Field notes from our teams — what's working in production, what we've learned the hard way, and where enterprise technology is heading next.
Salesforce is paying $5.66 billion for Braze with no combined roadmap and a close up to 467 days away. Why we say keep Marketing Cloud and Braze renewals short.
WhatsApp’s help pages carried 102 hidden notes for its support bot, and 8 of the 20 Redact summarised told it what to keep from users. Why we say audit your own bot’s rulebook this week.
JPMorgan and Schwab block outside agents that log in with a customer’s password. Schwab’s weekly count is one block a year per 27,000 to 77,000 accounts. Why we say measure your portal first.
OpenCode says provider keys stored in its Console may be gone. One victim logged 8,966 requests in a day. Why we say rotate at the provider before Monday.
OpenAI’s models took 148,940 records from a site that only had to give public figures. The owner’s first chance to know was 120 days later. Why we say PILOT for anyone with a public Salesforce site.
Meta opened an 816B model with a usable licence. Hosted, it costs 10 times Claude Haiku 5.5 per call before it reasons. Why we say PILOT only if you need private weights.
LangChain now hands agent housekeeping to ScaleDown’s small models. The split is right; the numbers are the vendor’s and it fails open. Why we say PILOT, on internal agents only.
$2 and $10 per million tokens until 30 November, then double. Per task it already costs more than Claude Opus 5.5 on one independent test. Why we say PILOT, priced at December rates.
718 firms: agents add 30% more code and no significant gain in shipped work. Pull requests waiting in review rise 83%. Why we say PILOT, and what to buy with the seats.
One supplier password, 8.75 million records, found by an invoice. The same pull from Salesforce is 0.2% of a day’s API calls. Why we say PILOT, and the audit to run this week.
Your customers’ agents want to sign in. 42 partners, 16 of them banks, payment firms, or insurers, and no CRM on the list. Why we say WATCH, and the policy to write first.
Up to 90% cheaper, and level with GPT-6 Luna to the cent. Past 100,000 tokens a new tokenizer and a higher tier cut the saving to about 35%. Why we say PILOT.
Commerce Cloud replatforms rarely fail on the page templates. The four integration boundaries that decide the timeline, the catalogue model mismatch, and the five questions to ask a partner.
A trillion parameters, open weights promised for late October. About 525 GB of GPU memory at 4-bit and 1,050 GB at 8-bit before a single prompt. Why we say WATCH.
A 27B local model won all six business-analysis cases against a 284B cloud model. The cloud model was about 19.8× faster and won on summaries. The narrative, with both halves.
Always-on agents, off by default in Enterprise. Boundary flags rise from 8.6% to 19.7% as task chains double. Why we say WATCH, and what to check on Monday.
OpenAI warned 100-plus organisations about its own agents. One trail self-deleted in 48 hours; free Salesforce logs keep one day. The drill to run this week.
Two decision models, two weeks apart, aimed at the same job. Status, list price, the independent evidence, cost at a million cases a month, and a free synthetic test set to run on your own stack.
A 27B model running locally won all six business-analysis cases against a 284B cloud model, and lost on summaries. Every reported aggregate, the derivations, the open CSV, and what we could not verify.
Every answer online stops at "it depends". The seven variables that actually move the number — and what a low quote usually leaves out.
Quote-to-cash rarely fails on the first quote. It fails on the first amendment. The five places Revenue Cloud implementations break, and the five questions that separate partners.
Salesforce retired its four-tier partner programme in March 2026. A current guide to the firms, what tier actually tells you, and the questions that separate them.
Most migrations don't fail during the load — they fail three weeks later. Load order, audit fields, automation and the failures that surface after go-live.
A four-rung value ladder from activity to business value, a five-level evidence maturity model, and a built-in skeptic that challenges inflated claims.
Hybrid retrieval, reciprocal-rank fusion, exact message provenance and 100% abstention on trick questions — running entirely on-premises.
Windmill, Supabase, Ollama and LangGraph combined into a sovereign platform — and why deterministic orchestration should wrap probabilistic agents.
Part 1. A silent schema change broke everything downstream. The detection, the guardrails and the fix that made our data pipeline resilient.
Your model can rot while every test still passes. How we instrumented drift detection to catch degradation before it reached users.
Reliability often isn't a model problem; it's a data problem. The pipeline and validation changes that made our predictions trustworthy.
The glamorous part is the model. The work that actually decides success happens long before — in the data and the maths underneath it.
Smart partitioning that conquers enterprise scale. How we restructured a massive Postgres database without taking the product offline.
What it really takes to push a Heroku stack to nine-figure traffic — the bottlenecks we hit and the architecture that held under load.
Credibility drives genuine conversions. How Prestanda's Double Diamond process turns enterprise UX from guesswork into a repeatable system.