Read time: 4 minutes

Leader, leader, welcome back!

A lot happened this week. I went deep on it so you don't have to.

Two years ago, AI for finance meant a chatbot that summarised a P&L.
The systems we now build with finance teams do real accounting work, end to end.

Picture this.
A controller types into Slack: "Did we accrue for the legal fees this month?"
The system reads the message, understands it is a close request, checks NetSuite for evidence of the accrual, finds none, drafts the journal entry, posts it back in the thread for review, and only writes it to the GL after explicit approval.
Audit trail intact, transaction linked.

The controller spent 60 seconds doing what used to take 30 minutes. The system did the rest.

The hard part of building these systems is not the AI.
The hard part is knowing which pieces of the workflow are AI's job and which pieces have to be code. Get that split wrong and you either end up with errors in the GL or a tool nobody trusts.

The pattern that works:
AI handles language and judgment.
Software handles controls and execution.

AI reads the messy Slack message, figures out the intent, drafts the conclusion, summarises the evidence in plain English.
Code validates account IDs, checks subsidiary mappings, prevents the same request being processed twice, records the audit trail, posts the journal entry through the API.

Neither layer pretends to do the other's job.

The same logic applies to your function.

The work your team keeps redoing every month (accruals, intercompany matching, variance commentary, vendor follow-ups) is full of judgement-heavy steps that AI handles well and control-heavy steps that have to be deterministic.
Most teams either over-trust AI (let it post directly) or under-trust it (use it only for summaries). Both miss the leverage.
The leverage sits in designing the split, then building the system around it.

The 5-minute action: list the 5 most repetitive monthly tasks your team does.
The split between AI and code lives inside that list.

The bigger shift sits underneath.
A workflow like the one above used to take a six-month software project.
It is now buildable in days by people who can describe the process and orchestrate the right tools.
The bottleneck is no longer engineering. It is knowing what to automate and how to design the controls around it.

That is the work Prime AI Solutions does. If your function is sitting on recurring work that should have been automated by now, reply to this email and we'll have that conversation.

When I started using Claude properly, I was hitting my usage limit by 10am.
Every single day.
I assumed I was just using it too much, so I'd close the tab and wait it out!

It took me longer than I'd like to admit to realise the problem wasn't how much I was using it. It was how!

Some tips below:

1. Pick the right model. Opus burns through usage 3 to 5x faster than the others. Most tasks do not need it. Haiku handles quick jobs. Sonnet handles most of your day. Opus is for the things that genuinely need heavy thinking. I default to Sonnet 4.6 for about 80% of my work.

2. Watch your thread length. Once a chat hits 100+ messages, Claude reprocesses the entire thread every time you send something. I had threads running for weeks, burning through credits with every message.

3. The reset trick. End every long chat with:
"Summarise this conversation in a way that lets me pick it back up in a new thread with full context."
Paste the summary into a fresh chat. Same context. A fraction of the tokens.
Even better, use Claude Projects for ongoing work and let project knowledge do the heavy lifting.

4. Use Skills. Skills are reusable instruction sets. You write them once, save them, and Claude triggers them automatically when you call them. I have one for newsletter writing, one for prospect research, one for blog posts. The blog Skill alone saves me 40 minutes per post because I'm not re-explaining the brief every time.
Setup: click Customize, then the + button, then Create Skill. Three parts.
A description that tells Claude when to use it, the actual instructions, and 2 to 3 examples of the output you want.

This is exactly the kind of system thinking and tips I help you with.
Just reply and let me know where I can help 🙂

Google I/O dropped a lot. Most of it was meh.
Gemini Omni Flash is their "any input, any output" model and the video generation is genuinely sharp. Gemini 3.5 Flash is fast but ships with a January 2025 knowledge cutoff, so it doesn't know what vibe coding is. Antigravity is their Codex clone, fine but nothing new. Gemini Spark, their 24/7 personal agent across Workspace, is "coming soon" so nobody knows if it actually works.
TLDR: stay on Claude or Codex for code, look into Spark when it ships, ignore Antigravity for now.

Granola Briefs launched. Granola (my meeting notes tool of choice) now searches your email, web, and previous meeting notes before a call, then hands you three bullets to walk in with. Genuinely useful if you live in back-to-backs.

The 100x to 1800x cost gap. Zapier's CEO Wade Foster open-sourced AutomationBench, which tests cheap vs expensive AI models on real production automation. Expensive models cost $1.80 to $6.31 per task. Cheap ones cost $0.003 to $0.02. For most workflows, the cheap models perform fine. Run 10,000 tasks a day on the expensive tier and that's $18,000. Cheap tier, $200. Worth a routing audit.

Microsoft Edge Copilot got serious. Reads across all your open tabs, generates podcasts from articles, builds quizzes for learning mode, and now uses your browsing history for context. Worth a look if you live in Edge.