
Custom software development: what gets built, four types with real examples, and how to pick a route
As a founder, operations lead or head of IT, you have a process that has outgrown its spreadsheet or rented tool, and someone has suggested building something. Four public systems, with their users and costs on the record, show what a build carries before you commit to one.
What is custom software development, and what actually gets built?#
Custom software development is designing and building software for one organisation's own users and processes, and what gets built falls into four types: customer-facing apps, integration platforms, extended packaged software and systems of record. In everyday speech, the result is called custom software, bespoke software or a custom program. Meanwhile, the full definition and the question of who owns the code sit in what custom software development is and what you own.
Even national accounts draw this line. The US Bureau of Economic Analysis (BEA), in its handbook chapter updated December 2024, splits software investment three ways. It counts "prepackaged software", "customized software" bought from software firms, and software a business "develops in-house". As a result, a business has more than a yes or no to make. It can buy a product, configure one, connect what it has or build.
The types of custom software differ in three plain ways. They are who uses the software, how long it must run, and how much of it you own.
- Customer-facing apps are used by people outside the organisation, so success is counted in sign-ups and use.
- Integration platforms are used by other software, and the connection is the product.
- Extended packaged software is a rented platform shaped to your process with settings, added code or low-code screens, which are built by arranging parts rather than writing code.
- Systems of record hold the data the organisation runs on, and they can stay in service for decades.
The UK government's Service Manual on choosing technology starts from the same split. In particular, it tells service teams to learn "what you'll need to build and what you can buy or get for free" before they start.
Which custom software examples can you check against a public record?#
Each of the four types has a public example with a date and a record: IRS Direct File, GOV.UK Notify, Microsoft Power Apps and the federal legacy systems GAO audits. Most lists of bespoke software examples name apps you cannot look inside. By contrast, these four publish their users, costs or limits, so you can judge each type on its own numbers.
| Type | Real example | What the public record shows |
|---|---|---|
| Customer-facing app | IRS Direct File, 2024 to 2025 | 296,531 returns in its second season, at $138.27 a return |
| Integration platform | GOV.UK Notify, since 2016 | 1,765 organisations and 12,979 services send messages through it |
| Extended packaged software | Microsoft Power Apps | Custom business apps on a rented platform, with a 180 second timeout on each outgoing request |
| System of record | GAO's 11 legacy systems, 2025 | About 23 to 60 years old, about $754 million a year to run |
Source: US Department of the Treasury, 2 October 2025; GOV.UK Notify performance page, read 1 November 2026; Microsoft Learn, 12 January 2026; GAO-25-107795, July 2025.
In short, the table answers what kinds of custom software applications exist and what each one carries. Each row then gets its own section below, in the same order.
What did IRS Direct File, a customer-facing custom app, actually serve?#
The US Department of the Treasury reported on 2 October 2025 that Direct File accepted 140,803 returns in its pilot and 296,531 in its second season. The pilot ran in 12 states, and the second season ran in 25.
Direct File let people file a federal tax return for free, straight with the IRS. It worked as an interview in the browser. The IRS also published its code on GitHub, so the build itself is open to read. However, the same Treasury report says "Direct File had low overall participation and relatively high costs". By the Treasury's 2025 count, Direct File carried less than 0.5 percent of about 146 million returns for tax year 2024.
Then the service stopped. In a report of 19 March 2026, the Treasury Inspector General for Tax Administration confirmed that the IRS suspended Direct File after the 2025 filing season. In short, use doubled and the service still closed. Doubling a small number left it small.
140,803
Pilot, tax year 2023
296,531
Second season, tax year 2024
Direct File returns more than doubled between seasons and still stayed under 0.5 percent of returns filed.
| Option | returns accepted by the IRS |
|---|---|
| Pilot, tax year 2023 | 140,803 |
| Second season, tax year 2024 | 296,531 |
That is the first lesson of a customer-facing build. Although it can work well, it can still fail on the one measure that counts for it, which is use.
Why does the number of users decide what custom software development costs per use?#
The Treasury's 2 October 2025 report puts Direct File at 225.85 US dollars a return in its pilot and 138.27 US dollars a return in its second season. In that report, the pilot cost at least $31.8 million and the second season reached $41 million.
So the total cost rose by less than a third while the returns more than doubled. As a result, each return carried a smaller share of the bill. Most custom builds follow the same shape. Because the team, the servers and the support desk cost much the same at any scale, more users mean less cost per user.
Show data table
| Step | Change | Running total |
|---|---|---|
| Pilot | 225.85$ | 225.85$ |
| Returns doubled, same cost | -118.61$ | 107.24$ |
| Cost rose to $41M | 31.03$ | 138.27$ |
| Second season | 138.27$ |
The cost per return fell from $225.85 to $138.27 as returns grew.
