Four Questions Every COO Should Ask Before Deploying AI

The AI era is full of promise, every corporation is reporting how much they’ve increased their efficiency and how much AI is doing that. As someone who has run operations in multiple AI startups and now runs an AI VC fund with over 120 portfolio companies I see a different picture. Lots and lots of useful AI tools and automation are being purchased, integrated, and introduced to no or little effect. According to recent McKinsey report on AI potential, nearly 70 percent of AI transformations fail. The problem is that if you introduce even the best AI tool into a messy human-run process, all you get is a messy process that is now also hallucinating and losing context.

One of our investors recently shared that their company has introduced AI agents into one of their operations and then has run a study to see how much efficiency they gained. The results were shocking — their employees were saving a lot of time on something they’ve been doing manually before, but spending the exact same amount of time trying to fix mistakes that AI made. Needless to say, the automation has been brought in by IT and the operations team was left out. Let’s speak about how COOs can leverage AI to actually improve operations.

At DVC we’re not only investing in AI startups, but we’re also early adopters of pretty much every new technology we see. We build our own agents and use our portfolio companies’ products in every aspect of VC work — from sourcing and scoring deals, assisting portfolio founders, or building tools our LPs use to look at angel investment opportunities. Our success in this comes from applying a very boring, but very useful framework.

Before any AI deployment, we ask these four questions:

1. Are There Clear Rules?

Can the process be defined by specific guidelines? If yes, it’s a great candidate for automation. Legal workflows, accounting rules, structured onboarding? Perfect. These are systems where outputs follow rules. AI thrives here.

But if your process is inherently creative — say, brand storytelling or strategic design — full autonomy won’t work, and the process has to be designed with people using copilots.  In brand marketing, breaking the rules often adds value. Don’t outsource that to an agent.

2. Does This Process Have a Single Source of Truth?

If your CRM says one thing, your order tracker another, and the real update lives in someone’s personal spreadsheet — pause. AI systems are only as good as the data you feed them.

Creating a single source of truth and eliminating data or knowledge silos is a gold standard of efficient process design, and for agentic AI it is more important than ever.

When all customer touchpoints and histories are logged in a unified database, AI can automate follow-ups, recommend next actions, and generate accurate reports. And even provide voice customer support or schedule client appointments. A lot of times we see startups succeed when they sell a solution with a built-in source of truth, especially when selling to small businesses, like Avoca AI, a telephone assistant for electricians, integrated with a built-in CRM, ensuring all customer data and interactions are centralized and up-to-date.

3. Is There Rich Data History?

Does every action get logged with examples of how decisions were made? AI learns from patterns in your historical data. No logs, no learnings. If your system doesn’t record what happened and why, it can’t generate patterns. It can’t improve. You’ll waste money.

But even if you’re recording every customer call, transcribing it with AI and storing it in a folder, it probably won’t be enough. Agents working with this should be configured to convert this unstructured data into summarized and structured, maybe even into graphs to better understand relationships, or it would quickly exceed their attention span. Imagine you’re an employee, who gets their memory wiped every time you come to work. You can read and write with superhuman speed, but you have to stare at megabytes of conversation logs and chat history trying to figure out what the company even does and how to do what the manager asked you to. That’s how an AI agent “feels” without a good database.

The best teams don’t just collect data — they structure and version it with the future in mind. That’s when learning loops form. That’s when AI gets smarter, even without having to do any model training.

In healthcare, Collectly applies this principle at scale: using years of annotated billing, payment, and patient interaction data, they optimize medical billing and revenue cycle management. Their AI learns from historical outcomes to reduce errors and speed up collections.

4. Is Your Tech Stack AI-Ready?

Can AI actually plug into your systems and tools, or are you stuck with that internal portal from 1988 that barely loads? We’ve seen cases where internal ops tools were so outdated they couldn’t generate structured outputs — let alone interface with APIs. In those situations, it was often faster and more effective to rebuild the system from scratch than to force AI into legacy infrastructure. If AI agents can use MCP, or a structured and documented API, it’s always better (and cheaper) than when it has to make screenshots of the interface and run them through image recognition to figure out which button to press.

AI is becoming infrastructure. But like electricity in the early 20th century, its potential only unlocks when you redesign the factory, not just install lightbulbs. Don’t retrofit. Reimagine. And, needless to say, a lot of internal tools that used to cost a million dollars to develop before can now be vibe coded from scratch by one of your engineers in their lunch break.

First Principles time.

Now the most interesting part. Let’s say we’ve designed an ideal process – it would be rule defined, will have a single source of truth, and will collect data in a structured way to self-improve. We’ve even persuaded our engineer to spend their lunch break vibe coding a new set of internal tools. But let’s look at this process one more time now. It is very likely that because of automation, it has become much, much cheaper to run. Now try to think what happens to your business with this cost shrunk so much. Try to see a bigger picture – how would this process coexist with other processes if they’re improved in the same way? Maybe it’s time to reimagine the whole thing with AI in mind.

A lot of times thinking about your business operation from the first principles can lead to identifying unexpected opportunities. For example, in DVC we automated deal analysis, due dilligence, and deal memo preparation, effectively going from 6 person/hours to 3 minutes of AI doing the job. Traditionally, VCs would only do all this work after they spoke with the founders and confirmed that the deal is worth spending these 6 person/hours on, and the firm would have a limited number of analysts. Now since it has become so cheap for us, we analyze the market, prepare a deal memo, and even do some due diligence BEFORE we speak to the founder. This enables us to only have calls with companies we know we can and want to invest in, saving time for our partners and founders alike.

But we can go even further with it. Since we effectively have an unlimited analyst, we can move these tools upstream to our investors and scouts, who refer new deal opportunities to us, so they can save their time, analyze every deal through the eyes of a professional VC analyst and reduce the number of times we would have to pass on a deal after reviewing it. We still collect all of the data, because we can use it to learn and make our tools better.

This allowed us to be roughly 8x more productive than a typical VC firm of our size. But we didn’t get here by chance. We mapped our internal operations, applied the four questions, and rebuilt from first principles.

This framework helps startup leaders and COOs shift their mindset: from “Can we use AI here?” — a question of technical possibility — to “Should we?”, which forces a deeper look at strategic value, data readiness, and long-term maintainability. It’s the difference between plugging in tools because they’re available and redesigning processes because it’s the right thing to do.