Custom Software Development in 2026: Why Large Enterprises Are Coming Back to Bespoke Builds Instead of Off-the-Shelf Products
Article date
11 11 2026
Article Author
Reading Time
8 minutes
The peak of emergency import substitution is behind us. Packaged products in critical environments have hit their architectural ceiling — not because they are bad, but because they were designed for a standard process. AI tools have narrowed the cost gap between a bespoke build and customizing someone else's product, though by less than vendors promise. In 2026 large enterprises choose custom development not out of a love of hand-built code, but because the alternative is an endlessly patched product that slows the business down. Below: what the numbers say, where custom development is justified, and what the customer has to bring to the table.
The market is large, but estimates differ by a factor of one and a half
Market participants in TAdviser's 2025 review put the size of custom software development at 200–300 billion rubles. Research by B1, cited by Vedomosti, gives 150 billion rubles for 2024 — including IT services, that is, on a broader methodology. The bond prospectus of FabricaONE.AI cites around 164 billion rubles for 2025 with average annual growth of 12% in 2022–2025. Three figures, three methodologies. If a consultant presents one of them as "the market size," ask what it includes.
On dynamics the sources diverge too, and that matters more than the absolute numbers. Mikhail Beloshapka, CEO of the digital production company Daleyе, tells TAdviser that average annual growth has been 20–25%, adding that "this growth is now harder for most players to sustain." TAdviser's forecast for 2026–2027 assumes 8–15% a year; Alexey Sushkov, technical director of the development department at Performance Lab, says 8–10%: the market is hitting a ceiling of available budgets and staff. KORUS Consulting estimates that IT market growth slowed to 10% against 21.9% in 2024.
The driver has changed, and here the sources agree. In 2022–2023 companies commissioned work to replace a departed vendor in a hurry. Now, in TAdviser's phrasing, the market has moved into a phase of "mature import substitution": companies have turned to rebuilding their platforms — migrations, integrations, reworking the architecture. Projects do not disappear, but they are split into stages and judged more strictly on outcomes.
On dynamics the sources diverge too, and that matters more than the absolute numbers. Mikhail Beloshapka, CEO of the digital production company Daleyе, tells TAdviser that average annual growth has been 20–25%, adding that "this growth is now harder for most players to sustain." TAdviser's forecast for 2026–2027 assumes 8–15% a year; Alexey Sushkov, technical director of the development department at Performance Lab, says 8–10%: the market is hitting a ceiling of available budgets and staff. KORUS Consulting estimates that IT market growth slowed to 10% against 21.9% in 2024.
The driver has changed, and here the sources agree. In 2022–2023 companies commissioned work to replace a departed vendor in a hurry. Now, in TAdviser's phrasing, the market has moved into a phase of "mature import substitution": companies have turned to rebuilding their platforms — migrations, integrations, reworking the architecture. Projects do not disappear, but they are split into stages and judged more strictly on outcomes.
A packaged product hits its ceiling wherever the process is non-standard
An off-the-shelf solution is built for a standard scenario. When a business has dozens of sites with different specifics, the product starts dictating the process instead of serving it. A public example is Petrovich, a DIY chain with 22 building and home-improvement centres: about 500,000 orders a month, more than 50% of orders placed online, tens of thousands of SKUs from small items to oversized goods. Its goods-flow system had been developed over years, but as volumes grew the existing approach to order management began to constrain development. KORUS Consulting and Petrovich-Tech considered options from ready-made products to custom development and chose their own OMS/IMS platform built on Temporal and a microservice architecture; the MVP is scheduled for launch in 8 months, with order-management logic being carved out of the running system without stopping operations.
The same logic applies at the level of entire system classes. TAdviser's review notes that a shortage of mature enterprise-grade packaged products in a number of classes pushes companies to build their own. X5 Group wrote its own WMS, Russian Post built its own CRM and corporate client services, and Magnit is developing its own customer data platform. For retail this is confirmed at the vendor level as well: KONCRIT from KORUS Consulting was created specifically for non-standard processes that fall outside conventional system classes. According to the developer, it handles more than 100,000 operations per hour and up to 200,000 SKUs, or up to 1 million for marketplaces; warehouse and yard management subsystems were implemented first, with transport on the roadmap.
The same logic applies at the level of entire system classes. TAdviser's review notes that a shortage of mature enterprise-grade packaged products in a number of classes pushes companies to build their own. X5 Group wrote its own WMS, Russian Post built its own CRM and corporate client services, and Magnit is developing its own customer data platform. For retail this is confirmed at the vendor level as well: KONCRIT from KORUS Consulting was created specifically for non-standard processes that fall outside conventional system classes. According to the developer, it handles more than 100,000 operations per hour and up to 200,000 SKUs, or up to 1 million for marketplaces; warehouse and yard management subsystems were implemented first, with transport on the roadmap.