The same 2025 report shows how far the plan drifted from the record. The Treasury's report says the early IRS estimates came to at most about $16 a return, depending on how many people took part. In practice, the cost per return was many times that. The Treasury's 2025 report also warns that its $41 million understates the true cost. Some IRS offices were left out of the count. Therefore, for any custom build, the expected number of users is the first number to test, before the build estimate or a software development RFP.
How does one fixed build cost spread across your own number of users?#
As a worked example, the same 41 million US dollar season spread over 1,000,000 returns would be 41 US dollars a return, and over 100,000 returns 410 US dollars. Those two cases are a modelled example. They divide the Treasury's $41 million of 2 October 2025 by a number of returns, so they are not reported figures.
The arithmetic is simple, and you can run it on your own case. First, take the yearly cost of building and running a system. Then divide it by the users you expect in a year. For example, the Treasury's 2025 count of 296,531 returns gives $138 a return. That matches the Treasury's own figure. In the same modelled example, 5,000,000 returns would bring the $41 million down to $8 a return.
Spread a yearly cost across your users
Put in the yearly cost of building and running a system and the users you expect in a year. The defaults are Direct File's $41 million season and its 296,531 returns.
Cost per user
$138
A modelled example built from the Treasury's $41 million figure of 2 October 2025. An estimate, not financial advice. Modelled, not measured.
Show data table
| Item | Value |
|---|---|
| 100,000 returns | 410$ |
| 296,531 returns (actual, tax year 2024) | 138$ |
| 1,000,000 returns | 41$ |
| 5,000,000 returns | 8$ |
The same $41 million costs $410 a return at 100,000 returns and $8 at 5,000,000.
Then hold that unit cost against the per-seat quote for a packaged product that does the same job. In particular, use the user count you can defend, not the one you hope for. Direct File's early estimates assumed far more people than came. Because the cost is fixed, a tenfold miss in users is a tenfold miss in unit cost.
What does custom software look like when its whole job is connecting other systems?#
The Government Digital Service's performance page shows GOV.UK Notify serving 1,765 organisations and 12,979 services, and on 28 October 2026 it sent 4,433,108 emails. By the same page, read on 1 November 2026, it also sent 4,235,545 text messages and 74,398 letters that day.
Notify has few screens of its own. Instead, service teams call it from their own software through an API. An API, or application programming interface, is a set of calls one program uses to ask another for work. The Government Digital Service's developer docs say "GOV.UK Notify is a Government-as-a-Platform service". Also, its features page describes sending emails, text messages and letters through that one service.
Show data table
| Item | Value |
|---|---|
| Emails | 4,433,108 |
| Text messages | 4,235,545 |
| Letters | 74,398 |
Emails and text messages made up nearly all of the day, and letters a small share.
That is what an integration platform means in practice. The custom part is the software between systems that moves work from one to the other. Each service team writes a few calls, and Notify owns the delivery. As a result, one build carries thousands of services. Its performance page, read 1 November 2026, counts 13.9 billion messages since May 2016. That is the scale Direct File never reached. For example, most businesses meet this type as the link between a store, a customer database and an accounts package, and a logistics firm meets it as custom logistics software between its warehouse system and its carriers. In practice, it is often the right first build, because the systems on either side already exist.
When is customized software enough, and where does configuration stop?#
Microsoft documents Power Apps, in January 2026 limits, as a way to build custom business apps on a rented platform, with a 180 second timeout on each outgoing request. Its Power Apps overview says you can "build custom business apps" that connect to your data.
Customized software systems like this are packaged platforms shaped to your process. However, the difference from building is ownership. Built software is made for you, and the code is yours. Customized software is a rented platform set up or extended to fit, and the platform's limits stay in charge. Customised software, in British spelling, is the same thing.
Those limits are written down. For example, Microsoft's limits page, updated 12 January 2026, sets that 180 second timeout. Meanwhile, the code apps overview, updated 24 September 2026, adds a code route. Developers can build with React or Vue, two common web frameworks, and still run on the managed platform. Also, Microsoft's September 2026 code apps page says those apps reach "1,500+ connectors" and follow the organisation's platform rules, such as app sharing limits.
In short, you can extend a long way, but the platform still sets the walls. Read its limits page before you design the process around it, not after.
| Built software | Customized software | |
|---|---|---|
| What it is | Made for you | A rented platform set up or extended to fit |
| Ownership | The code is yours | The platform's limits stay in charge |
| Example | IRS Direct File | Microsoft Power Apps |
| A written limit | None from a rented platform | 180 second timeout on each outgoing request |
How long does a custom system of record stay in service, and what does it cost to keep?#
The US Government Accountability Office reported in July 2025 that its 11 most critical legacy systems were 23 to 60 years old and cost about 754 million US dollars a year to run. In the same report, GAO-25-107795, the 2019 list ranged from 8 to 51 years and cost about $337 million a year.
