27 issues. 74,000 words. Half a year of Perspectives.
January through July, with a newsletter issue published every week. Here's what I wrote about, what I got right, and where the Power Platform and AI story actually went.
How this was built: Fable 5 went through my newsletter archive and selected a set of predictions that I had made as part of the issues. Then I sat down and reflected on how well I personally think I succeeded in the prediction. Comments in the "what happened" part are the answer from my current self to my past self.
Fourteen claims from the first half of the year, graded against what actually happened.
8 confirmed · 3 partial · 2 wrong · 1 too early
“A365 is more like CDM than any concrete product designed with a real customer in mind.”
From: Post-Ignite agentic hangover · Jan 16, 2026
What happened
Not even Microsoft understood A365 prerequisites upon GA, leading them to pull the rug on E3 customers since so much of the "product" functionality relies on existing products like Purview.
“The flow designer is being prepared to live inside the Copilot Studio chrome and all the AI assistants in there.”
From: The UX for automation still sux · Feb 20, 2026
What happened
Workflows are now as prominent inside the modern Copilot Studio agent experience as agents themselves. This is because the Power Virtual Agents era orchestrator with its "follow the rules" structure is being replaced by a purely agentic chain-of-thought LLM that is far less deterministic. Combined with the per-action Copilot Credits pricing model, expect to see all automation scenarios pushed into Copilot Studio primarily (well, at least non-RPA ones).
“We'll see a "Microsoft Claw 365" kind of a thing from Redmond in the near future.”
From: Clawpilot: no UI for AI · Feb 27, 2026
What happened
I literally posted about "ClawPilot" on LinkedIn two months before Omar Shahine from Microsoft shared that they are using it as the name for an internal version of what was later branded Microsoft Scout.
“This isn't going to magically get better. We've waited three years and the experience still sucks, just in novel ways.”
From: AI still can't figure out PowerPoint · Apr 9, 2026
What happened
I'll believe it when I see a coherent, single story from Microsoft on how to generate proper PowerPoint documents with LLMs. And since their agentic AI story consists of half a dozen competing tools, I won't have to change my mind in the foreseeable future.
“Dataflows Gen1 could have kept serving most customer needs for years. But it didn't happen — and now it never will.”
From: I see dead features, part 2: Dataflows · Apr 16, 2026
What happened
I simply see no way for Power Platform Dataflows to ever rise from their grave. Fabric is following a different path entirely, and if you thought integrating something small and simple like Power BI with Power Apps / Dynamics 365 was always a struggle, the Gen2 Dataflows are a whole different ballgame.
“Investments will rather be directed to building something brand new than making existing tooling like Power Fx work better with LLMs.”
From: Power Fx as a liability · Apr 24, 2026
What happened
Maybe Microsoft will eventually develop some kind of a bridge to port all the Power Fx dependencies into native code that LLMs love to write. But they've kept things like SSRS "alive" in Power Platform for 20 years since no one had KPIs that made touching them anything but pure poison, so it might also just never happen.
“There aren't any Copilot Credit costs during Frontier preview — I certainly wouldn't bet on that being the commercial model going forward.”
From: Scout, Claw, Cowork, Tasks, Opal, Agent 365... · Jun 12, 2026
What happened
This was too easy to call, and yet the arrival of the pay-per-use meter on Copilot Cowork has taken some of the M365 modern work AI advocates by surprise. The party is over when it comes to unlimited VC-sponsored tokens, and Microsoft isn't scared of being the early adopter of proper licensing models to address this, as we've seen with GitHub Copilot and later Copilot Cowork. Never forget: Microsoft literally invented the software licensing business and they're not going to sit back and watch frontier lab geeks like Sam Altman or Dario Amodei pretend like this isn't the path they'll need to eventually follow.
“If planning Power Apps per app licensing wasn't fun for customers, they're absolutely going to hate Copilot Credits.”
From: Markdown won, now what's Microsoft gonna do? · Jul 3, 2026
What happened
This is purely anecdotal and supported by random social comments, but I've yet to come across a representative of a business unit tasked with managing Copilot Credits reacting "this is fine". It's new, it's confusing, it's impossible to predict, it's complex to measure. It's absolutely everything that makes software licensing scary.
“The crystal ball I'm using has been flashing Anthropic's logo, and it appears to be constantly growing in size.”