The threshold beyond which customization stops being cheaper
The argument "I'll buy a packaged product and customize it" stops working at scale. While customization is local, it is cheaper. But past a certain volume of modifications you are paying as if for your own product while remaining inside someone else's architecture and someone else's roadmap. The vendor ships a release — you port the modifications again. The data model changes — you agree the changes with someone else's team.
This is calculated not by licence price but by three-year TCO: licences, modifications, integrations, maintenance costs and the cost of dependency. The threshold is different for every company, and it is set not by the volume of code but by how far the process departs from the standard one.
This is calculated not by licence price but by three-year TCO: licences, modifications, integrations, maintenance costs and the cost of dependency. The threshold is different for every company, and it is set not by the volume of code but by how far the process departs from the standard one.
AI closed the gap, but the gain is more modest than in the slide decks
This is the most overrated part of the story. Mikhail Beloshapka (Daleyе) estimates that copilot systems and models for log analysis speed up development "by 15–20 percent, no more." Alexey Kononov, head of development at BPMSoft, in Vedomosti puts time savings at the same 15–20% and suggests reinvesting it in refactoring and reducing technical debt. Internal testing at the T1 IT holding produced 20% across the development cycle. The numbers are notably higher where AI closes a narrow operation: at SberTech, AI agents in code analysis cut review time from five days to several hours.
The practical conclusion: building a business case on more than 20% productivity gain from AI means building in risk. And there is a downside vendors mention less often. Research by Veracode (as reported by Vedomosti) found that on average 45% of AI-generated code contains vulnerabilities from the OWASP list, and in Java up to 72%. Writing code faster is not the same as shipping a reliable system faster.
What AI does not change is the requirements on architecture. A solution designed for a standard process will not become yours just because code is generated faster.
The practical conclusion: building a business case on more than 20% productivity gain from AI means building in risk. And there is a downside vendors mention less often. Research by Veracode (as reported by Vedomosti) found that on average 45% of AI-generated code contains vulnerabilities from the OWASP list, and in Java up to 72%. Writing code faster is not the same as shipping a reliable system faster.
What AI does not change is the requirements on architecture. A solution designed for a standard process will not become yours just because code is generated faster.
Where to build and where to buy
Build when the process differs fr om the industry standard and that difference makes the business money; when you need control over architecture and data — especially in AI and regulatory environments, wh ere it matters where the data physically resides and who is accountable for the model; when the volume of modifications to a packaged product has already matched the cost of your own solution; when you need integration across disparate systems that the vendor will not take on.
Buy when the task is standard — ERP, electronic document management, HR records, standard CRM functions; when the decisive factor is speed to market with minimal investment; when the process does not affect the company's competitive position.
According to TAdviser, integration solutions and custom development for the corporate sector account for 60–70% of the market. The market already runs on this logic; the only question is where your own line between the two columns falls.
Buy when the task is standard — ERP, electronic document management, HR records, standard CRM functions; when the decisive factor is speed to market with minimal investment; when the process does not affect the company's competitive position.
According to TAdviser, integration solutions and custom development for the corporate sector account for 60–70% of the market. The market already runs on this logic; the only question is where your own line between the two columns falls.
When this will not work, and what the customer must provide
Custom development is not a sure bet. It will not work if:
• There is no product owner inside. Development without someone who sets priorities and is accountable for the outcome becomes an endless backlog. A contractor writes code but cannot stand in for the customer on decisions about business logic.
• The data is in poor shape. Any AI environment runs into the quality of reference data, duplicates in master data, the absence of a single customer key. That is a separate budget and separate months, and it has to be planned from the start.
• There is no support setup. Your own product has to be maintained: releases, monitoring, model updates, incident handling. If nobody owns that role, within a year the system is propped up by manual work.
• The budget covers an MVP while expectations cover production operations. The cost of a pilot and the cost of a system that holds the load in a closed environment with security requirements differ by a multiple.
• Requirements change faster than the team can design. What helps here is not custom development but short stages with measurable outcomes — and that is a matter of management, not technology.
What the customer must provide: a dedicated product owner with decision rights; access to real data and to process experts, not only to IT; willingness to accept staged delivery with fixed metrics at each stage; and an understanding that security and technology-independence requirements — the domestic software registry, operation in an isolated environment — add to the timeline.
