Your AI Business Can Grow While Your Margins Shrink

Most founders love seeing graphs like this.

Everything appears to be moving in the right direction.

More customers.

More subscriptions.

More AI requests.

More revenue.

For traditional SaaS businesses, those trends usually tell a reassuring story.

Growth often brings economies of scale.

Serving one additional customer typically costs very little.

Average metrics become reasonably good indicators of business health.

AI products are different.

An AI business can continue growing while quietly becoming less profitable.

Revenue increases.

Usage increases.

Customers remain happy.

Yet margins slowly deteriorate.

Not because pricing is wrong.

Not because customers stop paying.

But because the economics of AI are hidden inside the distribution of usage rather than the averages shown on most dashboards.

Growth tells you that the business is expanding.

It does not necessarily tell you that the business is becoming healthier.

Traditional SaaS trained us to think differently

For years, software companies learned to evaluate success through a relatively small set of financial metrics.

Dashboards typically focused on indicators such as: · Monthly Recurring Revenue (MRR) · Annual Recurring Revenue (ARR) · Average Revenue Per User (ARPU) · Gross Margin · Customer Growth

These metrics work remarkably well for many traditional SaaS products.

Once the software has been built, serving one additional customer usually introduces very little additional cost.

A customer logs in.

Views another dashboard.

Creates another project.

Runs another report.

The infrastructure certainly performs more work, but the marginal cost of those actions often remains relatively small.

That economic model shaped the way many founders learned to interpret growth.

If revenue increased while gross margins remained healthy, the business was generally moving in the right direction.

Average metrics provided a reliable picture of the company's overall health.

AI products challenge that assumption.

The cost of serving one additional request is no longer close to zero.

Every interaction may consume resources that directly affect profitability.

The averages still matter.

They simply stop telling the entire story.

AI products changed the economics

Unlike traditional SaaS, AI products often incur variable operational costs every time customers use the product.

A single interaction may trigger: · Model inference · Token consumption · GPU time · Retrieval pipelines · Agent execution · Image generation · Voice synthesis · External APIs

Each of those operations consumes infrastructure with a measurable financial cost.

Two customers paying the same subscription price may generate completely different operating costs.

One may submit a handful of lightweight requests each day.

Another may execute hundreds of long-running agent workflows.

From a revenue perspective, they appear identical.

From a profitability perspective, they may have almost nothing in common.

This is one of the biggest economic shifts introduced by AI.

Revenue becomes increasingly predictable.

This shift also changes how companies think about AI monetization after payment, something I explored in Why Payment Is Only the Beginning for AI Products.

Costs become increasingly variable.

That changes the questions founders need to ask.

Instead of looking only at how much revenue the business generates, they also need to understand where that revenue creates healthy margins—and where it quietly erodes them.

Why averages become dangerous

Most dashboards summarize a business using averages.

Average revenue per customer.

Average infrastructure cost.

Average gross margin.

Average usage.

Those metrics are useful.

The problem is that they can also be deeply misleading.

Imagine an AI product with one hundred paying customers.

The dashboard reports:

Everything appears healthy.

Now imagine looking beneath those averages.

The business hasn't changed.

The perspective has.

The averages suggested a healthy company.

The distribution reveals where profitability is actually being won—or lost.

This is one of the biggest differences between traditional SaaS and AI products.

AI costs rarely distribute evenly across customers or workflows.

A small number of interactions often accounts for a disproportionately large share of infrastructure spending.

Revenue grows across the entire customer base.

Costs tend to concentrate.

Without visibility into that distribution, healthy-looking metrics can hide unhealthy economics.

Averages tell you how the business looks. Distributions tell you how the business behaves.

Workflow profitability is becoming more useful than token counts

When AI costs begin increasing, many teams instinctively monitor token consumption.

That is a reasonable starting point.

Tokens are measurable.

Easy to aggregate.

Easy to visualize.

Unfortunately, they rarely answer the question founders actually care about.

The business doesn't exist to optimize token counts.

It exists to produce profitable outcomes.

Consider two AI workflows.

Looking only at token usage, Workflow A appears far more expensive.

Looking at business value, it may be significantly healthier.

The important question is no longer: How many tokens did we consume?

It becomes: Was the outcome worth the cost of producing it?

That shift changes what companies need to measure.

Instead of asking only: · Which model generated the cost? · How many tokens were consumed?

Teams increasingly need answers such as: · Which customer generated this cost? · Which workflow generated this cost? · Which feature generated this cost? · Did the completed workflow remain profitable? · Did retries significantly change its economics? · Would we make the same execution decision again?

