Beating Analysis Paralysis in Enterprise AI Projects
Why do enterprise AI projects never launch?
The single biggest reason, according to Abraham Gomez at Google, is analysis paralysis .
Teams pour effort into understanding everything and nothing ships — the product doesn't launch, the internal initiative doesn't start, and the team just waits.
The pattern shows up in both startups and enterprises.
Leaders see many potential options and hesitate, asking whether they're on the right path and focusing on the highest-ROI thing — and that hesitation freezes the work.
For larger companies there's a second drag: legacy systems .
With many active systems running in production all day, it's very difficult to plug AI in without a human in the loop, which limits how fully a company can adopt what's available now.
“there'll be a big effort in understanding everything, and nothing happens”
What is a realistic timeline for a first AI pilot?
The realistic answer is weeks, not quarters — and sometimes minutes.
The approach is to take the steps that are low hanging fruit and not expensive : define the use case, open a tool like Vertex AI or AI Studio, and start playing with a model against your own data.
Gomez describes processing your own documents with AI as a five minute exercise you might repeat every two months to see how much better it has gotten — until one day it's good enough to become a real product or feature.
Before you dive in, decide how you will evaluate the system at scale .
If today only one domain expert can judge the output, your project isn't just "build the AI" — it's also "replicate the evaluator," and that needs to be documented up front.
“culturally, you don't want your company to be scared of AI”
How do you build a culture that ships AI experiments fast?
Reframe AI development as an experiment , not a KPI commitment.
Through the experiment alone, the team learns a lot, and the barrier is low.
The cultural goal is for the company not to be scared of AI — it's a tool.
Two cultural commitments make this work: Allocate time to stay up to date.
Products built in 2023 already look dated; a slice of time has to be reserved for staying current in your specific domain.
Create and measure, fast.
Bring the idea to life internally as quickly as possible and measure the result right away, rather than pausing and waiting.
On sequencing the experiments, a simple impact vs. value matrix works: target the most time-consuming, lowest-value-add tasks first, and keep humans on the steps where judgment matters most.
“the steps you wanna take are the steps that are low hanging fruit and not expensive”
What happens after the first AI win — how do you avoid a new bottleneck?
Automating one workflow doesn't end the work; it moves the constraint.
As one operator put it, once you're far more productive at creating proposals, you suddenly have to handle the objections of 20x more prospects.
The discipline is to keep asking which step is now the most time-consuming and lowest value-add, and to deliberately keep humans on the steps where understanding the customer matters most — like handling objections, especially as deals get more technical and enterprise.
“you can basically automate a workflow and then the bottleneck becomes the next step after that”
Frequently asked questions.
- What is analysis paralysis in enterprise AI?
- It's when a team puts a big effort into understanding everything about AI — tools, models, risks — and as a result nothing ships. The product doesn't launch, the internal initiative doesn't start, and the team waits and waits. Abraham Gomez at Google calls it the biggest single problem he sees, affecting both startups and enterprises. The remedy is to create the idea and bring it to life internally as quickly as possible, then measure the result right away.
- How fast can a first AI pilot realistically run?
- The recommended approach is to open a tool like Vertex AI or AI Studio, point it at your own data, and start. Processing your own documents with AI can be a five-minute exercise that you revisit every couple of months to see how it has improved. The point is to take the low-hanging-fruit steps that aren't expensive, rather than going from zero to a hundred immediately.
- How should we choose what to automate first?
- Use a simple impact-versus-value matrix: focus on what takes the most time but adds the least value — the example given was adjusting a template in Canva. Keep humans on the higher-judgment steps, like handling objections, where understanding how the customer thinks and prioritizes matters most. As deals get more enterprise and technical, those human-led steps shift toward better product documentation rather than more automation.
- Why do so many AI projects reportedly fail?
- A widely cited MIT figure claimed around 95% of AI projects fail, but the author of that research has since said his words were transformed by the media and his analysis misinterpreted. The narrative stuck anyway. A more honest framing is that it's very easy to start an "AI project," so the failure rate depends heavily on how you define one in the first place.
- Why is evaluation the first thing to design?
- Because without it, the project doesn't scale. If today only one domain expert can look at an answer and judge whether it's right, you don't just have to build the AI system — you also have to replicate that evaluator. Documenting how that person decides is part of the build. Knowing how you'll evaluate at scale is described as really important before you dive in.
- Why is legacy harder for enterprises than startups?
- When you build from a blank slate, integrating AI is much easier. Larger companies have more legacy, more complex stacks, and more complex technology. The main limiter to fuller AI adoption is literally legacy systems: with many active systems running in production all day, it's very difficult to replace modules or plug in new AI capabilities without a human bridging the existing integrations.