Blog
How automations behave once they run, what we would rebuild first, and what actually mattered in what the industry shipped. Written from the work, monthly.
JULY 2026 · 9 MIN READ
July's material on computer-use and browser-use models, and what we make of it: how the loop works, what the benchmarks are actually measuring, the failure modes they add to a process, and how we decide whether one of them should take over a bot that already runs.
JULY 2026 · 6 MIN READ
Selectors describe one version of a screen, so the bots built on them fail in ways that are worse than stopping. We go through the four ways that happens, what the usual patches cost you, and what we would rebuild in their place.
JULY 2026 · 5 MIN READ
Deciding what happens to a bot is a sorting problem before it is an engineering one. Here are the six questions we ask about every process in an estate, in the order we ask them, and what each answer means for whether it gets rebuilt, replaced, left alone or switched off.
JULY 2026 · 5 MIN READ
Workflow files are XML. Read enough of them and you have an inventory, a dependency graph and a risk register before you book a meeting, and you know which questions are worth a person's time.
JUNE 2026 · 9 MIN READ
The number underneath every bot migration is how long the old bot keeps running once its replacement is built. People want one weekend, and that is not an answer we offer. Two of June's releases moved numbers that bear on the overlap, so this month we have written down how we size it.
MAY 2026 · 8 MIN READ
A vendor notice in May withdrew a set of audit values that people had been relying on. That is a fair argument for holding your own run record, so we have written ours out field by field: inputs, decisions, model confidence, validation results, approvals, outputs written and errors.
APRIL 2026 · 8 MIN READ
The April release notes were largely about unit consumption and metering, which says the platforms have found the same gap we keep hitting: most teams cannot say what one run of a process costs. Screen operation bills on every run, and an API integration bills mostly once, at build time. This is the arithmetic we use to choose between them, and the measurement it has to rest on.
MARCH 2026 · 8 MIN READ
March gave us a new OpenAI model, the shutdown of three older ones, and two papers that question how any of this gets evaluated. If you run a document intake workflow, the useful question after a month like that is how you would know yours still worked. This is the harness we use: where the deterministic steps split from the judgement step, what goes in the golden dataset, how we set the confidence thresholds, and what we run when a model version changes underneath us.
FEBRUARY 2026 · 8 MIN READ
February handed us a Salesforce release, a repriced Xero and two Anthropic models, and all three land on the same scoping question. A yes on whether the system has an API settles almost nothing. Here is what we check after that first yes: documentation, licence tiers, rate limits, authentication models and sandbox access.
JANUARY 2026 · 8 MIN READ
January brings the same question every year, where to start, and it usually arrives with an answer already attached: whichever process people find most irritating. We measure five properties before we agree to build anything.
DECEMBER 2025 · 8 MIN READ
December put the model releases, the survey data and two Australian policy documents into the same few weeks. We read all of it against one question: does any of it change what you should do with an RPA estate built five years ago?
NOVEMBER 2025 · 9 MIN READ
Approval keeps being discussed as a platform feature. We treat it as three decisions taken before the workflow is built: where the gate sits, what the approver can actually see, and what gets kept afterwards. Three of November's releases bear on the first one.
OCTOBER 2025 · 8 MIN READ
October put a run of new OCR research and model releases in front of the industry, and most of it points one way. We still maintain coordinate templates on some document streams, and the reason comes down to a number most finance teams have never counted. This is where a language model beats template and zonal OCR extraction, where templates hold, and what we run over model output in Python before anything is posted.
SEPTEMBER 2025 · 8 MIN READ
September 2025 handed anyone holding a UiPath renewal quote two useful documents, the quarterly results and the FUSION announcements, and read together they tell you what the account team is planning for before you reply to it. This is how we inventory an RPA estate from its source code and run history, rank each process by what it is worth, and work through the four positions available: reduce the entitlement, renew short, rebuild, or let the licence lapse.
AUGUST 2025 · 8 MIN READ
August was a rough month for agentic browsing, and the incident write-ups describe failures already familiar from RPA estates. Both kinds of automation depend on interfaces owned by somebody else, both keep reporting success after they start getting things wrong, and that is why we put the boundaries where we do.