What the December releases change about the reading step
December was heavy. OpenAI shipped GPT-5.2, G2 published its enterprise AI agents report, UiPath and Microsoft both moved on agent controls, and the federal government released the National AI Plan. We held the same question against all four, and it is not whether the technology got better. It is whether any of it changes what to do with an RPA estate built five years ago.
OpenAI released GPT-5.2 in December 2025 in three tiers, instant, thinking and pro, and published benchmark claims including 55.6 percent on SWE-Bench Pro against 43.3 percent for Gemini 3 Pro, alongside an internal evaluation claiming GPT-5.2 Thinking matches or beats human professionals on 70.9 percent of tasks. Coverage at the time noted that independent validation of benchmarks takes time and that many corporate performance claims are difficult to verify. Independent work on those figures has not appeared, so we carry them as vendor claims and nothing more.
Our position on benchmarks of that kind is that they are a weak proxy for the steps we care about in a business process, which are much narrower: reading a remittance advice PDF and pulling the invoice numbers, classifying an inbound email into one of nine queues, extracting seven fields from a photographed delivery docket. Those are the steps that keep a person sitting in the middle of an otherwise automated process. Our design position for that class of work is to build the step as a model call with per-field confidence scores and a Python validation layer behind it.
The pattern is the same one we have used for several years, and nothing in December moved us off it. A model does the interpretation, Python checks the result against rules held in configuration files, and anything under the confidence threshold routes to a person. The figure worth measuring in your own environment is the proportion of documents clearing the threshold on the first pass, because that number sets how much manual handling is left, and it varies with the quality and format of your own documents.
How much autonomy people say they are running
G2 published its enterprise AI agents report on 15 December 2025, and it needs handling with some care, because it carries two different populations. The report carries figures from G2's August 2025 survey, which put 57 percent of companies with AI agents already in production, 22 percent in pilot and 21 percent pre-pilot. The December report's own fieldwork was narrower, a structured questionnaire sent between October and November 2025 to five enterprise software companies building or deploying agents. On autonomy, 47 percent of verified agent buyers say they are at what the report calls autonomy with guardrails, and fewer than 10 percent report a full-autonomy mindset. We keep those two populations apart when we quote any of it. The August figures come from a survey. The December autonomy numbers come off a questionnaire to five companies, which is a small enough base that we would not read them as an industry position.
UiPath's December release notes point the same way from the engineering side. On 2 December it shipped a Max iterations setting for autonomous agents, a configurable bound between 1 and 100, documented as a way to control cost and execution duration and prevent infinite loops. An autonomous agent in production can loop, each iteration costs a model call, and the setting is the same bound a developer would put around a while loop in ordinary code. If you are running agents on UiPath, set that value deliberately as part of the design rather than leaving whatever came out of the box.
The shape we keep ending up with for a mid-sized operation has a defined input, a defined output, a step budget, a log of every tool call, and a person approving anything that leaves the organisation. The same G2 report recorded 34 percent of verified agent buyers already using an oversight model where agents act first and humans review after. That is a governance decision about what an organisation is willing to send out unreviewed, and our impression is that in a lot of places it is being made by default. Every platform ships a default for it, and we would go and look at what yours is set to rather than inherit it.
What the vendors actually shipped
UiPath reported third quarter fiscal 2026 results on 3 December, with revenue up 16 percent year on year and annual recurring revenue up 11 percent. It also announced it had become a founding technical contributor to AIUC-1, a security framework for enterprise AI agent adoption. Daniel Dines framed the demand as customers wanting one governed platform covering deterministic automation, agentic capability and orchestration. Whatever you make of vendor framing, that one points at a split we find in most ageing estates, where one team maintains the old bots and another runs model pilots with no shared logging between them.
UiPath's Agents release notes dated 19 December describe evaluation sets that run several model and temperature combinations in parallel and compare accuracy, latency and behaviour across test cases, plus a trace annotation feature for leaving feedback on individual spans, including tool calls, guardrail checks and model outputs. In general software terms those two features are a test harness and distributed tracing. We would want both in place before an agent runs against a process that touches customer records, and we would say the same about a build on any other platform.
Microsoft's Power Automate 2025 release wave 2 plan lists version control for desktop flows entering public preview on 12 December 2025, with general availability scheduled for 13 May 2026, and debugging into condition actions at runtime reaching general availability on 11 December 2025. Version control and a step debugger have been baseline tooling in general software development for two decades, and both reached Power Automate desktop flows in December 2025. That gap is the clearest measure we have of how far desktop RPA tooling has been running behind ordinary software practice. If your desktop flows currently sit as unversioned files on a network share, the December preview is the thing we would test first in the new year.
What still binds an old estate to the screen
UiPath selectors and Power Automate Desktop UI elements bind an automation to the structure of a screen. A vendor release that changes a DOM node, renames a control or moves a field breaks that binding, and the automation either fails or writes to the wrong place. Nothing in the 2025 releases changed that mechanism, so a process built on selectors carries the same maintenance exposure it carried a year ago.
Power Automate made browser automation through direct control, with no browser extension required, generally available on 6 November 2025. That removes the extension as a failure point, though the automation still binds to the selectors on the page. Microsoft has a self-healing agent for UI and web automation in desktop flows scheduled for public preview on 31 March 2026. We are interested in that one and we are not planning around it yet. Many of the systems these bots drive also publish documented REST APIs, and an API contract is versioned, so it does not change when the vendor restyles the screen.
The first piece of work we would do in January is to build two lists. The first records maintenance history per process, counting repairs over the last two years. The second records, for every system each process touches, whether there is a documented API, whether your licence includes access to it, and what the authentication model is. Processes that sit high on the first list and have good API coverage on the second are the rebuild candidates. A process with no repair history, running against a system that has not changed in years, should be recorded in the assessment as working and left alone. We put that in writing, because the useful part of an estate review is partly what it tells you not to touch.
The policy news, and what it does not require of you
The federal government released the National AI Plan on 2 December 2025. Reporting at the time recorded its approach as continuing to build on Australia's existing legal and regulatory frameworks, with established law remaining the foundation for addressing AI-related risk, and the plan does not propose AI-specific legislation. It also funds an AI Safety Institute to monitor AI capability and risk, scheduled to launch in early 2026.
The Productivity Commission handed its Harnessing data and digital technology inquiry report, report number 111, to the government on 10 December 2025, and it was released publicly on 19 December. Enabling AI's productivity potential is one of the five reform areas the inquiry covers.
For an operations leader the practical reading is that no automation-specific compliance regime is arriving to tell you what your automations must log, how long you keep run records, or which outputs need human sign-off. Your obligations under the Privacy Act, your records retention policy and your financial controls already cover work a model does on your behalf. What gets logged, how long it is kept and where approval sits are internal engineering decisions, and we write them into the design when we rebuild a process, alongside the maintenance and API lists that come out of the January assessment. Waiting for a regulator to specify them is not a plan we would put our name to.
Two things we are watching into next year. The self-healing agent for desktop flows is the one that would change our advice if it works, and whether it holds up against a real estate under vendor release pressure is untested, with the preview still months out. For the AI Safety Institute, the test once it launches is whether its output is usable by an operations team rather than a policy one.