From: Microsoft AI numbers: the good, the bad & the ugly · Feb 13, 2026
What happened
In the longer term, both OpenAI and Anthropic are equally problematic for Microsoft's ambitions to own the platform. The valuation of both startups hinges on their ability to deliver more than just frontier models, so Satya has been vocally advocating "every company should own their own LLM". The customer zero for this is of course Microsoft itself and what they hope to achieve with the MAI family of models.
“My sources tell me MS is actually going to bring the Per App license back to other channels, too.”
From: The licensing spaghetti on Microsoft's wall · Mar 20, 2026
What happened
A great example of how it doesn't matter what Microsoft representatives will privately tell partners. No global return of Power Apps Per App has been reported, but obviously the CSP lifeline still exists, making this a persistent reminder of the messy history of the SKU.
“The retention issue will be a far bigger concern than getting customers to subscribe to M365 Copilot for the first time.”
From: Microsoft AI numbers: the good, the bad & the ugly · Feb 13, 2026
What happened
Up until FY26 Q3, the number of licensed M365 Copilot seats reported by Microsoft has grown. How many have decided to not expand their pilots? That's very hard to tell.
“We are unlikely to get any clear roadmap explicitly calling out the differences between the orchestrators and how they'll be merged.”
From: The two engine options of Copilot agents · May 22, 2026
What happened
In June, Microsoft launched the modern agent experience for Copilot Studio (in preview, despite mistakenly claiming GA initially). They've naturally avoided directly addressing the real questions, but it seems like the Samba orchestrator and any Power Virtual Agents style agent created before June 2026 will be frozen in time with the old chatbot tech. The Power CAT docs clearly state that you shouldn't migrate the exact agent functionality over but rather rearchitect it, so this seems like postponing the hard questions of the tech merger to a later date.
“The more knowledge work resembles coding, the bigger the pressure to justify sticking with Office-era concepts for managing information.”
From: Markdown won, now what's Microsoft gonna do? · Jul 3, 2026
What happened
After this post, Microsoft first introduced the upcoming support for .TXT and .RTF files in Copilot Notebooks. Yeah, imagine the WTF reaction from that. A couple of days later, a separate message was published: "actually, we'll support Markdown files as knowledge sources, too". The jury is still out on how Microsoft 365 will avoid becoming just a wrapper on top of .md files while accepting the unstoppable flood of Markdown content everywhere outside their cloud.
“The per-agent licensing story is coming. It remains in Frontier preview and I have the trial licenses to prove it.”
From: The licensing spaghetti on Microsoft's wall · Mar 20, 2026
What happened
Microsoft renamed the per-agent Agent 365 license to "Microsoft 365 Frontier for AI teammates". I'm pretty sure they wouldn't have bothered if the agent license wasn't going to happen. Yet with the ever shifting story of AI agent licensing, even this doesn't yet guarantee it will become a real, transactable SKU.
Details →Six months of tracking what the adoption data, the missing metrics, and Microsoft's own internal memos say about Copilot as a business — as opposed to what the keynotes say.
PlusJanuary 16, 2026
Two months after Ignite, not one Agent 365 launch partner can show documentation that their integration exists. Is this a product or a "product"?
Agent 365 is a keynote "product" in the Common Data Model tradition — launched without shipping, with an ecosystem of partner solutions that cannot be verified to exist outside marketing videos.
PlusFebruary 13, 2026
15 million paid seats is 3.3% of the M365 base after two years. And when users can pick any AI tool, only 8% keep choosing Copilot.
The 15 million seat disclosure confirms rather than refutes the commercial failure thesis: companies stop reporting metrics that stop improving, and Copilot's real problem is that only one in twelve users keeps choosing it once they try alternatives.
PlusApril 9, 2026
There are five ways to make a PowerPoint with Copilot, and not one delivers both a good-looking deck and a real .pptx file. Three years of R&D: a box of chocolates.
After three years and hundreds of billions in AI spend, no Microsoft Copilot can generate a PowerPoint that is both well-designed and structurally a real PowerPoint — because LLMs are fluent in HTML, not OOXML, and the "headless Office" marketing claims don't survive inspection of the actual files.
PlusJuly 10, 2026
Copilot now needs to "earn the right to exist," per Microsoft's own internal memo. The fix: a new leadership team, a Super App, and yet more renaming.
Copilot's fragmentation isn't a product problem but Conway's law in action — Microsoft keeps shipping its org chart, and no leadership reshuffle or Super App will fix that before the AI capex reckoning forces a real refresh.
