# Custom software with AI: when to build it and when to buy

> With AI, a small business can build tailored tools in a matter of days. When it really pays off, which costs the prototype does not show and how to get started without surprises.

- Published: 2026-10-21
- Who it is for: SME owners and managers
- URL: https://cloudfactory.it/en/articles/custom-software-ai-smes/

## Should we build it ourselves?

In 2026 many SMEs are asking the same question: why pay a subscription for software we only half use, when AI lets us build one that fits us exactly? The short answer is that sometimes it genuinely pays off. The full answer depends on what the program has to do and on who will look after it afterwards.

Writing code has never been cheaper. Keeping a program running for years, securely and with customer data inside it, costs as much as it always did. Here we set out the data, a much-cited case and a practical grid for making the call.

## What happened to software in 2026

Software is not in crisis: the per-seat licensing model is. In February 2026, after the launch of new AI agents, around $285 billion of value was wiped off tech stocks in 24 hours ([UC Today](https://uctoday.com/will-ai-agents-break-the-project-management-software-model)). Analysts called it the "SaaSpocalypse".

Investors' reasoning is simple. If an AI agent does the work of several people, a company buys fewer licences, and if it can have an internal tool built for it, it buys none at all. The hardest hit were "horizontal" products: project management, CRM, marketing automation, HR ([Long Angle](https://www.longangle.com/alts-education/software-ai)).

The numbers also tell the other half of the story:

- corporate software spending grew by 15% in 2026, to $1.4 trillion, even as valuations collapsed ([UC Today](https://uctoday.com/will-ai-agents-break-the-project-management-software-model));
- management systems, ERP, security and data platforms held up, because they are embedded in business processes and replacing them is expensive ([Long Angle](https://www.longangle.com/alts-education/software-ai));
- according to the Wall Street Journal, cited by Long Angle, many large companies are using AI to build small apps and customisations, not to throw out their vendors.

One detail matters directly to SMEs. Monday.com chose to move towards large enterprises precisely because small businesses are the first to replace a subscription with a home-made tool ([UC Today](https://uctoday.com/will-ai-agents-break-the-project-management-software-model)). In short, the pressure is strongest where companies are most agile.

## Why building is now genuinely possible

With tools such as Claude Code, Lovable or Replit you describe what you need and within a few hours you have a first version to try. The most striking case of October 2026 is ArtCraft: a small team used AI to build open source alternatives to seven Adobe programs in about a week ([Open](https://open.online/2026/10/12/crafting-apps-adobe-photoshop-premiere-gratis/)). The authors themselves, however, describe them as alpha versions, not ready for professional work.

The interesting point, as [Luciano Ballerano](https://lucianoballerano.substack.com/p/e-se-il-software-ce-lo-facessimo) notes in his newsletter *Umani per ora*, is not remaking Photoshop. It is that small needs become manageable: too specific for an off-the-shelf product, too modest for a traditional project. An app that saves three people an hour a week has no market, but it can pay for itself in a few months.

The figures from the Retool 2026 report point the same way ([Retool](https://retool.com/blog/ai-build-vs-buy-report-2026)):

- 35% of respondents have already replaced at least one SaaS tool with one built in-house;
- the categories most often replaced are workflow automations (35%), internal admin tools (33%) and analytics dashboards (29%);
- around half of those who have put a tool into production say they save six or more hours a week.

They should be read with caution. The sample is 817 people who already use Retool, and so are more inclined to build than average. The same report also shows that only 31% generate complete apps "from a prompt": most use AI to write pieces of code that they then integrate and check.

## The costs the prototype does not show

The prototype works, and that is precisely the problem: it works even when it is not secure. There are four costs that never appear in the demo.

**Security.** IOActive analysed almost 20,000 code samples produced by 27 AI models and tools: 31.6% were fully exploitable by an attacker and no model produced 100% secure code ([IOActive](https://www.ioactive.com/the-security-gap-in-ai-generated-code/)). The weakest areas are exactly the ones that matter in a business: authentication, encryption, server configuration. This is not theory: in a scan of 1,645 apps built with Lovable, 170 left their database readable and writable by anyone, because nobody had set the permissions ([Cloud Security Alliance](https://labs.cloudsecurityalliance.org/research/csa-research-note-ai-codegen-vulnerability-debt-20260406-csa/)).

**Maintenance.** A program that is actually used has to be updated, backed up and adapted when a connected service changes. The same Retool report warns that the time saved can be absorbed by maintenance and by vulnerabilities to manage, and 26% cite the maintenance burden as a brake on automation ([Retool](https://retool.com/blog/ai-build-vs-buy-report-2026)).

**Shadow IT.** 60% of the builders surveyed by Retool created tools outside IT's control in the past year. In an SME this means spreadsheets and apps that only one person knows: when that person goes on holiday or changes jobs, the process stops.

**Platform dependency.** Many app generators also host the code and the data. Before starting, it is worth knowing whether they can be exported, where the data resides and how much it will cost to stay once the app has become indispensable.

## The Klarna lesson

Klarna is the most cited case of "we'll build it ourselves", and above all it teaches you to tell the announcement from the outcome. In September 2024 the CEO told investors the company was switching off Salesforce and Workday, with the idea of rebuilding their functions using AI ([Salesforce Ben](https://www.salesforceben.com/klarna-salesforce-workday-partnership-called-off-amidst-major-gen-ai-overhaul/)). The questions the industry raised straight away were the right ones: where the data lives, and who guarantees compliance and control.

