The last mile of ROI: why AI works for the developer and vanishes at the company

Uber burned its entire 2026 AI tooling budget in four months. Microsoft, a major backer of Anthropic, started canceling internal Claude Code licenses because one division consumed so many tokens that it blew through the area’s annual AI budget in a matter of months. When two of the most engineering-mature companies on the planet hit the cost ceiling at the same time, the whole room starts asking the same thing: what if this is a bubble about to pop?

I think that is the wrong question. Not because the cost is not real, it is, and it is measured. It is wrong because “bubble” implies the technology is fake and about to deflate. That is not what the data shows. What the data shows is more uncomfortable: the technology delivers, but the value leaks somewhere between the developer who got more productive and the result that was supposed to show up in the company. That leak has a name, and I am going to call it the last mile of ROI.

This post is about where the money disappears in that last mile, with real data, and about what separates those who capture value from those who only pay the bill.

The wrong debate

The internet loves two extremes: “AI works and will replace everyone” versus “AI is hype and will pop”. Nobody serious defends either one. The honest debate starts when you swap that false dilemma for two concrete questions.

First: does the value AI generates beat its total cost? Total cost, not subscription. Tokens, security risk and the friction of good people walking out all go into the bill.

Second: is whoever pays the bill actually capturing that value? Because, as you will see, the gain that shows up in your code does not always show up on the balance sheet of the company that bought the license.

These two questions have different answers, and that is exactly where the confusion lives. The answer to the first is “depends on who”. The answer to the second is “almost nobody”. Put them together and you understand why 7 in 10 companies are frustrated with AI while the developers who know how to use it will not let go of the tool.

The bill that never stops climbing

Let us talk real money first. Bill Staples, GitLab’s CEO, told customers that the per-developer tooling bill went from tens of dollars a month to hundreds, and is heading toward thousands. The reason is purely technical. An AI agent is not an autocomplete suggesting the next line. It is a process that opens several requests per task, reads files, runs tests, iterates all night and ships whole features. Every one of those iterations burns tokens.

Gartner reports that 29% of organizations already spend between US$200 and US$500 per developer per month on tokens alone, and that power users go past US$2,000 a month. I know people who got close to five figures in a single month. And the detail that breaks the business model: an NVIDIA VP, Bryan Catanzaro, said the most honest sentence in this entire debate. For his team, the cost of compute is already higher than the cost of the employees. Read that again. The machine got more expensive than the developer itself.

That is exactly what blew up the unlimited subscription model. GitHub paused new paid Copilot sign-ups in April 2026 because flat-rate unlimited pricing no longer closes when the agent consumes without limit. GitLab changed how it charges for a seat: it stopped being a unit of value and became a unit of task. The cost went up, that is a fact. I had already touched this wound in The Invisible Cost of Wasted Tokens, and what was an efficiency problem turned into a business-model problem.

The agentic paradox: cheaper token, bigger bill

Here the math seems not to add up. The price of a token is plummeting. Gartner itself projects that running inference on a one-trillion-parameter model will cost over 90% less by 2030. And yet the company bill is going to rise. Why?

Because consumption grows faster than the price falls. Goldman Sachs estimates that agentic usage will multiply token consumption by 24 times by 2030, reaching quadrillions of tokens per month. Cheaper token, bigger bill. That is the central paradox of the agentic era, and it is not new: it has had a name since 1865.

Jevons paradox. When a technology becomes more efficient, demand for it grows instead of shrinking, and total consumption rises even with a lower unit cost. William Jevons observed it with coal in Victorian England. Satya Nadella and Dario Amodei invoke the same principle for tokens today.

To ship quality code with an agent you need a harness that spends more tokens, not fewer: more context, more tests run, more verification. The better you want the result, the more you consume. Cheaper tokens are not relief on the invoice, they are an invitation to spend more. Anyone waiting for the bill to drop over time read the chart backwards.

The other side of the number

A high cost is not a problem if the return is higher. The problem is that the return stalls. In 2026 WRITER ran an AI adoption survey with 2,400 workers, half C-level, half employees. The result: 79% of companies admit facing serious adoption issues, and only 29% see significant ROI from generative AI. Seven in ten putting money in without seeing a clear return.

And when you swap the ruler from “executive perception” to “measured impact on the books”, the number collapses. The MIT NANDA report, “State of AI in Business 2025”, measured real P&L impact and concluded that 95% of generative AI pilots deliver no measurable financial impact. Roughly 5% deliver.

Ruler	Who measures	Success
ROI perceived by the executive	WRITER (2,400 workers)	29%
Measured P&L impact	MIT NANDA (State of AI in Business)	~5%

Notice these are different methodologies, with different severities, which is why the numbers differ. But both point to the same hole. When a source measuring perception and one measuring the balance sheet reach the same conclusion from opposite directions, something structural is happening. And here is the point that changes everything: if such a small fraction gets a real result with the same technology that is available to everyone, then the ceiling is not the model. The ceiling is how much the company is willing to change to get there.

