I went into a conversation about AI startup growth expecting to talk about distribution.
I left thinking about failure.
Not because growth and distribution stopped mattering.
Actually, the opposite.
The conversation made me realize that the more successfully an AI system scales, the more important it becomes to understand what happens when the system is wrong.
That sounds obvious.
But I don't think we talk about it enough.
Recently, I had a conversation with Praveen, founder of NEO SDR, about AI startups, product economics, distribution, AI commoditization, and where durable value might exist as foundation models become increasingly capable.
One idea from the conversation stayed with me:
Blast radius.
It gave me another way to think about AI startup economics.
AI economics aren't just about automation
A simple way to evaluate an AI product is to ask:
How much human work can this automate?
That's a useful question.
If an AI system can perform work that previously required five people, the potential economic value is significant.
But there's another question that becomes increasingly important as the system becomes more autonomous:
What happens when the AI is wrong?
Consider a human making an error in a workflow.
Perhaps they make one mistake.
Someone notices it
The mistake gets corrected.
The process continues.
Now imagine an AI system performing the same workflow across 100,000 transactions.
If the system has a 5% failure rate, the question is no longer simply whether the system is "95% accurate."
You have to ask:
What does that 5% actually mean?
How many transactions are affected?
How expensive is each failure?
Can the mistake be detected?
Can it be reversed?
Does a human review the output?
Does the failure affect customers?
Does it damage trust?
Does it create regulatory, financial, or reputational consequences?
The percentage alone doesn't tell you the whole story.
The idea of blast radius
During our conversation, Praveen described a simple way of thinking about this:
Blast Radius = Cost of Mistake × Volume of Transactions
It isn't a formal accounting equation.
I see it more as a thinking tool.
The important insight is that AI can amplify both productivity and failure.
Suppose a human completes 20 transactions and gets one wrong.
That's one mistake.
Now imagine an AI system performing 100,000 transactions with roughly the same error rate.
The underlying error rate might not look dramatically different.
But the consequences can be.
That's the blast radius.
The system's ability to operate at scale becomes the same mechanism that allows an error to propagate at scale.
Scale doesn't only amplify success
This is where I think the conversation connects directly to startup growth.
Growth teams naturally think about leverage.
How do we acquire more users?
How do we increase activation?
How do we improve retention?
How do we automate more of the workflow?
How do we reduce the cost of serving each customer?
These are important questions.
But AI introduces another form of leverage.
Automation leverage.
An AI agent doesn't just help one person complete one task.
It can potentially execute the same type of task repeatedly.
That changes the economics.
If the workflow works, that's powerful.
If the workflow fails, the failure can potentially propagate just as efficiently.
So perhaps one of the questions AI startups should ask before scaling isn't only:
"How much value can this create?"
but also:
"How much damage can this create if we're wrong?"
This changed how I think about AI startup growth
My usual lens for evaluating startups starts somewhere around:
Problem → ICP → Positioning → Distribution → Adoption → Growth
Who has the problem?
How painful is it?
Who is the ideal customer?
Why this product?
Why now?
How does the startup reach those customers?
How does it convert attention into adoption?
How does it retain users?
Those questions still matter.
But after this conversation, I think there's another dimension worth adding.
Failure economics.
The question becomes:
What happens when the product succeeds at scale but the underlying system isn't always right?
That creates a different set of questions.
A simple AI Growth + Risk Audit
I'm experimenting with a framework that combines the traditional growth questions with questions around reliability and failure.
It's still a work in progress.
Demand
Who has the problem?
How painful is it?
What evidence shows that people actually want the solution?
Are people paying for it?
Are they returning?Positioning
Why this product?
Why this customer?
Why now?
What outcome is being promised?
Can the customer understand the value without needing a technical explanation?
Distribution
How does the product reach customers?
What acquisition channels exist?
Is there a repeatable distribution loop?
Does the founder have a distribution advantage?
Can growth scale without the acquisition cost becoming unsustainable?Value
What measurable business outcome improves?
Does the product generate revenue?
Reduce cost?
Save time?
Increase conversion?
Reduce operational workload?