Show data table
| Item | Value |
|---|---|
| 2019 list, youngest system | 8 years |
| 2019 list, oldest system | 51 years |
| 2025 list, youngest system | 23 years |
| 2025 list, oldest system | 60 years |
The youngest system on the 2025 list, at 23 years, is older than the youngest on the 2019 list, at 8.
The two lists are not the same systems, so the rise does not track one system getting dearer. Instead, it shows what the oldest custom software systems cost as a group, and that group got older and dearer.
about $337 million
2019 list
about $754 million
2025 list
GAO's most critical legacy systems cost about $337 million a year in 2019 and about $754 million in 2025, across two different lists.
| Option | US dollars a year, two different lists |
|---|---|
| 2019 list | about $337 million |
| 2025 list | about $754 million |
Source: GAO-25-107795, July 2025
Age brings a staffing problem as well as a technical one. In the same July 2025 report, GAO says eight of the 11 systems use outdated languages. Four run on unsupported hardware or software, and seven run with known security holes. For instance, the Treasury systems on the list run on COBOL, the Common Business Oriented Language, and on Assembly, a language close to the machine. GAO notes these languages have "a dwindling number" of people with the skills to support them.
The wider picture points the same way. GAO's July 2025 report says about $83 billion of planned federal information technology (IT) spending for fiscal year 2025 was for running and maintaining systems. That is 79 percent of the total, by the same report. Therefore a build decision is also a decision about who will look after the system for decades, a choice the custom enterprise software development process puts to the client at launch. A business with an ageing core faces the same sum at a smaller scale. That is also why maintenance and support is its own line of work.
How does a business pick its custom software development route: buy, configure, integrate or build?#
Three answers pick the route: how many people will use the software, how long it must stay in service, and how closely a platform you can rent already fits the process. Each answer points back to one of the records above.
First, the user count. Direct File shows that a build's unit cost falls only when use grows. So a small user count favours a product you rent. Second, the life span. GAO's 2025 list shows a core system can run for 60 years. So a long life favours owning what you can maintain. Third, the platform fit. Power Apps shows how far a platform stretches before its limits decide for you.
Then read the answers together:
- Buy when the process is common to many businesses, such as payroll or email.
- Configure or extend when a platform covers most of the process and the users are few to moderate.
- Integrate when the systems you need already exist and the gap is the link between them, as with Notify.
- Build only when the user count is high, the life is long and no platform fits the process well.
Each route also has a cost. A bought product follows its vendor's roadmap and renewals, not yours. A configured platform stops at its limits, and an integration needs fixing whenever a system on either side changes. However, a build is the only route where you own the code, and that ownership includes the upkeep for its whole life.
Similarly, the UK government's Service Manual gives the same order for public services. Map the parts of a service, then build only what you cannot buy or reuse. For the money side, what custom software development costs works through the figures. Meanwhile, the custom software development timeline covers the weeks. Also, when a product and a build are both on the table, the custom software versus off-the-shelf comparison sets them side by side. Finally, when all three answers point to a build, how to choose a custom software development company covers the vendor, and a custom software development engagement sets the scope, the team and the upkeep plan.
When is custom software development the wrong fit?#
Custom software development is the wrong fit when a rented product already covers the process, when few people will use the result, or when nobody will own its upkeep. Each case has a better route.
When the process is common, buy it. Direct File is the case in point. The Treasury's report of 2 October 2025 recommends that the IRS refocus on Free File, the existing program that "operates at little cost to the government". The packaged route was there all along.
When the users are few, configure a platform. For instance, a team app for a few dozen people fits a platform such as Power Apps. The vendor then keeps that platform patched. Although the platform's limits will bind in time, they rarely bind a small internal tool first.
When the system will have no owner, do not build it. In July 2025, GAO found seven of its 11 legacy systems running with known security holes. A system with no owner ages into that list. Instead, rent a managed product and let its vendor carry the upkeep, or connect the tools you already pay for.
| Case | Better route | What the record shows |
|---|---|---|
| The process is common | Buy it | The Treasury recommends that the IRS refocus on Free File |
| The users are few | Configure a platform such as Power Apps | The vendor keeps the platform patched |
| The system will have no owner | Rent a managed product, or connect the tools you already pay for | GAO found seven of its 11 legacy systems running with known security holes |
Where should your custom software decision go next?#
The next reads cover the full definition and ownership, the cost, the timeline and the build-versus-buy comparison that this guide only touches. So pick the one that matches the question the route test left open.
For who owns the code and what the term covers, read what custom software development is. For an ageing system of record, choosing a legacy modernisation partner starts from where GAO's systems are now. When the gap is a link between tools you already run, API integration work covers that route.
You may need none of these. The four public records above are enough on their own to place your case on one route.