The incumbent published the number we said didn't exist — because we read a roundup instead of the source
Weekly Research Update · Monday, November 2, 2026 · a 4–5 minute catch-up
Seven cards, about 30 seconds each. The week's biggest finding is a correction to our own record, the second is a break-even figure we can test against our own ticket log in one sitting, and the most urgent line on the page is a one-line fix with two days left.
EDGEPOINTE · VANTAGE
The headline · the week's real finding
The biggest service-desk vendor shipped four AI agents with partner-measured economics attached — and we filed it as having none. Game-changer
On October 29 the incumbent put four AI agents into general availability — one for remote monitoring, one that answers inbound support calls, and two that close information gaps before a ticket exists.
Our board recorded, three days later, that no economics were disclosed, so the buy-versus-build line could not be re-costed. ❌ That was wrong. The vendor's own release carried five quantified partner outcomes.
The number that matters
Break-even at roughly sixteen minutes per technician per day
The supporting figures
Over seven thousand tickets evaluated in thirty days; ninety-one percent useful diagnosis at about a minute each; first-touch time from an hour to five minutes
⚠ Provenance, stated plainly
A vendor press release; every figure is partner-supplied. Treat as marketing until corroborated — but marketing and absent are different things
Bottom line: sixteen minutes a day is testable against our own ticket log in one sitting. Below it automation loses money; above it the build pays.
The defect · one layer below coverage
The scan list held. Every source was checked by name. The failure was reading a lane at a trade roundup while the vendor's own release sat one day older. Method
This is worth more than the finding it produced, because it is the kind of mistake that repeats quietly.
✅ New standing rule, adopted: when a roundup names a vendor event, fetch the vendor's own release before filing the row. A roundup is a pointer, not a source.
Three lanes are honestly short this week and are stated rather than implied clear:
One trade publication has been unreadable at the fetch layer for seven consecutive runs. It is not being watched — it is being listed.
One vendor source carries sixty-seven catalog entries and no page date anywhere; the only date is a filename, and the published artifact is two years stale. Sixteen runs. It should be demoted to a quarterly re-read.
✅ One correction in our favour: a post we had recorded as not existing does exist — dated October 22, seven items, all payments and invoicing. That confirms last week's read from the primary source instead of inferring it.
Bottom line: the scan was complete and the sourcing was shallow. Those are different defects and this one is cheaper to fix.
The urgent line · not a finding, a measurement
One line of code, five identical copies, and two days until the platform starts failing the query. Overdue
Our content-history query still runs unbounded, with no limit clause, measured verbatim identical in five of five code trees this morning. The platform begins enforcing a fifty-megabyte result ceiling in two days.
What the table is
The per-tenant live-edits history — the one table that grows without bound, and the one our own house rules name as the recovery mechanism
The fix
One line plus pagination, through the normal build gate
❌ The record
Fourth consecutive run measured, fourth consecutive run outstanding
✅ Nothing else on the board has a two-day clock and a one-line fix. That is why it leads the stack — not because of anything that happened this week.
Bottom line: this is the single most actionable item in the report and it has been for a month.
The opportunity · build once, bills twice
Per-client AI cost attribution is now two endpoints and a header — against a service line a distributor just named and is building tooling for. Sellable now
Two things arrived in the same window and fit together.
The instrument
Our model vendor shipped an organisation-level usage and cost analytics API in beta. Paired with per-end-customer request profiles, that gives model cost per tenant without building a metering layer
The market
A major distributor is naming AI cost management as an MSP's next recurring service — customers cannot connect token activity to cost — and is shipping the tooling for it
The pitch
"We tell you what your AI actually costs, per department, monthly." A dashboard read and a monthly review
How to price it
As a share of savings — which means it needs no rate card
⚠ Distributor content marketing: the opportunity is real, the urgency is invented. And the same build is our own cost control.
Bottom line: small effort, recurring revenue, and we are the first client of it.
Our stack and our competitors · measured, not asserted
Four breaking changes and a model deprecation landed on us this week. All five are no-ops — and a builder shipped the feature venues ask for first. Zeros held
Greped across all five trees rather than re-flagged, and recorded as zeros so they are never raised again:
Deferred code updates, an agent-SDK removal, a jurisdiction restriction, a subrequest change — zero files, zero occurrences, each one.
A fast-mode removal with a hard date — zero files. Our pinned-model inventory is entirely on current models.
⚠ The pattern worth keeping, not either item: that jurisdiction property is create-time-only and irreversible — the second such one-way door this board has seen. When a residency-sensitive client asks, it has to be decided before the resource exists. That belongs in the isolation white paper as a paragraph.
❗ On the competitive side: a major builder turned late-cancellation and no-show fees from a developer API into a toggle an owner can switch on — charged per guest or per party. That is the single most-requested thing in a venue booking flow, and the gap between our flow and a builder's narrowed on exactly that feature.
Bottom line: nothing broke on us this week. The builder gap closing on booking deposits is the thing to watch.
So what · the stack and the honest scorecard
The order on top did not change because of a finding. It changed because the clock moved. Act on it
Ranking principle, applied without exception: a date outranks a capability, and a decaying evidence window counts as a date.
1 · Two days
The one-line database limit. Small effort
2 · Ten days
The AI-disclosure line on the receptionist — the only dated compliance obligation on a shipped product
3 · No deadline, which is the trap
A credential sweep that cannot answer an October question if it is run in November. Carry two firewall rules on the same visit — one errand, not three
4 · New
Re-cost buy-versus-build against that published break-even
5 · New
Stand up per-client AI cost attribution, then sell it
❌ What did not change: the database fix unshipped four runs. A finished receptionist build unshipped ten weeks — and this is the first week not shipping it carries a competitive cost rather than an opportunity cost. Nothing this week removed an excuse.
Bottom line: two small commits clear the top two lines. The rest is one client visit and one calculation.