What moved in tech this week — and what it means for EdgePointe
Week of October 5, 2026 · a 5–7 minute catch-up
Nine cards, ~30 seconds each. You get the so-what; Paul carries the detail. Hit Play narrated to listen.
EDGEPOINTE · VANTAGE
The big picture · our bet
Cheaper AI keeps proving our bet: the value is in whose data it runs on Bet holds
Our bet has never been on which AI company wins. It is that somebody operates this for a small business, on that business's own material. Two fresh pieces of evidence this fortnight.
The market is real
Our voice supplier says its agents now handle more than 15 million conversations a week, tripled since February. Category confirmed — and it is moving without us.
The pricing theory
A Stanford economics paper says AI pushes service firms toward asset-based pricing, not cheaper hours. Vantage is the asset.
Hourly work races to a floor. A platform the client holds does not.
Bottom line: stop selling hours made cheaper by AI. Sell the platform and the memory it builds from the venue's own material.
Where we stand
The finished phone AI is still parked — and the unbuilt piece is now twenty-one weeks old Clock running
The phone AI
Built, tested, answering a real number — and still unshipped. It was parked eighteen days on September 28; it is now about twenty-five.
Venue-trained receptionist
The version competitors cannot copy is unbuilt for the twenty-first straight week.
Voice vendor check
One login to read the new voice model's credit cost. Six check-ins now, still open — and the whole cost model waits on it.
Venue-software watch
Seven quiet weeks, and we learned why: one rival's update page has been dead since March 2024 and another has none. We are changing how we watch.
Bottom line: nothing this fortnight removed an excuse for the phone AI, and nothing added an obstacle.
Claude · our core tool this week
Claude Code changed its own safety default on us — one fix, one sitting Do this week
On September 28, Claude Code began starting in auto mode whenever no permission setting is chosen. We never chose one — so every session picked up the vendor's default.
Then on October 3, four loopholes in that same mode were patched. Our Mac still runs the version from before those fixes.
Ten of our unattended runner scripts operate with permission checks switched off. No loss has occurred, and none is claimed.
The fix is two small steps: upgrade the Mac, and write a one-line settings file that pins the mode we actually want. Then rotate stored credentials once, in that order.
Bottom line: a safety setting changed itself without a line in anyone's notes. Pinning it makes the next change a decision, not a surprise.
Platform · Cloudflare
Vantage is fine — and two Cloudflare changes make our reporting easier to build All calm
Nothing on our platform broke or is scheduled to break. No new retirement notices touch what we run.
Client reporting
As of October 2, thirty days of traffic history on every plan, plus our worker's own logs, can sit on one dashboard. The quarterly client report is now mostly assembled from parts we already pay for.
Easier setup
An ordinary team member can now create the scoped access key for the worker — no top-level admin needed.
A date to watch
The smaller of our two pinned AI models is guaranteed only until October 15. We re-check the table right after.
Bottom line: calm platform, one small date. Build one sample client dashboard and we have a report we can bill for.
Biggest growth move · Managed IT
Microsoft's November 2 billing change is our cleanest chance to be useful before the invoice lands Act now
Microsoft moves Copilot Business to usage-based billing by default on November 2 — four weeks out. A fixed per-seat line becomes a variable one on our clients' invoices.
Any client on Copilot gets a bill they did not expect and will call us. Managing that spend is exactly what a Managed IT provider is for — and it is a billable line.
This week: name the affected clients and draft a one-paragraph heads-up. It comes from one trade outlet, so we confirm it at Microsoft's own announcement before anything goes to a client.
The sales angle moved too. Industry voices say roughly 80% of what clients need is automation, not AI. Lead the governance one-pager with outcomes; AI is the mechanism, not the pitch.
Bottom line: a dated event, a billable fix, and we get to be the first call instead of the last.
Client-facing · venues & Vantage
A free, fast tool from Cloudflare could make our receptionist and Ask Vantage smarter and cheaper Measure first
On October 1, Cloudflare released Clef, an open model built to sort messages by urgency and route them. It runs on the AI service our platform already uses in production.
Near-free: roughly nine cents per million words, and under forty milliseconds on the small version. Fast enough for a live phone call.
The test is cheap: run fifty real messages through it and score the routing against what a person would do. Inside the free daily allowance.
Honest limits: the benchmarks are Cloudflare's own. Nothing goes in front of a client until we measure it on our data.
Also noted: Webflow can now show customers which AI engines cite their site. Vantage has no equivalent yet — a visible gap in a sales conversation.
Bottom line: the model was never the expensive part, and now it is nearly free. Our edge is the venue's own memory and running it.
The wider AI race · so what
Three honest cautions about the AI scoreboard — including one about our own design critic Context
This fortnight's research corrected our own measuring stick.
Our design critic is a tiebreaker, not a grader. The best published picture-judging systems agree with humans only about 45% on the usual scale. We stop calling its score a quality verdict, and we measure its noise before tuning it again.
Test results can be made to look better or worse just by how they are reported — one study swung accuracy from 55% to 90% without changing a single verdict.
The so-what: the win is a tester that clicks through the page and then scores it. That is a later build, not this week's task.
A caveat from the research: one striking figure on order-bias could not be re-verified, and our conclusions do not rest on it.
Bottom line: never show a client a design score we cannot defend. Show the page.
The bottom line
Two decisions need Bob and Niven this week Decisions
1 · Set a ship date for the phone AI
Finished and parked about twenty-five days. Every week it sits, the generic version gets given away by someone else. It needs a date on a calendar, not more build.
2 · Approve the Copilot heads-up
Microsoft's billing change lands November 2. Agree that we name the affected clients and send a short, confirmed note before then — and offer to manage the spend.
What Paul does without you this week: upgrade the Mac and pin the safety setting, rotate credentials once after that, build one sample client dashboard, and run the fifty-message routing test.
Bottom line: two yes-or-no answers keep us ahead of a dated change instead of behind it.
Card 1 / 9 · ~20s
Narration audio not found yet — run the render step, then reopen.