• There is no product owner inside. Development without someone who sets priorities and is accountable for the outcome becomes an endless backlog. A contractor writes code but cannot stand in for the customer on decisions about business logic.
• The data is in poor shape. Any AI environment runs into the quality of reference data, duplicates in master data, the absence of a single customer key. That is a separate budget and separate months, and it has to be planned from the start.
• There is no support setup. Your own product has to be maintained: releases, monitoring, model updates, incident handling. If nobody owns that role, within a year the system is propped up by manual work.
• The budget covers an MVP while expectations cover production operations. The cost of a pilot and the cost of a system that holds the load in a closed environment with security requirements differ by a multiple.
• Requirements change faster than the team can design. What helps here is not custom development but short stages with measurable outcomes — and that is a matter of management, not technology.
What the customer must provide: a dedicated product owner with decision rights; access to real data and to process experts, not only to IT; willingness to accept staged delivery with fixed metrics at each stage; and an understanding that security and technology-independence requirements — the domestic software registry, operation in an isolated environment — add to the timeline.
Approach to choosing IT solutions
Split the portfolio into two columns: commodity functions go to packaged products; anything that departs from the industry standard and affects revenue or cost is a candidate for custom development. Calculate TCO over three years rather than the licence price: modifications, integrations, the cost of porting during upgrades, dependency on the vendor. Treat AI as a tool inside the process, not as the goal of the project: 15–20% acceleration is useful but does not replace architecture and data quality. And split projects into stages with a metric — in 2026 the customer pays for the outcome, not the process.
Questions and answers
"We run an isolated environment and external services are prohibited. Is AI even applicable?" Yes, if the model runs on your own infrastructure. We deploy PlanExpert and ProtokolExpert on-prem: speech recognition and data processing happen inside the perimeter and nothing leaves it. There is one constraint — hardware: local models require GPUs, and that has to be budgeted from the start.
"We already invested in a packaged product and customization. Do we throw it all away?" No. A working packaged product with modifications is not rewritten from scratch — which is exactly why public cases introduce new platforms in stages. At Petrovich, order-management logic is being carved out of the running system without stopping operations, with an MVP in 8 months. The same order makes sense for you: move a specific area into a separate service that over time replaces part of the functionality.
"How is this different from what our in-house team would build?" No difference in the outcome, but a difference in speed and cost of ownership. An in-house team knows the process better, but it is expensive to scale and hard to retain. The reasonable option is a hybrid: your own product owners and process experts, an external team doing the development, with competencies gradually transferred in-house.
"What if custom development does not pay off?" That becomes visible in the first stage. We do not start with a large project but with a narrow area with a measurable metric: processing time, error rate, cost per operation. If there is no effect in that area, a large project will certainly not produce one — and it is better to learn that in the first months than in a year.
"How do we justify ROI if we have no data on the cost of manual operations?" Then ROI cannot be calculated, and that is the first piece of work, not the last. Without a baseline — time per operation, error rate, cost of an hour of downtime — any business case rests on belief. Establishing a baseline for one process is usually cheaper than one development stage.
"We already invested in a packaged product and customization. Do we throw it all away?" No. A working packaged product with modifications is not rewritten from scratch — which is exactly why public cases introduce new platforms in stages. At Petrovich, order-management logic is being carved out of the running system without stopping operations, with an MVP in 8 months. The same order makes sense for you: move a specific area into a separate service that over time replaces part of the functionality.
"How is this different from what our in-house team would build?" No difference in the outcome, but a difference in speed and cost of ownership. An in-house team knows the process better, but it is expensive to scale and hard to retain. The reasonable option is a hybrid: your own product owners and process experts, an external team doing the development, with competencies gradually transferred in-house.
"What if custom development does not pay off?" That becomes visible in the first stage. We do not start with a large project but with a narrow area with a measurable metric: processing time, error rate, cost per operation. If there is no effect in that area, a large project will certainly not produce one — and it is better to learn that in the first months than in a year.
"How do we justify ROI if we have no data on the cost of manual operations?" Then ROI cannot be calculated, and that is the first piece of work, not the last. Without a baseline — time per operation, error rate, cost of an hour of downtime — any business case rests on belief. Establishing a baseline for one process is usually cheaper than one development stage.
Next step
Take the one system that slows the business down the most and calculate two numbers for it: the volume of modifications over the last 12 months and the cost of an hour of downtime or of a manual operation. If the first is growing faster than the second, it is a candidate for an architecture review. We are ready to go through such a system in a single meeting: you show the landscape, and we say where custom development will deliver an effect and where you are better off staying on a packaged product.