Companies replaced people with AI and are now quietly rehiring them. On why the last 20% of the work still needs humans, and what "good enough" is doing to SaaS.
The promise of instant productivity has been reshaping the SaaS industry, leaving professional casualties behind. What looks like efficiency in a quarterly earnings call often turns out to be deferred cost; in rehiring, in technical debt, or in customer trust, a year or two later.
SaaS stands for software-as-a-service, the rentier model where you subscribe to your favourite piece of software instead of buying it once. By ecosystem I mean that these services, more often than not, rely on each other, in a symbiotic relationship. For example, a blog site may use a translation service, or a shop processes payments with an established vendor. But what happens when teams are given time and tokens to develop such services internally? A poisoned apple spoils and spreads through the bunch.
The rehiring wave nobody wants to share
Companies convinced themselves that AI could handle the judgment-heavy work their support and content teams did. "Why should we keep humans around when the model does some of the work?"
Klarna stopped hiring for more than a year and cut staff by 22% after its AI chatbot handled 2.3 million conversations in 23 markets in a single month.
IBM replaced around 8,000 HR workers with its "Ask HR" AI.
Commonwealth Bank of Australia cut 45 customer service jobs claiming its AI voice bot had reduced call volumes by 2,000 a week. Calls rose instead, managers were drafted onto the phones, and the bank apologised and offered every job back.
Then the reversals started. Forrester's 2026 Future of Work report found 55% of companies that laid off staff for AI now regret it.
"Artificial intelligence is a fantastic tool, but it's only as good as the information you use to train it."
Charles Poon, Ford
The case studies where lessons were learned
Klarna and Duolingo saw real short-term gains. The problem is that they only measured what was easy to track. Klarna's assistant handled status updates and simple disputes well. Duolingo's GPT-4 system generated language exercises at a fraction of the cost of professional translators.
Both CEOs publicly sold the AI-first story as a competitive edge. Then the quality gap showed the cracks in the system. Klarna's CEO Sebastian Siemiatkowski acknowledged it directly:
"Cost unfortunately seems to have been a too predominant evaluation factor... what you end up having is lower quality."
Sebastian Siemiatkowski, CEO of Klarna
Klarna started rehiring people to handle the conversations the bot couldn't. Duolingo users noticed a drop in translation quality. When cancellations increased, the CEO doubled down with an AI-first message that drew hundreds of critical comments and later backtracked, saying he "did not give enough context".
There's a footnote worth keeping just for the irony: Duolingo's founder, Luis von Ahn, was involved in inventing CAPTCHA, built specifically to tell humans apart from machines.
Adobe spent a decade treating captive Creative Cloud subscribers as a stable revenue base, then leaned into generative AI by revising its terms of service in ways that appeared to grant it rights to train models on customer stock work, especially to customers that would sell on Adobe Stock. The clarification came too late to stop the perception that Adobe extracts from its userbase rather than serve it. Instead of competing on product quality, it tried to buy its way past Figma for $20 billion, lost the deal to EU and UK regulators along with a $1 billion breakup fee, and was left with no competitive design tool. Adobe's stock has lost roughly two-thirds of its value since. Locking users into a subscription while using AI to cut costs rather than improve the product works only until your audience feels alienated.
Not every company felt an ominous boomerang, however. At Shopify, AI now writes more than half of the company's internal code. Their revenue grew 34% in Q1 2026 while headcount declined, pushing revenue per employee to around $1.63 million. Still, even Shopify's own reporting calls the productivity evidence "genuinely mixed," leaving the approach entirely at management's discretion. It's less proof that AI-driven choices are risk-free, and more a case of preparation beating panic.
The 80/20 rule
AI tools produce decent first drafts fast. For managers with access but limited hands-on experience, "80% done" feels like a real handoff rather than a rough start. The team receiving the work hears "AI did the heavy lifting, just clean it up", which makes the remaining work sound like polishing. To no one's surprise, it isn't.
The 80/20 rule, or Pareto Principle, is usually cited as "80% of value comes from 20% of effort." In software, the mirror version matters more: the last 20% of a feature takes 80% of the time. That's when the polishing phase of a project begins. Defining error handling, edge cases, accessibility, security audits, and unknown unknowns that may cause breaks in production.
Research published in early 2026 found that after teams adopted AI coding tools extensively, code duplication rose by 48%, refactoring decreased by 60%, and technical debt increased by ~41%.
"The flow-debt trade-off arises when seamless code generation occurs, leading to the accumulation of technical debt through architectural inconsistencies, security vulnerabilities, and increased maintenance overhead."
