What we're learning

Playbooks and field notes from inside real rollouts.

Memos on AI operations in regulated companies — what shipped, what stalled, and the numbers we're allowed to publish.

Frontier APISovereign track
Board packs, memos, policy draftsYesenterprise terms: no training, DPAnot needed
Financial models, Excel workYesClaude for Excel, Copilot — check retention termsnot needed
Contracts, vendor documentsYeswith DPA; zero-retention where eligiblenot needed
Aggregated management reportingYesno customer-level PII in promptsnot needed
Cardholder dataPAN — PCI DSS scopeNever rawtokenize or mask before any promptYes
Transaction-level data with PIIRegulator-dependentEU-region, zero-retention at bestYes
AML cases, safeguarding filesdon't tryYes
The split that makes everything else simple: classify first, then two tracks run in parallel. Enterprise agreements from all four frontier vendors exclude your data from training by default and offer DPAs; zero-data-retention is available per-organization but read the exceptions — Claude for Excel, batch and file APIs carry their own retention terms. Nothing here absolves PCI scope.

Featured · Playbook · Jul 14, 2026

The FinTech CFO's AI stack: deploy this quarter, build next quarter.

Your close runs nine days, reconciliation lives in a departed analyst's spreadsheet, and the CPA reqs have been open for a quarter. The split that makes AI tractable: frontier models today for everything that may leave the building, a sovereign track for what may not — plus the tools, the hidden engineering tickets, and what actually happens to the accounting team.

CFOFinance opsReconciliation12 min read
Read article

Latest · 03

Notes from inside the org chart.

Written by the partners running the engagements — no ghostwriters, no thought-leadership filler. If we publish a number, a client signed off on it.

GLM-5.2DeepSeek V4-Pro
Intelligence Indexcomposite · Artificial Analysis v4.1 · independent51open-weight #144
Real-world agentic workGDPval-AA v2, Elo · independent15241328
Repo-scale codingSWE-bench Pro · vendor-reported62.1%55.4%
Terminal agentsTerminal-Bench Hard · independent50.846.2
Reasoning with toolsHumanity's Last Exam · vendor-reported54.748.2
Long-context reasoningAA-LCR · independent71.366.3
Instruction followingIFBench · independent73.376.5
Competition mathHMMT Feb 2026 · vendor-reported92.595.2
Competitive programmingLiveCodeBench · vendor-reportednot published93.5#1 global claim
Serving speedmedian across providers · Artificial Analysis174 tok/sTTFT 1.4s62 tok/sTTFT 1.9s
Output priceper 1M tokens, first-party API$4.40$0.87
Both models at maximum reasoning effort, July 2026. "Independent" = measured by Artificial Analysis; everything else is the vendor's own harness. Where a cell says not published, no honest comparison exists — beware of aggregators that fill the gap with mismatched benchmark versions.

Research · Jul 13, 2026

GLM-5.2 vs DeepSeek V4-Pro: the open-weight title fight, scored honestly.

The two strongest open-weight models ever released, eight weeks apart, both MIT. GLM-5.2 wins every independent agentic composite; DeepSeek V4 wins every line of the price sheet. The head-to-head numbers — vendor claims separated from independent measurements — and the verdict by workload.

GLM-5.2DeepSeek V4Benchmarks
10 min

The trust gap, measured

% of consumers
81
81
75
44
Expect a human escape hatchWant the human to pick up with contextWant to know it's an AIActually trust AI with their issue
Sources: Zendesk CX Trends 2026 (11,000+ consumers, 22 countries), Salesforce State of the AI Connected Customer 2025 (16,585 respondents), aggregated escalation surveys. Service leaders guess the last bar at 65%. It's 44.

Playbook · Jul 12, 2026

The escape hatch: automating support without burning customer trust.

AI can take two-thirds of your support queue this year — and 64% of customers wish it wouldn't. The platforms worth shortlisting, the memory layer (Graphiti, Mem0, Zep) that kills the "repeat your order number" problem, and the hard gates that keep one hallucinated refund from becoming case law.

Customer supportAgent memoryEscalation design
12 min

Two true lines

One dispute overlay, first quarter in production · indexed to week 1 = 100

Monthly token spendCost per resolved case
Index
0100200300400W1W4W8W12×3.2−60%

Week in production

Both lines are true at once: the token bill tripled while the cost of resolving a case fell 60%. A scorecard that only sees the first line kills a working program; one that only sees the second gets killed by finance.

Playbook · Jul 11, 2026

The AI scorecard: prove the leverage before the token bill arrives.

Accuracy dashboards don't survive a CFO review, and unmetered token burn doesn't survive a quarter. The nine numbers we put on every overlay — queue, cost, guardrails — and the thresholds that trigger a scale-up, a fix, or a kill.

KPIsToken economicsLeadership reporting
8 min

The archive · 02

Earlier memos.

Jul 10, 2026

The sovereign stack: what to run when the data can't leave the building.

Banks and payment companies can't pipe cardholder data into a commercial API, however good the DPA looks. Which open-weight models to deploy instead, what hardware each tier actually needs, and the order to roll them out — without buying a supercomputer first.

Playbook · 11 min

Jul 2, 2026

The pilot graveyard: why enterprise AI dies after the demo.

Most of the AI pilots we audit never touch a production queue. The failure isn't the model — it's that nobody rebuilt the workflow around it. The operating cadence that gets an overlay out of the sandbox, week by week.

Positions · 6 min

Where to start

Book an org assessment.
Ninety minutes, no slides.

We'll spend an hour with your leadership team mapping where AI creates the most leverage in your operation — and what an overlay would look like in practice.

Book the assessmentmail@orgops.net