The more measurable the outcome, the easier it becomes to evaluate whether the AI is actually creating economic value.Reliability
How often is the system wrong?
Does performance vary significantly between different contexts?
Can errors be detected automatically?
How quickly are failures identified?
What happens when the model encounters something outside the expected workflow?
- Blast Radius This is the part I'm most interested in exploring. If the system makes a mistake:
How many actions can that mistake affect?
How quickly can the failure spread?
What does one failure cost?
Is the failure reversible?
Can the system stop itself?
A low-frequency error can still be economically significant if the system operates at enormous volume.
Human Control
Where does a human remain involved?
Does someone review the output?
Does someone approve high-risk actions?
Can a human override the system?
Can the workflow be stopped?
The goal isn't necessarily to keep humans involved everywhere.
The question is:
Where is human intervention economically and operationally justified?Economics
This is where everything comes together.
What does one successful workflow cost?
What does one failed workflow cost?
What does human oversight cost?
How much does automation actually save?
What happens to margins as usage grows?
And perhaps most importantly:
Does the economic model still work when failure is included?
Reliability and growth aren't separate conversations
This is probably the biggest shift in my thinking.
I've traditionally thought about growth as something that happens after you establish that a product solves a meaningful problem.
But with AI systems, reliability can become part of the growth equation itself.
Imagine two products.
Product A automates 1,000 workflows per day.
Product B automates 100,000.
At first glance, Product B looks much more scalable.
But suppose Product A has strong monitoring, human review for high-risk actions, and cheap recovery when something goes wrong.
Product B operates almost completely autonomously, but failures can propagate across thousands of transactions before anyone notices.
Which one has the stronger growth model?
The answer isn't obvious from user count or automation volume alone.
This is why I think scale and reliability have to be evaluated together.
The AI commoditization problem
Another part of the conversation that stood out to me was Praveen's shift from a traditional SaaS model toward white-label.
The reasoning was tied to something many AI founders are already experiencing:
Foundation models are commoditizing parts of the execution layer.
If a general-purpose model can increasingly perform functionality that once required a specialized SaaS product, then simply building another interface around similar capabilities may not create much defensibility.
That raises a harder question:
Where does durable value live when the underlying intelligence becomes cheaper and more accessible?
Potential answers include:
- Distribution
- Customer relationships
- Workflow integration
- Proprietary data
- Domain expertise
- Trust
- Reliability
- Specialized workflows
- Implementation
- Brand
- Operational knowledge
And this brings the conversation back to growth.
If execution becomes increasingly commoditized, distribution and customer relationships may matter more.
But if autonomous execution becomes increasingly powerful, reliability and trust may matter more too.
The interesting opportunity may exist at the intersection.
What I would ask an AI startup today
After this conversation, my questions for an AI startup would look slightly different.
I'd still ask:
Who is the customer?
What problem are you solving?
Why does the customer care?
How are you acquiring users?
What measurable outcome are you producing?
But I'd also ask:
What happens when the AI is wrong?
How many actions can one failure affect?
What does one failure cost?
How quickly can you detect it?
Where does a human remain in control?
What happens to your economics when those failures are included?
And:
What remains valuable if the underlying models become dramatically better and cheaper?
I'm not convinced I have the answers yet
That's probably the most important part.
I'm not presenting this as a finished framework.
I'm testing an idea.
The concept of blast radius made me realize that my previous growth lens was missing something.
I was asking:
Can this startup create demand and distribute the product?
Now I'm also interested in asking:
If this startup succeeds, what exactly gets amplified?
Revenue?
Productivity?
Customer value?
Or potentially failure?
Because AI creates an unusual kind of leverage.
It can make a workflow dramatically more valuable.
But it can also make the consequences of a bad workflow dramatically larger.
That's why I think the next generation of AI startup thinking needs to connect:
Demand + Distribution + Value + Reliability + Failure Economics + Defensibility.
I'm still working through what that framework should actually look like.
But this conversation changed the question I ask.
And sometimes that's the more useful outcome.
What do you think?
When evaluating an AI product, should "blast radius" become a standard consideration alongside ROI, distribution, and defensibility?
Or is there a better way to think about the economics of AI failure at scale?
Top comments (0)