This is why workflow profitability is becoming a far more meaningful metric than raw infrastructure usage.

Customers purchase outcomes.

Healthy AI businesses increasingly optimize the economics of those outcomes rather than the cost of individual model calls.

Visibility must come before optimization

When founders notice AI costs increasing, the first instinct is often to revisit pricing.

Should subscriptions become more expensive?

Should credits be introduced?

Should usage limits change?

Should top-ups become mandatory?

Those are reasonable questions.

But they all assume the company already understands where its costs come from.

In practice, many teams don't.

They know the cloud bill increased.

They know model usage is growing.

They know margins are changing.

They often don't know why.

Pricing decisions made without operational visibility are mostly educated guesses.

Before changing pricing, companies increasingly need answers to questions such as: · Which customers are consistently profitable? · Which workflows generate the highest costs? · Which AI features create the most business value? · Which execution paths require repeated retries? · Which providers contribute most to operational costs? · Which outcomes become unprofitable after infrastructure overhead is included?

Without this visibility, optimisation becomes reactive.

The business adjusts prices without understanding the underlying economics.

That may improve revenue.

It rarely fixes the real problem.

Understanding the distribution of costs should come before attempting to redistribute them.

Engineering is becoming part of business strategy

For many years, engineering and business strategy were largely separate conversations.

Engineering focused on building reliable systems.

Finance focused on revenue, margins and profitability.

Product teams focused on customer experience.

AI products increasingly blur those boundaries.

Today, infrastructure decisions directly influence business performance.

A metering error can distort customer profitability.

A retry policy can change workflow economics.

I've previously written about why retry behaviour is increasingly becoming an economic concern rather than just a reliability concern in Why Retry Safety Is Becoming a Business Problem for AI Products.

Incorrect usage accounting can make pricing appear successful when margins are quietly deteriorating.

A missing authorization check can turn profitable customers into unprofitable ones.

None of these issues begins in a pricing spreadsheet.

They begin inside the runtime.

This is one of the most significant shifts introduced by AI.

Infrastructure is no longer just responsible for delivering software.

It increasingly determines how healthy the business becomes as it scales.

Engineering decisions now influence questions that traditionally belonged to finance: · Which customers are profitable? · Which workflows should be encouraged? · Which features justify their operational cost? · Which execution paths should be redesigned?

The conversation is no longer simply about building systems that work.

It is about building systems whose economics continue to work as the company grows.

Engineering no longer supports business strategy. In AI products, it increasingly shapes it.

A new way of thinking

Traditional SaaS businesses often begin with a familiar question: How much revenue are we generating?

That question still matters.

But for AI products, it is no longer sufficient.

A more useful question is: Which customers and workflows are actually profitable?

Those questions produce very different decisions.

The first encourages growth.

The second encourages sustainable growth.

One useful mental model looks like this:

Each layer provides context that averages alone cannot reveal.

Revenue tells you that customers are paying.

Runtime events explain what actually happened.

Infrastructure costs show what the company spent.

Business outcomes reveal whether those costs created value.

Healthy AI companies increasingly optimise across the entire chain—not just the first metric.

Growth remains important.

But understanding how growth is generated is becoming just as important as measuring how much growth exists.

Revenue tells you how fast your AI business is growing. Profitability distribution tells you how healthy it is becoming.

Closing

The first generation of AI products taught us how to integrate models.

The second generation is teaching us how to operate them sustainably.

Those are different challenges.

Building an AI application is becoming increasingly accessible.

Building an economically healthy AI business remains considerably harder.

The companies that succeed won't necessarily be the ones with the largest models or the lowest inference costs.

They'll be the ones that understand the economics hidden inside their own infrastructure.

That means looking beyond averages.

Looking beyond monthly revenue.

Looking beyond total token consumption.

Instead, they'll understand: · Which customers generate healthy margins. · Which workflows create sustainable value. · Which runtime behaviours quietly erode profitability. · Which engineering decisions influence business outcomes.

Growth is still important.

But growth without visibility can be misleading.

Revenue can continue increasing while the underlying economics quietly deteriorate.

Healthy AI businesses increasingly optimise for something more fundamental than growth alone.

They optimise for sustainable profitability.

Learn More

As AI products mature, a new infrastructure pattern is beginning to emerge.

Rather than treating payments, runtime execution and business analytics as completely separate systems, many teams are starting to connect them into a single operational layer.

Its objective isn't simply to process payments or record usage.

It's to keep commercial state, runtime execution and business economics aligned throughout the lifecycle of every AI request.

That increasingly includes capabilities such as: · Runtime aut

... (truncated at the site's per-page length limit; see the source link below for the full text.)