue
"I'm iterating with Blue late at night, telling her: OK Blue, here's the task list for you to work on while we're all asleep."
That line comes from Rana El Kaliouby — AI scientist, serial entrepreneur, and now partner at venture firm Blue Tulip Ventures. She left MIT in 2009 to found Affectiva, bringing human emotion recognition into the product testing systems of half the Fortune 500, before selling the company to a Swedish automotive firm listed on the public exchange. Now, as an investor watching AI adoption unfold, she told a story on the Microsoft WorkLab podcast that sounds like bragging at first — and deeply unsettling on reflection.
Blue is her fund's first in-house AI agent, trained to handle day-to-day operations. She's not fully mature yet — "she still gets stuck in certain situations." Even so, Rana has already moved to a working mode where she talks to Blue on her phone late at night, assigning tasks for the next day, letting Blue finish them while everyone else sleeps.
That detail made me pause. She wasn't saying "I'm trying out a tool." She was saying "I'm already collaborating with a colleague who doesn't sleep." The first is experimentation; the second is dependence — and dependence is what real adoption looks like.
But Blue still forgets things. AI's memory isn't stable yet, and its sense of time is fuzzy. Tell it "Monday morning," and it doesn't really know what Monday morning means. Rana says these technical issues will be fixed — it just requires ongoing磨合 (calibration) for now.
And that calibration is exactly where most enterprises get stuck on AI adoption.
The Pilot Succeeded. Now What?
Rana sees a recurring pattern in her investments: companies can start AI experiments easily, but struggle to embed the results into daily workflows.
"Experiments are easy to start. But integrating what you've built into your everyday workflow — that's the hard part."
That's followed by an even more sobering judgment: many companies are actually using AI to optimize a process that was already broken. She's blunt about it — "If your workflow doesn't work already, putting AI on top of it won't make it work."
A broken process plus AI still gives you a broken process.
This isn't a novel insight, but it explains why so many AI pilot projects quietly die. It's not that the AI tools aren't good enough — it's that nobody bothered to dig out the real problems in the business process before AI showed up. AI becomes a patch slapped onto a pipe that was already leaking.
Rana's Blue Tulip has a clear investment criterion: the solution must be deeply integrated into the workflow, not "sitting on the side." "If the solution you're building is bolted on to the side, the friction is too high. Customers might do a proof of concept, they might try it, but they won't end up using it in their daily work."
What fits into the workflow lives; what doesn't, dies. For B2B AI products, this is almost a survival line.
The Remaining 30% Is the Real Work
Rana describes what she calls the "60-70% problem."
Many AI coding platforms and collaboration tools can get you 60% or even 70% of the way there. But the remaining 30% — the part that requires ensuring accuracy, preventing hallucinations, and not breaking existing workflows — turns out to be even more work.
For anyone transforming business with AI, this is a critical signal: the gains AI delivers are most easily wasted at the last mile.
It's not a technology failure — it's a handoff failure. AI does 60%, then hands an unvalidated output to a human who isn't mentally prepared to take it over. That seam is where most AI projects carry their real risk today.
That's why she keeps emphasizing the value of human oversight: not excluding people, but intelligently deciding which steps need human involvement and where AI autonomy delivers the most value. Oversight isn't about distrusting AI — it's about understanding where AI's boundaries lie.
When AI Agents Become "Colleagues," Who Manages Them?
There's a question in the podcast that almost nobody is seriously thinking about.
When a team has both human employees and AI agents, how do you enforce company culture? How do you define accountability? What's the escalation path when an AI agent gets something wrong?
Rana says one of their investment theses is the "AI colleague" form — AI existing as a collaborator rather than a tool within a team. But her concern lies here too: "We're so focused on what this colleague can do, and we're not thinking about how it will coexist alongside real human colleagues."
More subtle still is the bias problem she raises. We have explicit and implicit biases in hiring. And as we "create" AI agents, the same biases are quietly at work — we just haven't realized this needs to be treated as a serious organizational issue.
"We haven't given these questions enough attention when it comes to creating AI colleagues."
This is an organizational question, not a technical one. Most companies talking about AI adoption today are still discussing tool selection, cost, and accuracy. The questions of how AI agents fit into organizational culture, how they bear responsibility, and how they find their place in team collaboration — these remain almost entirely blank.
