Edition #3. This week I built two things: a tool that calculates how replaceable your job is, and an AI agent that does my full bookkeeping. I also launched my company AI Agent. Beyond that: Google launched Gemini 3.1 Pro, OpenClaw was acquired by OpenAI, and Anthropic launched Claude in PowerPoint.
AI Agent is founded
I founded AI Agent. We help people and companies build and implement AI agents through training, advisory and hands-on implementation. No PowerPoints about the future of AI, we just build it. If you're not technical but want your own AI agent that can do everything, get in touch.
BenIkVervangbaar.nl: how replaceable is the average job?
I built BenIkVervangbaar.nl, a tool where you enter your LinkedIn profile and get an honest analysis of how replaceable your role is, per task, with a timeline, and concrete tips to stay relevant. The first results are in and the average sits at 78 percent. A nurse scores 20 percent, an accountant 88 percent, a software developer 87 percent. The more people do the check, the sharper the picture gets, because all results are anonymously collected and you see live statistics from everyone participating.
And that bookkeeping agent
I don't have an accountant, because everything that can be done systematically can be taken over by AI Agents. So I built my own AI Agent that handles my bookkeeping completely. That agent now creates invoices, processes receipts, generates VAT returns and exports audit files, all via the well-known chat apps, without an interface or browser.
The secret sits in two things: the AI Agent that learns from its mistakes, and a good backend. The agent is allowed to make mistakes (rarely happens), but those get caught by the validation built into the software. A wrong amount gets rejected, an invalid VAT number throws an error after which the agent corrects itself and learns from its mistakes, and that's how you build something you can actually trust.
Why this is possible now: OpenClaw
OpenAI hired Peter Steinberger, the solo developer behind OpenClaw, which stays open-source via a foundation and runs agents that are permanently active with memory, connected to your systems and reachable via Telegram, WhatsApp or Slack so they can execute tasks without you having to be present.
That's the fundamental difference with chatbots: a chatbot answers questions while an agent executes tasks in your systems, and those two things sound similar but in practice are a world apart.
Google launches Gemini 3.1 Pro
The numbers: released yesterday and it scores 77.1 percent on ARC-AGI-2 where Opus 4.6 sits at 68.8 percent, 94.3 percent on GPQA Diamond, and first place on SWE-Bench for agentic coding, which adds up to first place on 13 of 16 benchmarks and the signal that Google is seriously back in the race.
Sonnet 4.6: Opus quality at a fifth of the price
Sonnet 4.6 just dropped and beats Opus 4.6 on the tasks that matter most for businesses. Office tasks scores 1633 against Opus's 1606, and on agentic financial analysis 63.3 percent against 60.1 percent. Only on computer use is it nearly equal: 72.5 against 72.7 percent.
Two weeks ago Opus was the best model in the world. Now it's being overtaken by a model that costs a fifth of the price. The race is no longer about the biggest model, but about specialized performance at lower cost, and for everyone deploying AI agents this changes the math.
Claude in PowerPoint
Anthropic launched Claude in PowerPoint as a research preview where you install it as an add-in and Claude then works directly next to you in your deck. Build slides from your own template, make edits to selected slides, and generate native charts you can keep editing without static images or off-brand styling.
The real win
It's not about the hours you save but about the tasks that fully disappear, because my bookkeeping costs zero hours per month, not less than before, but zero.