Where the money leaks

Now the pivot of the whole post. The same survey that says only 29% see ROI also says superusers save almost nine hours a week and are up to five times more productive. Ninety-seven percent of executives report a clear individual benefit. In other words: the senior dev, the staff, the tech lead who knows how to use AI has a huge, measured, undeniable gain.

Stop and look at the contrast. The gain at the person level is enormous and measured. The gain at the company level does not show up, or takes a long time to. There is only one possible explanation for that gap. The money is leaking on the way between the individual’s gain and the organization’s result. That is what I call the last mile of ROI.

The Last Mile of ROI. The distance between the productivity AI generates in the developer and the value that reaches the company’s result. The first is short and measured in hours saved. The second is long and full of friction: integration, process, governance, culture. Almost all the promised ROI is lost on that stretch, and no new license shortens it.

MIT’s own diagnosis is surgical: the problem is not model quality, it is integration and the learning gap of the tool inside the organization. The model is the truck. The last mile is the potholed road between the warehouse and the customer’s door. You can buy the fastest truck in the world and the package still does not arrive.

Harvesting the fruit without planting the tree

Why does the individual gain not become the company’s gain? Because turning productivity into real ROI requires three things nobody wants to pay for: process redesign, governance and culture change. None of the three can be bought with a credit card or solved with hype.

What most companies do is the opposite. They buy a license, hand it to the devs, wait for magic and chase the same deadlines as before. That is harvesting the fruit without planting the tree. The tree is the boring work: redesigning the review flow, defining where the agent can and cannot act, training the team, rebuilding the harness, changing how the team measures success. Without that, the tool amplifies the chaos that was already there instead of generating new value.

And it is exactly the part nobody wants to touch, because there is no shortcut. You do not skip the hard part. Whoever skips it ends up with the token bill and no return, then concludes “AI does not work”. AI works. What was missing was planting the tree.

The real bubble is firing before you know how to earn

There is one behavior in the market that I would call a bubble. Look at the contrast. Sixty-nine percent of companies report layoffs attributed to AI, but 39% admit having no formal strategy to generate revenue from those tools. Translation: they are firing before knowing how to make money with the technology. That makes no sense at all.

And Gartner buries the argument that cutting generates return. In a survey of 350 executives, 80% reported workforce reductions, and there was no correlation whatsoever between cutting and having more ROI. The companies that cut the most are not the ones that profit the most. Forrester went further: 55% of companies that laid off citing AI already regret it, because they lost institutional domain knowledge, senior people, and gained rework and rehiring costs. Sam Altman gave part of this a name: AI washing, using AI as an excuse for a cut that was going to happen anyway. Out of more than a hundred thousand layoffs analyzed in one cut, only about 7% actually cited AI.

The GitLab case is the perfect portrait of the contradiction. A profitable company, growing 16%, with US$220 million in free cash flow, and still it cut 7% of headcount wearing the narrative of the agentic era.

The layoff narrative	The reality in the data
Cutting headcount frees margin and proves AI maturity	No correlation between cuts and ROI (Gartner, 350 execs)
AI replaces the senior	55% regret it over lost knowledge (Forrester)
The market rewards whoever cuts early	GitLab stock fell over 8% on announcement day
It is an efficiency decision driven by AI	Much of it is AI washing, a cut that was coming anyway

Notice the last line of the reality column. GitLab’s stock fell over 8% that same day. The market that used to reward these cuts started to distrust them. We are entering the reality phase, and it is less generous with the easy narrative.

The total cost nobody adds up

There is a cost of AI that shows up on no invoice, and it is the cultural one. Fifty-four percent of executives say adopting AI is tearing the company apart from the inside. Twenty-nine percent of employees admit sabotaging the company’s AI strategy, a number that rises to 44% among the youngest. And 67% report that data has already leaked through some unapproved AI tool.

Add that to the token. The total cost I mentioned at the start is not just inference. It is tokens, plus security risk, plus the friction of good people leaving because adoption was imposed without process. When you add it all up, you understand why, on average, today the cost beats the captured return even in places where the technology technically works. The problem is rarely the tool. It is everything around it that nobody put on the spreadsheet.

Eight months to the return

If the last mile is so expensive to pave, is it worth it? Google’s data, through the DORA metrics, says yes, with patience. Their model for an engineering organization estimates a payback of about eight months for the AI investment to pay for itself, and a return in the region of 39% in the first year for those who get through that curve.

Eight months doing the work that slows you down in the short term: training people, removing bottlenecks, paying down technical debt, fixing process. It is the classic J-curve. You get worse before you get better. Most give up in the valley of the curve, conclude it did not work and go back to chasing deadlines. Whoever endures the valley is the one who reaps the return on the other side.