Features don't get killed at Microsoft anymore — they get abandoned, locked, or silently disconnected. These are the issues that documented the body count.
February 6, 2026
One managed solution update locked Planner Premium's Plan table for every customer overnight. Five days on, Microsoft has said nothing in public.
Silently flipping the customization kill switch on Planner Premium's Plan table — after years of inviting partners to build on it — is the most reckless managed-solution move Microsoft has ever made, and the total silence afterwards makes it worse.
PlusMarch 12, 2026
"The most common cause is a changed password." It is NOT. It's an expired OAuth token — and if a Workflows agent built the flow, you can't fix it at all.
Microsoft's Frontier "products" like the Workflows agent ship with architecture embarrassing enough to hide — misleading alert emails, unfixable connections, and thousands of broken artifacts that the maker would get blamed for.
April 16, 2026
Power Platform Dataflows were never deprecated — just abandoned. With their parent product now stamped "no new features, ever," it's time to stop pretending.
Power Platform Dataflows are a zombie feature — never deprecated, never finished, with only the maker-facing tooling ever built — and with Power BI Gen1 officially legacy, the wise customer plans an exit before the zombie bites.
Coding agents started delivering what low-code only promised. What that does to Power Fx, app builders, and the whole calculus of who builds software.
PlusJanuary 23, 2026
Anthropic skipped building an M365 and caught everyone's attention with text files and a terminal. The GUI is no longer the enabler — it's the garden wall.
Apps and their GUIs have become the wall keeping AI away from where the value lives — data and code — which is why a terminal agent on plain text files beats the entire M365 Copilot suite.
PlusJanuary 30, 2026
The XKCD "is it worth the time?" matrix is obsolete: a monthly task only needs to save 5 seconds to justify an hour spent with an LLM and a terminal window.
Cloud convenience has deskilled us, and LLMs plus a command line are how you take back control of real computers — because every outage, bad Windows update, and cut undersea cable reminds you whose infrastructure you actually depend on.
March 27, 2026
Microsoft's release plan API broke, so I burned expiring Azure credits on GPT 5.4 Pro and built the search UI Microsoft should have shipped years ago.
When Microsoft can't be bothered to make its own release plan data searchable, a solo consultant with an AI coding agent and a few days can ship a better product experience than the trillion-dollar vendor's official site.
PlusApril 24, 2026
"Test Engine has near-zero usage." Microsoft's own deprecation note admits it: generic code plus frontier LLMs beat the low-code language built to save citizen developers.
Power Fx has already been beaten — frontier LLMs are too good at common programming languages for a niche Excel-flavored low-code language to compete, and Microsoft's own Test Engine deprecation is the written admission.
June 5, 2026
There are 100x more people who can write .md than .tsx. Once information workers touch something that looks like code, they're suddenly coding — without being programmers.
Coding agents are delivering what low-code only promised — democratized code — and the old dividing line of "you don't write much code" is gone: code work is becoming like driving, where everyone does it but few do it professionally.
June 26, 2026
Lovable built my entire app in 15 minutes on free daily credits. I spent longer than that fighting SharePoint's list formatting JSON.
Microsoft customers shouldn't settle for first-party app generators: Lovable shipped in 15 minutes what Power Apps tooling makes tedious, because its agent builds in the language LLMs actually speak instead of vendor-specific abstractions.
The gap between the agentic vision on stage and what a practitioner can actually deploy, govern, and afford today.
PlusFebruary 20, 2026
Microsoft shipped a flow designer marked 'moved to next release wave' straight into production, rolled it back in two days, and deleted the announcement. Who approved this?
Microsoft won't give Power Automate makers a real code view or AI development support because customers getting more productive with deterministic automation doesn't feed the Copilot-first story — so community tools have to do the vendor's job.
PlusFebruary 27, 2026
Copilot is all talk, no hands. OpenClaw has hands — and every instance is a ticking bomb whose blast radius equals the credentials you handed it.
Claws prove users want AI with hands rather than another chat sidebar, but tools plus triggers plus an unreliable LLM makes every deployment a ticking bomb — and Microsoft's tame, waitlisted answer won't capture the moment.
PlusApril 2, 2026
Satya says a Microsoft-built OpenClaw would be like launching a virus. Four weeks later, the exec who wired OpenClaw into his smart home is CVP of OpenClaw + M365.