Flow, Technical Debt, and Guidelines for Sustainable Use
The death of SaaS
A friend who works in SaaS put it well: there's growing internal pressure to build tools in-house rather than buy them, because a capable-enough AI model removes the need to shop for a specialised SaaS product.
Some examples:
An automated translation layer for multilingual publishing, instead of paying a subscription for an internationalisation service.
A homegrown content management system to bridge the gap between sales and marketing, instead of finding a suitable platform.
An internal project management tool rather than a licensed one.
None of these replacements is as good as the dedicated products they're displacing. But they're "good enough", and good enough is cheaper than a recurring bill, once a team has Claude Code and a few days to spare.
A SaaS is a customer too: it pays for hosting, analytics, a payment provider. Every cancelled subscription starves a vendor, which then cancels subscriptions of its own, and the web gets a little thinner. This isn't hypothetical: a third of enterprises have already replaced at least one SaaS tool with something built in-house, and Sanofi is cutting 80% of its ServiceNow usage with agents built on Claude Code. Meanwhile the homegrown replacements rot the way anything built to 80% rots, with no vendor left to fix them and, in time, no product left to crawl back to. The money that used to feed thousands of vendors now feeds Anthropic, OpenAI, and Microsoft. The line goes up, over the corpses of many decent products.
Line goes up
LLM "labs" (Anthropic, Google, OpenAI, and the rest) need ongoing investment, attention, and the promise that the next breakthrough is close. The claims have been consistent: Sam Altman's June 2025 essay "The Gentle Singularity" declared the AGI takeoff was already underway. Dario Amodei wrote that "exceptionally powerful AI" could arrive as soon as 2026:
"Smarter than a Nobel Prize winner across most relevant fields."
Dario Amodei, CEO of Anthropic
The benchmarks tell a quieter story. Each new AI model is only getting a little better at standard tests, moving from 86% to ~90% correct on benchmark tests for coding and knowledge.
An arXiv paper titled "The AI Scaling Wall of Diminishing Returns" laid out the mathematical case. A January 2026 MIT paper on low-budget models found that exponentially increasing inference compute produces diminishing benchmark gains.
Yann LeCun, a Turing Award winner and one of the fathers of deep learning, called LLMs "a dead end" architecturally. Thomas Wolf of Hugging Face pointed out that Nobel-level breakthroughs come from asking new questions, not answering known ones, which is what current models are trained to do. Sam Altman has since quietly admitted that a new architecture is needed.
The AGI deadline has moved so many times that it reminds me of the decades-old promised theory of everything, pushed by string theorists to mould known mathematics and physics theories to make their model work. It's a noble cause, but it ends up taking funding from other theories, due to its sheer popularity. It's sexy, it sells.
Switching to a local model
The move toward AI that runs on your own phone or laptop, instead of some distant server, tells you something. The cloud-everything approach isn't working, not financially and not technically. Running AI at scale in remote data centres is expensive. Waiting on a response from a far-away server slows things down for whoever's using the product. And if your product always depends on someone else's servers, you don't really own the edge over competitors. You're just renting it.
At WWDC 2025, Apple shipped an AI model that runs entirely on your iPhone, with no external server needed. That was only possible because Apple has spent 15 years building its own chips specifically for this kind of task. In 2026, it went further, running a significantly more powerful model locally on the iPhone 17 Pro and newer Macs.
NVIDIA is doing something similar for PCs. Its DGX Spark, a small desktop machine, can run a 120-billion-parameter model entirely offline. Qualcomm and Samsung are moving in the same direction on mobile chips.
And while American labs sell subscriptions, Chinese labs (DeepSeek, Qwen, Kimi, and GLM) give the models away under MIT or Apache licences, allowing you to run them on your machine. Qwen alone was downloaded over one billion times and overtook Meta's Llama as the most used open model family.
What actually works
AI genuinely helps with first drafts, summarising, routing simple requests, running specialised skills like UX and accessibility analysis, and catching basic bugs. It's not good enough for the last 20% of any feature. Hell, hardly for the first 80%. That's what we're here for.
The companies rehiring laid-off staff aren't giving up on AI. They're recalibrating it as a helper rather than a replacement, as serious practitioners said before the cost-cutting trend took over.
"AI is more than up to the task. What these companies discovered is that they weren't. They skipped the hard part."
Eric Vaughan, IgniteTech
Maybe it's time to invite back the golden apples, the people who actually dedicated their careers to the last 20%.
There is comfort in the chaos of human craft.
Hail Eris.
Originally published at jlopes.eu.
Top comments (0)