And this is where the harness matters. I already argued in From SonarQube to Agentic Process that a metric becomes a reward function when the agent optimizes against it, and in Rat in the Maze that the agent’s performance depends more on the maze than on the rat. Paving the last mile is exactly that: building the rails that turn individual gain into structural gain. Without rails, the developer’s productivity evaporates before it reaches the balance sheet.

Why there is no going back

Despite all of this, I have no doubt it is a one-way road. And that is not faith, it is what the data supports.

First, the individual gain is not perception, it is measured time. Nine hours a week is more than a full working day. For a senior, a staff, a principal, that is a superpower. Second, the market for AI coding tools more than doubled in a short time. People paying a lot and complaining about the price, but not letting go, because they see value. That is not the behavior of a bursting bubble, it is the behavior of someone who found leverage.

Third, the historical argument. By Jevons paradox, greater efficiency pulls greater demand, not mass unemployment. David Solomon, CEO of Goldman Sachs, wrote that the job apocalypse is overstated and that the economy adapts as it always has. Dario Amodei himself, of Anthropic, walked back his prediction that AI would eliminate half of junior jobs, reframing it through Jevons logic: if you automate 90% of the task, everyone moves to the remaining 10%, and that creates new work. A Harvard Business School study shows exactly this shift: a 17% drop in postings for automatable roles and a 22% rise in amplification roles.

Think about the electric car and the self-driving car. Why are they not everywhere yet? Because cities have to adapt, infrastructure has to change, the system has to train on every new context. But it is inevitable. Nobody is going back to the past. AI in development is on the same curve: adoption is uneven, painful, full of people climbing the wrong way, but the direction has no return.

What separates the 29%

So what do the 29% who capture value do differently? They treat AI as process redesign and people amplification, not as replacement. And technically, they have a harness.

To translate for those who do not follow my other posts: if your project has tests with good coverage, clear context, readable architecture and defined conventions, AI becomes an amplifier of those patterns. It applies the same convention across two hundred files without tiring. It writes more tests than a human would with the same patience. It propagates senior patterns into junior code automatically. A team with maturity and a good harness ships more code, cleaner, with more quality. A team without a harness ships more chaos, faster.

It is the same thesis I defended in Hallucination Comes From the Code, Not the Model: the vector that most changes an agent’s output is not the model, it is the engineering around it. The 29% paved the last mile. The other 71% bought the truck and left the road full of potholes.

Conclusion

Put it all together. The cost of AI is real and rising, in some teams it has already passed the cost of people. The return exists, but it stalls at the individual level and does not climb to the company, because process, governance and culture are missing. Firing is not a source of return, most of the time it is the opposite. And even with all that, there is no going back, because the base value is real and the market has already decided.

That is why the bubble metaphor fails. A bubble pops and disappears. What is happening is something else: an adoption curve most are climbing the wrong way, paying the bill without paving the last mile. AI is not going to deflate. The ones who fall behind are the ones who thought they could skip the homework, and they will keep paying for tokens while complaining the technology does not deliver.

Now I want to hear from you, the one living this firsthand. Is AI amplifying your team and your work, or has it become just an excuse for cuts and more pressure? The answer tells you which side of the last mile you are on.

References

WRITER (2026). “Enterprise AI Adoption Survey 2026”. 2,400 workers. writer.com/blog/enterprise-ai-adoption-2026
MIT NANDA (2025). “The GenAI Divide: State of AI in Business 2025”. Coverage: fortune.com
Staples, B. / GitLab. “GitLab CEO sees developer tool bill increasing 100-fold”. infoworld.com
Gartner (2026). Inference cost projection for a 1-trillion-parameter LLM by 2030. gartner.com
Goldman Sachs (2025). “AI agents forecast to boost tech cash flow as usage soars”. goldmansachs.com
Catanzaro, B. / NVIDIA. “Cost of AI is greater than cost of employees”. fortune.com
The Next Web (2026). “GitHub pauses Copilot sign-ups over agentic AI usage”. thenextweb.com
Gartner (2026). “AI layoffs may create budget room but do not deliver returns”. 350 executives. gartner.com
Forrester (2026). “The AI layoff trap: why half will be quietly rehired”. hrexecutive.com
Altman, S. (2026). “AI washing” and layoffs. fortune.com
The Next Web (2026). “GitLab layoffs and the agentic era”. thenextweb.com
WRITER / Workplace Intelligence (2026). “AI adoption is tearing companies apart”. hrgrapevine.com
Google DORA (2025). “The ROI of AI-assisted software development”. cloud.google.com
Solomon, D. / Goldman Sachs (2026). “AI job apocalypse is overblown”. webpronews.com
Amodei, D. / Anthropic (2026). Walking back the prediction on junior jobs. fortune.com
Harvard Business School (2026). “How AI is changing the labor market”. hbr.org
NPR Planet Money (2025). “Jevons paradox and the economics of AI”. npr.org
Yahoo Finance (2026). “Uber burned its entire 2026 AI budget”. finance.yahoo.com
Windows Central (2026). “Microsoft cancels Claude Code licenses”. windowscentral.com