Promoting the executive who publicly bypassed Microsoft's own OpenClaw security guidance exposes the Secure Future Initiative as theater — the reinforced behavior is that security warnings are for other people, not visionaries.
PlusMay 22, 2026
The same prompt retrieves 10 documents in a Declarative Agent and 3 in Copilot Studio. Sydney vs. Samba is the divide Microsoft won't document.
Which orchestration engine runs your Copilot agent — LLM-era Sydney or Bot Framework-era Samba — is decided by opaque platform barriers rather than conscious customer choice, and it materially changes what the agent can see, do, and how it's governed.
PlusJune 12, 2026
One prompt to Microsoft Scout cost me $4 in GitHub Copilot credits — and the agent doesn't show up in a single Agent 365 admin dashboard.
Microsoft's agentic product map is growing faster than anyone can track, while the flagship Autopilot burns $4 a prompt and is invisible to Microsoft's own agent governance layer — what's needed now is focus and stability, not more reimagining.
PlusJune 18, 2026
MCP is the USB-C of AI: beautiful when it works. The rest of the time you're yelling "NO, USE ADO, NOT GITHUB!" at an agent ignoring Microsoft's own MCP server.
AI agents aren't doing the work yet, but they've become extremely good at building tools — and MCP is simultaneously the answer and the problem, which is why purpose-built deterministic GUIs still beat negotiating with a chat surface.
Licensing spaghetti, capacity math, and the end of the CoE Starter Kit — the operational reality organizations have to manage while the product story keeps changing.
PlusMarch 20, 2026
The $5 Per App license was killed with one line in a PDF, then resurrected by sneaking "and new" into an edited blog post. Microsoft licensing in 2026, everyone.
Microsoft's licensing operation runs on saying as little as possible and never admitting a reversal — the half-GA E7, the Per App resurrection, and a 2,000-seat Copilot Chat cutoff are spaghetti thrown at the wall, not strategy.
PlusMay 15, 2026
The CoE Starter Kit is officially no longer maintained. In its place: raw inventory data with no templates, and guardian agents you won't be able to extend.
Managed Environments' commercial story is what killed the CoE Starter Kit, and its replacement — an inventory API with no patterns and unextendable guardian agents — proves your governance priorities must outlive the vendor's.
Brand damage, a decade of Dynamics 365, and why Markdown winning matters — the issues that stepped back from the weekly product churn.
January 9, 2026
Microsoft didn't plan for 2026 to be "the year of Microslop." But after a year of self-inflicted decisions, it starts the year weaker with every audience that matters.
Microsoft's reputational hole isn't one misstep but a series of deliberate choices that have left it weaker with business customers, consumers, recruits, employees, and society all at once.
March 6, 2026
Ten years ago you could buy Dynamics CRM 2016 and install it. Today the blockers are knowing what's actually available, what it costs, and whether it still exists tomorrow.
A decade of Dynamics 365 built a genuinely better platform underneath, but the product layer keeps going in circles — from the Insights bingo era to 15 sales agents that you need an LLM to parse.
PlusJuly 3, 2026
Markdown has become the source code of information work — and Microsoft's answer is a notes stack where the only export option is print to PDF.
Markdown won as the native language of AI-era information work, and Microsoft's half-hearted support isn't an accident — the Copilot Credits business model now profits from keeping content locked in token-hungry Office formats behind M365 toll booths.
UDPP26 and DynamicsMinds — what practitioners were saying to each other when the vendor wasn't steering the agenda.
April 30, 2026
"What's the point of low-code anymore?" was the question in every UDPP26 hallway discussion. My panel answer: most apps should not be built to begin with.
Building apps is now the easy part, so low-code's defensible value has moved to everything around the generated code — governance, lifecycle, auditability — and most apps shouldn't be built at all.
May 29, 2026
DynamicsMinds isn't a conference — it's a festival. And the biggest on-stage moment: someone finally said "canvas apps are dead" out loud.
DynamicsMinds shows the BizApps community is healthiest where the vendor isn't steering the agenda — honest "canvas apps are dead" talk, MCP in every other session, and Microsoft itself using Lovable instead of its own tools.
The free edition — Power Platform, Copilot, and the fine print.
Extended analysis, weekly issues, and the deep-dive reference pieces.
Direct advisory on Power Platform and Copilot strategy, governance and licensing.
