by Luis Rodrigues
What I’m building with AI + one hot take + four links worth your time.
The Take
My IKEA reskilling post went viral this week. But a friend sent me another side of that story, and it's messier than LinkedIn wants it to be.
Back in 2022, they launched a new chatbot. Instead of cutting 8,500 customer support roles, they decided to retrain them as interior designers. That created a new business line worth billions.
Salesforce made the opposite call.
They cut 4,000 positions in 2025 and went all in on chatbots.
Same technology, opposite strategies. One decided to produce more output with the same number of employees, while the other reduced employees and kept output.
One increased revenues, the other cut costs.
Almost one year later, the data is out. Salesforce leadership admitted publicly they were "too confident" that AI could replace human judgment.
Forrester found 55% of companies that cut staff, citing AI efficiencies, now regret the decision. More than 30% spent more rehiring than they saved.
The new pattern is: Fire loud, rehire quiet.
IKEA is also not perfect; just recently, in 2026, Ingka and Inter IKEA cut 1,650 corporate roles due to declining sales (not AI).
The real lesson is not "reskill good, replace bad".
AI changes cost structures either way. Your choice is whether you use the margin to grow capacity or shrink headcount.
Each feature you and the state-of-the-art lab release gives your users the same choice you're making. Does the tool make them more capable, or do they feel they can replace you?
The Build Log
Last week I wrote about how thoughtled.ai analyses voice using structured traits instead of descriptions. This week, we look at a failure that even those structured traits cannot prevent.
The pipeline generated a post for one of our users that sounded exactly like him. Good sentence length, humor, and just enough emojis. His colleagues would have believed he wrote it.
The problem was that it started with "when I was scaling the engineering team of my previous startup...". This person never led an engineering team or worked in a startup.
The voice was perfect but with a hallucinated bio.
There's a big difference between how someone writes and who that person is.
The voice profile captures patterns. It doesn't know your job, your company, nor your real-world experience.
Without access to this information, the generation agent will complete the post with plausible-sounding data.
Plausible-sounding fabrication is much worse than incorrect content, as there's a higher likelihood it will be published.
The solution is not complicated. LLMs are good at respecting the data in the context, so if we collect profile data from LinkedIn and pass it to the generation and critique agents, the problem is solved.
The main work is done by the critique agent, which checks whether any generated content mentions experiences or positions that contradict the person's actual background. If it does, those sections need to be rewritten.
We quickly learned two interesting things. First, LinkedIn profiles almost always sound amazing.
"AI Visionary | Helping Founders 10x" is marketing speak and not real facts. So we treat company and position as hard signals and about and headline as soft ones.
Second, people know things about themselves that aren't listed on LinkedIn, so we allow them to edit the context we found there.
The principle is simple: voice and identity are separate inputs. Never mix them.
If you're using AI for content generation, ask yourself: does the model know enough about you to generate exactly what you need? If it's not in context, you're trusting it to infer your biography, and it will do so confidently.
On My Radar
Amazon jailbroke the model, flagged it to the US government, and access vanished for all of us.
Kimi generated the same 12 high-quality landing pages as Fable, but at 94% less cost. Open-source models are approaching frontier quality on real-world outputs.
Faros tracked 22,000 developers and saw the per-developer defect rates rise from 9% to 54%. Coding agents made code production easier; now the difficulty shifted to code review.
OpenAI rewrote ChatGPT's memory system, turning past chats into a sorted user profile and boosting factual recall from 41.5% to 82.8%
What are you building?
Got something you built? Hit reply, the best ones get featured next week.
Know someone who needs to write better content? Forward them this.
New here? Someone forwarded this to you? → Subscribe to Build What Matters: https://luisrodrigues.ai/
How was this issue?
