The skill that matters is shifting

Knowing which story to tell is the most important skill.

Thanks for reading issue two. Every month you'll get the same three things: what mattered, one workflow with the prompt, and one thing we built. If you have any feedback, please let us know.

Cole and the team at North Group

The three minute version

Dashboards just got easier than they have ever been. This month, get everyone on your team using AI as more than a search engine.
The cheaper models got a lot better. Haiku 5.5 beats June's Sonnet 5 on Anthropic's own tests.
Hiring and pay decisions made with Claude need a person to sign off. Anthropic's updated policy starts November 12.
In-person training is impactful. Two afternoons with a client this week, solving problems in the moment.

What mattered

There's so much news to keep up with in AI that it's fatiguing. A lot of different models have come out since the last issue, like Gemini 4 Argon and Claude Haiku 5.5. Haiku 5.5 is the one I'd pay attention to. It's Anthropic's cheap, fast model. On Anthropic's own numbers, it now beats June's Sonnet 5 on a few of their tests, at a fraction of the price. The lower-end models are getting insanely better and even more efficient. For a lot of everyday work, they're what most people should be using now. Work you'd have handed to Sonnet a few months ago can go to Haiku.

Data visualization and storytelling just got easier than they have ever been. Anthropic released Claude Dashboards and Motion on October 8, and Motion handles the storytelling side.

If you've built reports in Power BI, you know the painful routine. Hours on data models, DAX and formatting, and the result still looks like every other report. Now you connect your data, ask a question in plain English, and Claude builds the dashboard. You can click any number to see the query behind it.

My reaction when it launched was pure excitement. Dashboards have been the biggest thing since sliced bread over the past couple of months. Anthropic's own help page says to open a chart's query and check it's counting what you meant. The skill that matters is shifting towards knowing what question to ask and how to tell the story in the data.

We've somehow hit the point where people are fighting over which model is best again. Meanwhile, who onboarded their AI assistant properly? Most people still type "just help me out" and wonder why the answer is so generic.

We've done a lot of trainings. How quickly people have ramped up and started using Claude has been insane. People who were technologically challenged before, and got overwhelmed every time they had to learn a new system, have really taken this on. Ramping up on a legacy system like Excel is so much harder.

No one is behind. It's very new. This month, get everyone on your team to a baseline, using AI as more than a search engine. Give it a spreadsheet and ask it to turn that into a document or a presentation.

Since the last issue

If you run Claude for a team

Start with one team. On Enterprise, Dashboards is off until an owner turns it on, and can be limited to groups.
Check how your warehouse connection signs in. Anthropic says a shared dashboard uses each viewer's own access, but one shared login reaches whatever it can access.
Keep it for quick questions. Anthropic pitches it alongside BI tools, on warehouse or Salesforce data. Spreadsheets aren't listed. For those, use the prompt below.

Also this month

Recording calls or meetings? Check your consent policy. Canada's Privacy Commissioner says businesses covered by PIPEDA must follow the Act when recording calls, whoever placed the call. Amazon's new Bee wearables record the whole time they're on. Make sure your policy covers AI note-takers (Privacy Commissioner, October 8, CNBC, October 15).
Anthropic says hiring and pay decisions made with Claude already need a qualified human reviewer. Its updated usage policy, effective November 12, now lists shifts and performance too. The person affected must be told AI was used, and customer bots must say they're AI (Anthropic, October 8).
Ask who can reach your coolers. Hackers took over fridges in hundreds of French stores through old, often unsupported controllers. If a contractor runs your coolers, HVAC or point of sale remotely, ask if it's reachable online, still updated, and who signs in (CNN, October 14).
Just interesting: Google says Gemini 4 Argon, in rare instances during training, "attempted to escalate container privileges to alter grading scripts" and once "crashed the evaluation sandbox to avoid getting scored for a failed run" (Google, October 13).

Run this

Turn last month's export into a dashboard that shows how Claude got every number. Use it on sales, invoices, expenses or jobs.

Below is last month's export from one of our systems. Turn it into a dashboard in this order: what you read (how many rows, first and last date), the overall total, the count, the average, totals by category, totals by week, and the five largest entries. Under every number, add one line saying how you got it: which rows you added or counted, and which you left out and why. Use only what is in the export. If anything is missing, repeated, negative, inconsistent or looks like a total, do not guess or fix it quietly. Say what you did with it and list it under a final heading called Flags. Do not add targets, forecasts or comparisons with other months.

Paste it into Claude, then paste last month's export as rows. Take customer names and contact details out first.

Read the line under the total first. We ran the prompt five times on Claude Sonnet 5.5, using a made-up September invoice export with seven planted problems. All five runs flagged all seven. Four left a $1,240.00 credit out of the total and one took it off, and each said so in the line under the total. So two people running the same export can get two different top numbers.

Where it stops working, from those five runs

It made one slip in its working. One run said a category was "$20.00 below" another when it was $220 above.
It won't pick a date for you. All five would not decide whether "10/09/2026" meant September 10 or October 9.
It runs long. The five ran 952 to 1,112 words.

Something we built

Two afternoons in person

This week. Laptops open all afternoon.

With one client, we grouped people by department at their request, so everyone would get a say. One of us presented while the other walked the room for anyone who got stuck. I think in-person training is impactful because you can problem-solve in the moment and interact with people directly. It also puts a real human in the loop while people are learning.

The first project was built around an interview. Claude asks about your job one question at a time, then writes up what it learned as the project's instructions. We emailed the prompt to each room the moment we reached it.

The part you can copy: group people by team, keep laptops open, and have one person walk the room. Build each first project from a Claude interview, and add a line to its instructions whenever Claude gets something wrong.

If you know someone who'd benefit from this, forward it. They can sign up at northgroup.ai/summit.