Another company Rana invested in, Tough Day, offers an interesting angle: their conversational AI agent Tuffy helps people navigate workplace challenges — dealing with difficult colleagues, handling tough situations when support isn't available. This positions AI as a "workplace emotional coach" rather than a productivity tool.
Emotion Is the Foundation of Trust, but AI Can't See It Yet
Rana brings a unique perspective from her 10+ years at Affectiva: over 90% of human communication happens outside of language — facial expressions, body language, gestures, tone of voice.
Current AI can't see any of this. It only processes what you said, not how you said it.
She gives an example: a truly empathetic AI colleague should be able to sense that you're not in a good headspace today and know now isn't the time to press that complex question. "Empathy is actually one of the most core ways humans build trust. I see you're not OK today, and I ask: what's wrong, is everything alright? That kind of empathy is what matters."
This isn't a soft topic. Trust is the underlying variable determining whether AI agents get truly adopted. A tool people trust gets used. A tool people don't trust gets worked around, no matter how capable it is.
And trust between people is built largely through those 90% of nonverbal signals. AI can't read them yet, so its responses always fall slightly short. That gap is negligible in some scenarios and fatal in others. Rana believes emotional perception capability will be a core differentiator for future AI agents, especially as AI begins to enter the physical world and highly sensitive workplace contexts.
The Moat Must Survive the Next Model Version
When asked about competitive advantage for AI companies, Rana gives a very practical standard.
Data is a moat. Proprietary IP is a moat. But she emphasizes the moat's "lifespan."
"If you're worried that the next version of an AI model will make your product obsolete, that's not a defensible business, and I wouldn't want to invest in it."
This judgment is especially sharp today. AI foundation model capabilities make significant jumps every few months. A product moat built on the limitations of an older model version can disappear entirely with the next model update. What has real longevity are defenses that don't depend on the model's current limitations — unique industry data accumulations, deep switching costs from embedding into customer workflows, specialized knowledge graphs built in vertical domains that external models can't easily replicate.
She adds an investment preference: she focuses on companies that completely reimagine vertical industry workflows with AI. Not adding AI to old processes, but redesigning the entire workflow from scratch with AI as the fundamental assumption. "Transform those legacy industries, reimagine the entire workflow with AI." That's where they're placing their bets.
Learned Through Play, Not Training
In the lightning round, Rana was asked: what capabilities should individuals develop to stay valuable in the AI era?
Her answer: "Play. Stay playful, stay experimental."
This response is easy to dismiss as a motivational soundbite. But the context she gives it more weight. She has two children: a 22-year-old daughter who refuses to use any AI tools, insisting human connection is what matters most; and a 17-year-old son who does the opposite, always trying the latest AI products, staying at the frontier.
"The right answer is in the middle." She agrees with her daughter's insistence on human connection, but also believes her son's curiosity will give him real power to influence where the technology goes. Understanding it and using it is what gives you the chance to shape it.
For companies, this is an organizational challenge. In any organization, there will always be "daughter types" and "son types" — refusal and full commitment coexist. A leader's job isn't to force everyone to become the son, but to create an environment where people can play safely — where the cost of failure is low, the pace of experimentation is fast, and the path to seeing results is short.
Rana's final three pieces of advice for leaders: commitments must be clear enough to send a strong signal to the organization; you must build a culture of experimentation; and you must stay current with the pace of technology, because AI moves too fast to face this year's tools with last year's thinking.
What entrepreneurs most need to take from Rana isn't a specific tool recommendation — it's a shift in the judgment framework.
Not "have we adopted AI," but "have we fixed our process first, then embedded AI into it."
Not "have we run a pilot," but "has the AI in our pilot truly penetrated the capillaries of daily work."
Not "what features does our AI product have," but "after the next stronger foundation model comes out, what do we have left."
Blue works at night, but Blue doesn't yet know what Monday morning means. That detail isn't the point — what matters is that Rana is already iterating with her.
About MindsLeap
MindsLeap is an AI-native organizational transformation accelerator.
We partner closely with Founders Space, the Silicon Valley innovation incubator, continuously connecting global AI frontier knowledge, Silicon Valley tech entrepreneurship ecosystems, and the real transformation scenarios of Chinese business leaders.
Around AI-native organizational development, MindsLeap is building a transformation ecosystem for entrepreneurs, founders, AI engineers, industry experts, and investors — helping enterprises move AI from cognition, strategy, and tools into real organizational capability, business processes, product innovation, and growth systems.
This article was translated and adapted from the Chinese original with AI assistance.