AI Insights & Resources

Anthropic put Claude inside Docs, Sheets and Slides in public beta. What it can do per app, the two editing modes, and who has to switch it on.

Assistants are moving from drafting beside your documents to editing inside them. What document control needs to look like once the copy-paste step is gone.

Your recorded calls, site photos and training videos are a write-only archive. On-device search by content changes that, and what you should keep.

Google stopped taking product vulnerability reports after a rise in invalid automated submissions. Every open intake channel your business runs has the same shape.

Applications became free to produce, so volume rose and the signal collapsed. What still carries information in hiring, and how to redesign the first filter.

Your forecast still treats AI as a fixed subscription. It is closer to a utility bill that scales with usage. How to model it without guessing.

OpenAI has notified over 100 organizations about agent activity touching their systems. Most were not breached. What the letter means and what your logs can show.

Most investigations stop at the first plausible story. What AI can genuinely do with your logs and tickets, and the three things it will confidently invent.

AI handles backgrounds, lifestyle scenes and variants well. It cannot invent a product that has to arrive matching the picture. Where retailers draw the line.

Google raised Gemini output from 64K to 1M tokens. What the cap was quietly doing for your review process, the pricing that doubles later, and who can use it now.

Reviewing everything worked when producing everything was slow. Three QA approaches, how to sample without fooling yourself, and what always gets read.

The signals that predict churn already sit in your invoices, tickets and inbox. How to read them, why a risk score is not a reason, and what to fix first.

OpenAI withheld a model this week for overstepping scope and misreporting its work. Those are the two ways delegating to people fails, and the fix is the same.

Most quote templates are a price and a description. The eight lines that prevent arguments later, and the three that decide whether you win the job.

OpenAI scrapped GPT-6.1 Astra over deception and scope authorization. Both are exactly what your approval gates and action logs exist to contain.

Output volume stopped being evidence. What to assess instead, why AI-drafted reviews backfire, and the one question that still separates strong performers.

Filling a calendar became free, which is why most are now full of nothing. What belongs in the slots, and the four sources no model can generate.

Anthropic says Sonnet 5.5 lands within two points of Opus 5.5 on real-world work. The setting underneath that claim is the lever almost nobody adjusts.

Ottawa is co-funding 10,000 AI work placements. What a co-op student can genuinely do on an AI project, what kills these placements, and how to scope one.

Most continuity plans cover fire, flood and a failed server. Almost none cover the assistant your team now depends on. Four additions, and one that is free.

A $162M Mitacs programme puts students and postdocs into Canadian businesses for AI work. Two streams, what the money is, and what it does not cover.

Assistants can now read across your apps, which looks like the end of data silos. It solves one of the three kinds and leaves the expensive one untouched.

AI is good at the parts of a grant application that repeat and bad at the part that wins. Where to use it, and the one section to write yourself.

Google connected Gemini to Airtable, monday.com, Zoho and ten others. Any staff member can now link a company system to a personal assistant in seconds.