What moved in tech this week — and what it means for EdgePointe

Somebody else just wrote our business plan in a trade magazine — and gave us until October to claim it Bet intact

Our bet hasn't changed: we don't bet on which AI wins. We own the client relationship, and Vantage is the front of the business — website, leads, phone, and the AI layer as one thing.

On September 10 the MSP trade press named a service category that does not yet have a name: customers deploy the AI agents, but their IT provider should govern them. The shape it lays out is a one-time setup fee plus a recurring oversight retainer — monitoring how the agents perform, reviewing their output for errors, and keeping an audit trail the compliance people will accept.

We already hold all three of those. Our phone AI now reports on every call it takes. Our house rule is that AI proposes and a person posts. Vantage has shipped an audit log for months. We built them for ourselves and never packaged them.

Finished work is still parked, and one five-minute check runs out of road tomorrow Clock running

We finally have a way to prove our quality scoring works instead of arguing about it Win

For seven weeks running, every briefing has included another tweak to how we grade the websites we build. Not one of those tweaks has ever been proven, because there was no way to run the same test twice and compare.

On September 11 our AI vendor shipped exactly that: a command that runs a scoring suite and returns a score and a report you can keep. The word that matters is reproducible. It turns an opinion about quality into something closer to a regression test — run it, change one thing, run it again, and the difference is a number.

Same week, three ways around our own safety rules were closed — and one rule was deliberately narrowed. A guard rail written expecting wide reach now only binds inside the file that declared it. That one gets re-verified, not assumed.

Vantage is fine — and our phone AI just learned how to report on itself All calm

Nothing broke and nothing is about to. The one genuinely useful change is that every call our AI receptionist takes now produces a result you can read.

• Per-call outcomes, live. Each conversation now reports whether it completed, stalled, or failed — and how long it took to start speaking. Until this week, a bad call was just "a bad call." Four different faults looked identical.
• New security detections land tomorrow in watch-only mode. Nothing blocks, nothing breaks. One of them will eventually flag any client site that accidentally leaves developer files public — which makes it a useful audit prompt now, not an alarm.
• One quiet trap: new background jobs keep their error history for seven days instead of thirty. Old ones are unchanged, and nothing in the code tells you which is which. Paul sets it explicitly rather than inheriting it.

Governing our clients' AI is a service we can sell now — and the pricing argument is no longer ours alone Act now

Card 2's category has a delivery shape we can run: a one-time setup engagement, then a recurring oversight retainer. Setup is understanding what the client is using the AI for, defining what good looks like, mapping the risk and configuring it. The retainer is watching it, checking its work, and keeping the record.

The market numbers behind it are now third-party, not our opinion. McKinsey: AI-enabled service desks cut interactions 40–50% and cost-to-serve by more than 20%. Forrester: virtual agents resolve up to 65% of first contacts without a person. Everest Group says managed-service pricing is moving off hourly and onto outcomes — uptime, time-to-resolution, security results.

One honest caution, and it reorders our own plan: the same reporting says get your own house in order first. Clean ticket data and written procedures come before the automation — automation pointed at an undocumented process just automates the mess.

A competitor shipped an owner dashboard. One sentence has to leave our pitch this week. Watch

On September 11, Duda shipped a client-facing site performance dashboard — not an agency tab, the thing the venue owner logs into. That is the fifth owner-portal piece they've shipped in fourteen days, after conversational site building, logins, forms, and search optimization.

The bundle is still genuinely ours. They have no lead inbox with alerting, no AI answering a phone number, no reviews engine, and nothing that ties a phone call to a lead to a booking. But the sentence "they can't build an owner portal" is no longer true, and if a prospect notices before we correct it, we look behind instead of ahead.

The replacement line: "They've built the pieces. Nobody but us has connected them." The venue-software gap holds for a different reason than we've been saying — venue software is sold to an events manager, the website is bought by the owner. That's a buyer gap, and it's durable.

A year-old study says our quality scoring is built backwards — and tells us to cancel a build Redirect

The most useful thing found this week is a paper from last October that nobody here had read. It benchmarks AI judges on exactly our problem — judging the quality of a built web page — and it contradicts our design in four specific ways.

• Compare two pages, don't score one. Head-to-head grading beats absolute scoring by more than 8 points. We already have the comparator sitting there — the previous version, or the original site we're improving on.
• Show it the code, not just a picture. Judgments are anchored in the source. This does not replace a human looking at the pixels.
• Yes/no beats one-to-five. Graded scales are where these things fall apart.
• Don't build the elaborate version. The multi-step "agent" judge loses to a single well-aimed call, because errors pile up at each handoff. That's the design we had planned.

Three decisions need Bob and Niven — and two of them have dates attached Decisions

Paul carries the rest: the Cloudflare check tomorrow, the three sixty-second greps, and the first before-and-after quality number using this week's new scoring command.