A product leader asked me recently:

On the surface, this sounds obvious.

Everyone is saying AI is bringing the cost of development down to near zero. Faster dev cycles, smaller teams, more output = less cost and higher margins.

Right?

Not so fast.

The more I’ve thought about this, the more I think that logic rests on a few assumptions that don’t really hold up in practice.

If you’re leading a product org right now, this matters. Because it changes how you think about your product strategy, pricing, roadmap, P&L, your leverage with the executive team, and your relationship with the CTO.

Let’s break it down.

This is the biggest trap. Most people think AI is a clean swap:

Less engineering labor → lower cost → better margins.

But that’s not what’s happening. We’re not eliminating cost. We’re shifting it.

Historically, most R&D budgets looked something like this:

That was the envelope.

With AI, you’re introducing a new cost center: Inference. And it’s not cheap. Running LLMs costs more than it does to buy 30 GitHub licenses. As a result, in many cases, teams are moving toward something closer to:

And the total spend is often higher than before.

Even when you reduce headcount, you don’t “save” that money cleanly. You replace it with agents that can burn through token usage at scale.

A senior engineer might cost $256K a year, fully loaded, which is $123 per hour or $985 a day spread across working days. That may sound like a lot except that an autonomous agent can burn through thousands of tokens in a week in compute, expanding context windows, retries, rewrites, and looping on a problem — unless you’re tightly managing usage.

This means engineering labor costs and token costs are not a clean swap. You didn’t remove the cost. You just changed its shape.

Tomasz Tunguz captured this shift well. He explained he’s getting 31 tasks a day at $12K a year. By that math, an engineer on the payroll at $100K needs to be producing 8x more output to justify the delta.

The question used to be:

“How many engineers do we need?”

Tomasz is proposing the new question will be:

“How much productive work are we getting per dollar of inference?”

That’s an entirely different conversation. It’s the conversation CFOs, boards, and investors are starting to have.

And most engineering teams aren’t ready for it.

At first glance, this feels true. AI makes coding more accessible. Anyone can use Claude to build little pet projects and or Lovable to vibe code functional software. Anyone can be an engineer now. Ergo, engineers get cheaper.

But there’s a massive difference between that and building production-grade software that scales. One that deals with SOC2, compliance, security, etc.

This means the job of an engineer is evolving. An engineer armed with the right tools and a generous LLM budget can do the work of the full front-end, back-end, QA, and DevSecOps squad. The best engineers won’t just write code. They’re building the full stack:

So instead of hiring 10 engineers at 1x pay, you're paying one engineer at 2x. The total headcount goes down, the price per head goes up.

And your dependency on their judgment increases significantly.

This is the one most people miss.

If AI makes building cheaper, then everyone builds more, which makes it more expensive to compete, not less. Because everyone has the same tools.

This is a classic dynamic:

Lower cost of production → higher level of competition → more total spend.

This is exactly what happened in manufacturing. As the cost of manufacturing shriveled, the cost of competing went up.

This is now hitting software development.

Think of it like this: AI doesn’t just increase your capacity; it increases the market’s expectations. So whatever efficiency you gain gets reinvested into the next roadmap bet. Immediately.

From a product standpoint, this shows up as:

So, while you may spend less on individual product initiatives, you'll likely spend more in aggregate because competition demands it. The ante goes up.

AI may make product development more efficient while simultaneously risking competing the margin right out of product P&Ls!

In the short term, you may see some relief. You don’t backfill every open role. You slow hiring. You get more output from smaller teams. So R&D as a % of revenue might dip.

Temporarily.

But underneath that, three things are happening:

Over time, those forces tend to offset any initial savings. And in many cases, they increase total R&D spend. In fact, CFOs are coming to grips with the fact that profit margins may get compressed from 35-40% to 15-25%. (AI is already compressing SaaS gross margins from 75-90% to 50-60%.)

Worse, left unchecked, AI risks blowing up CAC payback. CAC payback is driven by one thing: How much gross profit you generate per customer, per period. So when gross margin drops, your gross profit per customer drops, and your payback period increases. Even if revenue stays the same.

It’s not overly dramatic to say that a single “efficient” AI feature can put the entire business model at risk.

Expect CFOs to scrutinize product roadmaps more closely.

Corollary: This is upending how software companies were traditionally valued. Tomasz Tunguz argues that Gross Profit per Token will be the new valuation metric. Take note if you’re in a startup.

Before AI, the cost of a feature was mostly fixed: Build it. Maintain it. Done.

Now? Cost scales with usage. Every time a user interacts with an AI-powered feature, tokens are consumed, models are called, and systems are triggered. That means if adoption doubles, your cost may more than double.

In other words, success = higher cost.

So in your product proposals, you need to answer:

If you can’t model that, your product won’t be profitable, no matter how much your customers love it.

It’s no longer enough in product proposals to talk about value in terms of improving conversion or engagement. You now need to pair it with factors like cost per query, cost per workflow, or cost per customer.

For example, instead of:

“This AI assistant will improve onboarding.”

You need:

“This assistant costs $0.08 per session and improves activation by 6%. At our volume, that yields $X incremental revenue at $Y cost.”

That’s how execs will evaluate it. Not as a feature. As a return profile.

If cost per user goes up, price or revenue per user has to follow.

Consider introducing premium tiers for AI features, and usage guardrails, usage-based pricing, or limits with paid expansion.

If you don’t do this, you’re compressing margin with every new feature.

Every serious proposal should include:

This is new discipline for Product.

Today, most roadmaps still think in engineering capacity and team allocation. Instead, you now have at least three constrained resources:

And they trade off. Two features might take the same engineering effort but one could be 10x more expensive to run. If you ignore that, you’ll make bad bets.

Every major initiative should answer:

If it pushes payback from 15 → 22 months, that’s not a small detail. That’s a strategic tradeoff.

Product teams have always lamented not having the license to experiment. The cost of development and the need to compress CAC payback periods made it difficult to justify “experimentation” as a line item in the budget. (Agile didn’t solve this.) As much as execs say they want to encourage learning, when it comes time to put the money on the table, they want to fund “wins,” not experiments.

AI gives product leaders an opportunity to be more strategic in their product plans. In budget proposals, consider separating Cost to Learn (experiments) from Cost to Win (at scale).

The latter is tied to short-term revenue targets (and margin efficiency). The former is tied to longer-term outcome bets (differentiation and growth).

This is where the strongest product leaders will separate themselves. Instead of defending features, roadmaps, and headcount, you start framing:

These align directly with how finance is thinking about AI.

Before AI, the question was:

“Can we afford to build this?”

With AI making build affordability next to zero, the question will become:

“Can we afford to run this at scale?”

That’s a fundamentally different question.

You should be aligned with your CFO on acceptable payback ranges, margin floors, and where AI investment is worth margin tradeoffs. Otherwise, you’ll end up in reactive conversations later.

A strong product proposal now includes:

That’s what makes it investable. And defensible.

AI isn’t just making building cheaper. It’s making running the product a core economic decision.

That means product leaders now own a much bigger piece of the P&L equation.

CFOs everywhere are grappling with the impact of AI on the P&L. This is happening right now. If you don’t bring that into your plans, Finance will bring it into the conversation for you.

And you may not like how that goes.

That’s all for today.

Have a joyful week, and, if you can, make it joyful for someone else too.

cheers,
shardul

Shardul Mehta
I ❤️ product managers.

​Follow me on LinkedIn