Skip to content
Codex · ChatGPT · OpenAIEpisode 155 · 26 August 2026 · 19:39

Is the era of apps ending? How AI is already replacing everyday software

Central question

Are separate apps, dashboards and CRMs still needed, when AI finds the data, takes the action and returns the result right there in the chat?

What you take away

The reader gets six cases from live projects where no interface was needed — and a working test for when a separate program is still justified: not the habit of seeing a list, but how fast the data has to arrive and how much of it has to be processed.

Main threads

What to watch for

1Take one report you produce regularly and assemble it from a chat with access to the data — before commissioning a screen for it.
2Check where an interface in your process is there for a reason and where out of habit: replace the list with the task «show me this data» and see what is actually missing.
3Before building a control panel, price it in servers, layout work and speed — and compare that with a chat on top of the same API.
4Verify a model's output on a sample rather than exhaustively: forty examples checked by hand give you more than demanding a hundred-per-cent match.
5Stop filling in SEO tags and page descriptions by hand — compare what a model produces with what you have now.
Signals to track afterwards
Partnerships between large models and app developers: whose code sits inside the finished product.
The fate of intermediate services between models and messengers — how long they last before being absorbed.
The arrival of devices and recorders you simply speak into, instead of screens.
How corporations approach rewriting internal software — and what it costs them.
What happens to search promotion if the familiar results page stops being the main entrance.
Most useful for
Founders and product leads deciding whether to build a separate interface for a process.People already assembling solutions in Codex or Claude Code who want to compare notes.Sales and finance managers whose work is tied to CRMs and accounting systems.Engineers and tech leads weighing the cost of maintaining an interface.Anyone who uses a chat daily and wants to see where it already replaces familiar programs.

Key takeaways

01:23A Hypothesis From Three Years Ago: the Layers Fade

The author recalls saying three years ago that with systems like ChatGPT, specialised tools would start to disappear. He named CRMs first — Salesforce, and amoCRM on the Russian-speaking market — and then the simpler layers: PDF generators, reporting-visualisation systems such as Power BI, marketing analytics.

02:08What Disappears Is Not the Company but the Detour

The caveat comes immediately: Salesforce and amoCRM are not going away tomorrow. What goes is not the product but the necessity of the step. Why should a sales manager open a CRM to find out whom to call, when the same answer comes from a chat — telephony integration included.

03:10A Layer Between a Model and a Messenger Does Not Last

OpenClove appeared as an adapter between strong models — OpenAI, Anthropic, Google Gemini — and messengers. The author said at the time that a great many processes do not need a messenger at all. The story faded as fast as it arrived, and the work moved into OpenAI Codex and Claude Code.

06:01There Is No Interface, and None Was Needed

The financial-market monitoring system lives on a server, integrated with Telegram and finance APIs, and pushes notifications into a bot. It has no control panel at all: updating a portfolio or pulling a report is a phrase in Telegram or a task in Codex.

08:03Where an Interface Is Nearly Impossible to Design, the Chat Wins

Voice-analysis projects run into the verification interface: whose voice is this. Designing one is months of work, and it still ends up as a single fixed layout. From a chat it works differently: asked to push the match to a hundred per cent, the system returns six hundred cases; the author takes forty and listens to short excerpts right inside Codex.

09:27The Program Written for the Sensor Turned Out to Be Redundant

A Mac program was written for the heart-rate sensor — and it kept having memory and synchronisation problems. Now the synchronisation runs through Codex: start the sensor for the night, download the data in the morning, build a report, compare it against three years of cardiograms, send it to the doctor.

10:41Reporting Better Than the Specialised System's

The author connected to the American QuickBooks over its API, and Codex downloaded the data and reconciled it against bank statements and PDF exports. Then: receipts pulled from Expedia, personal expenses separated from business ones, unfamiliar contractors identified across twenty or thirty transactions. No system, in his assessment, does this as easily.

13:32The More You Ship, the Less You Need a CMS

For the real-estate sites, a manageable CMS took a long time to build. On his own projects — ToTheMoon and Besolid — the opposite happened: the more that shipped, the clearer it became that a CMS was not needed. Filling in page tags and parameters by hand is odd now; a model such as Fable describes it all more carefully, re-checks the texts and indexes the images.

15:36The Caveat: There Are Places Where an Interface Is Mandatory

The author deliberately stops to say he is not removing the interface everywhere. There is the habit of a list — leads from a site, production and service requests, support tickets. The question is whether the list is genuinely needed, or whether a terminal you tell «show me this data» is enough.

16:42An Interface Is a Cost

The more projects there are, the more it pays to go without an interface: it is extra load, extra servers, and always problems with visualisation, layout and speed. The limits are real ones, though — how fast the data arrives and how much of it has to be processed.

17:14Corporations Will Stay on Their Own Software for a Long Time

The mass market and large companies will diverge. Corporations had trouble rewriting internal software long before AI. The example is Russian Railways: for years tickets were issued from terminal software where you all but typed the command by hand, and it was never rewritten.

What this episode is about

A solo episode about what happens to familiar software once there is a chat next to you that can not only answer but act. The author recalls the hypothesis he stated three years ago: specialised tools — CRMs, report generators, visualisation systems — will fade as a layer between a person and a result. Not vanish tomorrow, but stop being a mandatory step.

Six cases from his own practice follow, and the question in each is the same: was an interface needed here? Financial-market monitoring lives on a server and talks to its owner through Telegram — there is no control panel at all. A camera system recognises people, animals and plates and runs the house — no screen required. Voice-analysis projects run into the fact that a good interface for verifying voices takes months to design, while from a chat it happens immediately: forty short excerpts, an answer for each. A heart-rate sensor stopped needing a program of its own. Accounting assembled itself from statements and receipts better than the specialised system does it. Websites stopped needing a CMS, because a model fills in tags and descriptions more carefully than a person.

The author is explicit about the other side: he does not claim an interface is never needed. There is the habit of seeing a list of leads and tickets, and there are real limits — speed and volume. But an interface is always extra load, servers, layout and speed problems, and more and more often it pays to go without one. The mass market and large corporations will diverge here: big companies will sit on internal software they cannot rewrite for a long time yet.

The argument is not about whether apps will disappear, but about what stops being a mandatory step. As long as an interface is there for speed and volume, it stays; when it is there only out of the habit of seeing a list, it turns into extra servers, layout work and lag. The author leaves the question open and asks viewers for their own examples: where software migrates into a chat, and where it cannot migrate at all.

Episode transcript

The episode is in Russian; below is an English reading guide to the transcript (the full EN transcript is a machine translation). Voice matching applied to 34 segments: 34 identified, 0 mixed, 0 probable, and 0 unresolved.

Loading…