Three snapshots of the same vendor pricing pages, June 2023, mid 2024 and August 2026, show identical numbers. Retool Business is still 50 US dollars per builder and 15 per internal user. Power Apps is still 20. What changed instead was a second meter, for AI, bolted on top of the unchanged seat. This report prices what it actually costs to ship and maintain a signed, auto-updating desktop client instead, and puts the crossover at 52 seats against Power Apps Premium and 62 against Retool Business.
The claim under test was simple enough to check. AI-assisted coding collapsed the cost of writing an internal application, so the per-seat platforms that host those applications should have repriced. They did not. Retool's Business plan billed 50 US dollars per builder per month and 15 per internal user on the June 2023 archived pricing page, the same on the May 2024 page, and the same today. Power Apps has charged 20 dollars per user per month across the June 2023, June 2024 and current pages. Dataverse database capacity has been 40 dollars per gigabyte per month since at least June 2024. Three snapshots, three years, three flat lines.
What arrived instead was a second meter. Retool now allocates AI credits by tier, sells additional packs, and bills agent time by the hour at rates that vary by model. Microsoft ends seeded AI Builder credits on 1 November 2026 and directs customers to buy Copilot Credits at one US cent each on the pay-as-you-go meter, with no automatic conversion and no reduction in the seat price to compensate. From that date a Power Apps Premium customer pays the unchanged 20 dollars per seat, loses the AI allowance that seat used to carry, and buys AI capacity as a third line. Paying rent twice is now the published architecture of the product rather than an inference drawn from it.
The honest version of the build case is narrower than the one usually made, and it belongs here rather than at the end. A self-built client does not replace the platform. It replaces the interface. Seat rent also buys a permission model administered outside the application, an audit trail of who changed what and when, an availability commitment, a vendor with a security questionnaire already on file, and a data loss prevention integration compliance already accepts. For regulated data, or anywhere an auditor will ask who approved a change in March, that is worth real money and this report's arithmetic does not apply.
Vendor pricing pages were read on 25 August 2026 and compared against Internet Archive snapshots of the same pages from June 2023 and mid 2024. Everything traced to a vendor page, a project repository, an operating system vendor's own documentation, a standards body, a published independent audit, or a survey with a stated sample is labeled fact. Everything else is labeled as my assessment or as a hypothesis. No third-party estimate label appears anywhere in the report, because no research firm publishes a number it needed.
Retool Business billed 50 US dollars per builder per month and 15 per internal user in June 2023, in May 2024 and today. Retool Team was 10 and 5 across all three. Power Apps has charged 20 dollars per user per month on every one of the three pages. Dataverse database capacity has been 40 dollars per gigabyte per month since at least June 2024. The rate card did not respond to anything that happened in AI-assisted development.
Retool's current plans allocate AI credits by tier, 250 on Free through 3,000 on Business, pooled at account level, renewing monthly and not rolling over. Additional packs are sold to paid plans and agent time is billed by the hour at model-dependent rates. Microsoft ends seeded AI Builder credits on 1 November 2026 and requires Copilot Credits instead at one US cent per credit. Neither vendor cut the seat price to compensate.
Copilot Studio's documentation records that on 1 September 2025 the common currency for agents changed from messages to Copilot Credits, and states plainly that there is no change in the quantity per prepaid pack or to the pay-as-you-go rate. The 200 dollar pack bought 25,000 messages in 2024 and buys 25,000 credits now. The dimension was renamed and re-denominated while the money stayed where it was.
The Power Apps per-app plan at 5 US dollars per user per app per month sat on the June 2023 pricing page and is absent from the current one. Licensing practitioners report it reached end of sale in January 2026 through a line in a licensing guide changelog rather than an announcement. The direction of travel is consolidation upward into a single premium seat, which is the opposite of repricing away from seats toward usage.
Published costs to run a signed, auto-updating Windows desktop client come to roughly 492 US dollars a year: code signing, crash reporting and a paid Workers plan. Everything else is engineering labour, which nobody publishes. At an estimated fifteen engineer-days a year the annual total is about 12,500 dollars, and the client becomes cheaper than Power Apps Premium at 52 seats and cheaper than Retool Business at 62.
The first is whether the application needs a browser capability WebKit does not implement, because Tauri binds to the operating system webview and therefore ships WebKit on macOS, iOS and Linux. The second is whether the team ships Linux, where Tauri's own documentation records blank windows, flicker and silent software rasterisation on Nvidia hardware. Two noes means Tauri. Either yes means Electron.
Chromium runs on Windows and Android through the WebView2 runtime, which Microsoft's own distribution documentation states is not the Edge browser and cannot be substituted for it in production. WebKit runs on macOS, iOS and Linux. The Linux build therefore behaves like the Mac build, not like the Windows one. A team that tests on Windows and Linux and assumes macOS will follow has tested one engine twice.
Microsoft's documentation now states that Extended Validation certificates no longer bypass SmartScreen, that paying a premium for EV solely to avoid warnings is no longer justified, and that reputation builds only through download volume, taking several weeks and hundreds of clean installs from a wide audience. An internal application with sixty users will never reach that threshold on the consumer path. The same page names the enterprise escape hatches, which is where the answer actually lives.
Trusted intranet locations are not subject to SmartScreen review, IT administrators can submit files for accelerated trust, and Windows Defender Application Control treats a managed installer such as Intune as an allowlisting authority. Every one of those is an IT function. A business unit cannot ship a binary; IT can ship it trivially. The seat rent was purchasing an implicit exemption from that internal conversation, priced at hundreds of dollars per user per year without anyone naming it.
Pew Research Center, using a probability-based panel, found in September 2025 that 21 percent of US workers say at least some of their work is done with AI, and in February 2025 that 9 percent use an AI chatbot at work weekly or more while 29 percent had not heard of them. Microsoft's Work Trend Index, a survey the vendor commissioned and published, reported 75 percent of knowledge workers using AI. Both are cited with their sponsor and method named.
A desktop client costs a fixed amount per year regardless of headcount. A per-seat platform costs a variable amount that scales with it. Somewhere they cross. Published fixed costs are 492 US dollars a year. Everything above that is annual maintenance labour at 800 dollars a fully loaded engineer-day, and that single unmeasured number decides the whole comparison. Power Apps at 240 dollars per seat per year, Retool Business at three builders on 600 and end users on 180. List rates only, steady state only, year-one build cost excluded.
A client costing about 5,300 US dollars a year to keep running crosses under twenty seats against both Power Apps Premium and Retool Business. This assumes the runtime upgrade treadmill goes smoothly and the installer needs little attention.
Roughly 5,300 USD a year to run
About 12,500 US dollars a year: eight days on two forced Electron major upgrades plus dependency patching, three on the signing pipeline and installer, four on client-attributable support. Cheaper than Power Apps Premium at 52 seats and cheaper than Retool Business at 62.
Roughly 12,500 USD a year to run
At about 24,500 US dollars a year the crossover rises past a hundred seats and the platform is the right answer for most internal applications. The band is wide because the deciding variable is unmeasured, and any report that gives a single number here is hiding that.
Roughly 24,500 USD a year to run
Getting from a generated scaffold to a signed, packaged, auto-updating client that survives an endpoint security review is estimated at 16,000 to 32,000 US dollars. It does not recur, and a finance function should see it alongside the recurring figure rather than instead of it.
16,000 to 32,000 USD, non-recurring
At 200 seats the arithmetic is not close. Power Apps Premium is 48,000 US dollars a year against a client at roughly 12,500, and the backend those seats were hosting costs 60 dollars a year on a Cloudflare Workers paid plan, or effectively nothing on Lambda, at a reference workload of 176,000 requests a month. At 2,000 seats the volume rate of 12 dollars per user per month is 288,000 a year against a client cost that has not moved. The self-built client does not become 23 times better at that scale; it becomes 23 times cheaper relative to the alternative, because one side of the comparison is a fixed cost and the other is not. Below about twenty seats the platform is simply the right answer and no amount of architectural preference changes that. The largest single source of error for any specific buyer is the enterprise discount off list, which is invisible from outside: at a negotiated rate half of list, every seat count here doubles.
For a meaningful share of internal applications the answer is a browser tab, and saying so is more useful than the rest of the report. An installed client earns its cost when at least one of four things is true: the application must work without the network, it must reach the local filesystem or a device, it must run or notify in the background, or it must be present as an application rather than as a tab the user closes by accident. A team that cannot name which of those four applies should not be building a client.
Any application that is read, filter, enter, submit. No install, no signing, no allowlisting entry, no update mechanism to build. It fails on offline use, on the local filesystem beyond downloads, on protocol handling, on tray or background work, and on presence in the taskbar.
The right answer more often than anyone admits
Adds a dock or taskbar entry, a window without browser chrome and an icon. Supported in Chrome and Edge on Windows and macOS, and in Safari on macOS 14 and later through Add to Dock. Still a web page: storage remains subject to browser eviction, and on WebKit that means the seven-day script-writeable storage rule applies.
The strongest objection to the whole thesis
Small binaries, a real Rust privilege boundary, and two published independent audits by Radically Open Security under NLnet funding. The cost is two rendering engines to test, Linux graphics failures the project documents itself, and a mobile plugin set with named gaps including the updater.
Pick it when Linux is out of scope
One bundled Chromium everywhere deletes the divergence problem class, and a risk committee will recognise Visual Studio Code, Slack, Signal and Figma on the official showcase. The cost is a larger footprint, at least two forced major upgrades a year on the eight-week cadence, and a security history in which context isolation has been bypassed repeatedly, most recently across thirteen advisories published in July 2026.
Pick it when Linux is in scope
A natural fit for a team that already writes Go and does not want Rust in the stack. Version 3 is beta as of mid 2026, published with alpha release tags. Choosing pre-release infrastructure for a system somebody has to own for five years needs a reason beyond language preference.
Ruled out on maturity, not on merit
A Microsoft-supported path for a Microsoft-shop team, sharing Razor components with an existing web application. It ties the client to the .NET release cadence and the Microsoft toolchain, which is either the point or the objection depending on the organisation.
The default if the shop is already .NET
One architectural rule removes most of the divergence problem and it is made in the first week of the first project: the webview is a rendering surface, and anything touching the machine goes through the native layer. Adopted at the start it costs nothing, because a team that never reaches for the File System Access API never discovers WebKit lacks it. Adopted after the interface is written it is a rewrite. The report also names the sequencing rule that decides whether any of this ships at all: get the signing identity, the Intune deployment ring and the endpoint security exception agreed before writing the application, not after. The framework decision can wait; the distribution decision cannot, because it is the one with a queue in front of it and the one that can return a no. Three scenarios run to 2028, and the tripwire worth watching is the cheapest to check: whether Power Apps Premium list price changes for the first time since 2023.
This isn't a vendor summary. Every sentence is labeled by what stands behind it: verified fact, vendor claim, third-party estimate, my assessment, hypothesis, or scenario. Sources are numbered and clickable. Forward-looking sections use scenarios with observable tripwires, not forecasts. It's the same method behind every market assessment I write.
Thirty-six pages, built from public sources with no client brief and no interviews. Read it in the browser or take the PDF.
Each report here answers a real question, directed and researched against public sources and evaluated against a stated assumption, then delivered as Word and PDF. If you're weighing a platform, sizing a category, or defending a number to a board, tell me the decision behind it and I'll tell you honestly whether a report is the right tool.
Commission an assessment