On customer service the story is clearer. After handing most chats to an AI assistant and freezing hiring, in May 2025 Klarna started recruiting human agents again: the CEO admitted that AI cost less but delivered lower-quality service ([Entrepreneur](https://www.entrepreneur.com/business-news/klarna-ceo-reverses-course-by-hiring-more-humans-not-ai/491396)). Gartner predicts that by 2027 half of the companies that cut staff because of AI will rehire for similar roles ([Gartner](https://www.gartner.com/en/newsroom/press-releases/2026-02-03-gartner-predicts-half-of-companies-that-cut-customer-service-staff-due-to-ai-will-rehire-by-2027)).

For an SME the lesson is a practical one. Replacing is fine, as long as you know what you lose: the hidden functions of the old system, human oversight at sensitive moments, the compliance the vendor used to guarantee.

## Build or buy: a grid for deciding

In most cases the right answer is neither "everything custom" nor "everything on subscription". It makes sense to keep the solid systems and build around them whatever wastes time. The table helps show which way each individual tool leans.

| Question | Leans towards building | Leans towards buying |
| --- | --- | --- |
| Is the process yours, or the same for everyone? | Specific to the way you work | Standard: accounting, payroll, e-invoicing |
| How many of the product's features do you use? | A small fraction, but you pay for all of them | Most of them, and you need them |
| What data does it handle? | Internal and not very sensitive | Customer, health or financial data with legal obligations |
| What happens if it gets it wrong? | You lose an hour | Fines, operational downtime, exposed data |
| Does it need to connect other systems? | Yes, today it is done by copy and paste | No, the integration already exists |
| Do the rules change often? | No, the process is stable | Yes, and the vendor updates on your behalf |
| Who will maintain it? | There is an owner, in-house or external | Nobody has the time or the skills |

If most answers fall on the left, building is a reasonable choice. If they fall on the right, it is better to buy and, if anything, add a custom screen or automation on top.

## How to get started without getting hurt

The approach that works is to start small, with dummy data, and raise the level of control as the tool becomes important.

1. **Pick a small, measurable problem.** A repetitive task that currently costs a few hours a week, not the management system.
2. **Write down how you do it today.** What data goes in, what result has to come out, what exceptions occur. This is the most important part, and only the people who know the work can do it.
3. **Build a prototype with sample data.** Test it on cases where you already know the right answer.
4. **Before using real data, have the code reviewed.** Permissions, access, credentials, backups and where the data resides.
5. **Appoint an owner.** Someone, in-house or external, who knows how it works and who updates what.
6. **Measure after three months.** Hours saved on one side, hours of maintenance on the other.

At Cloudfactory we work in exactly this middle ground. We help companies turn an idea or a home-made prototype into a tool that is secure, documented and connected to existing systems, and to decide when it is better to stay with an off-the-shelf product. It is the same approach we use on ourselves: several tools we rely on every day were built this way, replacing subscriptions we only half used.

## Frequently asked questions

### Can an SME really replace subscription software with something built with AI?

Yes, especially for automations, internal admin tools and dashboards. It makes less sense for accounting, payroll and systems subject to frequently changing regulations.

### Is AI-generated code secure?

Not by default. In the most recent studies, from around a third to over half of generated code has security problems, especially around access and authentication. It needs to be reviewed before it is connected to real data.

### How much does it cost to build a custom tool?

The prototype may cost a few hours. The real cost is making it secure and maintaining it: estimate it over at least three years and compare it with the subscription it replaces.

### What is shadow IT and why is it a risk?

It is software created or adopted without the knowledge of the people who run IT. The risk is that data and processes come to depend on tools nobody controls or knows how to fix.

### Where is the best place to start?

With a small, repetitive task that is easy to check, tested first on sample data.

## Sources

- [Luciano Ballerano, "E se il software ce lo facessimo da soli?" (What if we built the software ourselves?), Umani per ora](https://lucianoballerano.substack.com/p/e-se-il-software-ce-lo-facessimo)
- [Retool, The Build vs. Buy Shift 2026](https://retool.com/blog/ai-build-vs-buy-report-2026)
- [UC Today, Will AI Agents Break the Project Management Software Model?](https://uctoday.com/will-ai-agents-break-the-project-management-software-model)
- [Long Angle, Software vs. AI Q1 2026](https://www.longangle.com/alts-education/software-ai)
- [IOActive, The Security Gap in AI-Generated Code](https://www.ioactive.com/the-security-gap-in-ai-generated-code/)
- [Cloud Security Alliance, Vibe Coding Security Debt](https://labs.cloudsecurityalliance.org/research/csa-research-note-ai-codegen-vulnerability-debt-20260406-csa/)
- [Open, AI-built alternatives to Adobe software (in Italian)](https://open.online/2026/10/12/crafting-apps-adobe-photoshop-premiere-gratis/)
- [Salesforce Ben, Klarna and Salesforce/Workday](https://www.salesforceben.com/klarna-salesforce-workday-partnership-called-off-amidst-major-gen-ai-overhaul/)
- [Entrepreneur, Klarna goes back to hiring human agents](https://www.entrepreneur.com/business-news/klarna-ceo-reverses-course-by-hiring-more-humans-not-ai/491396)
- [Gartner, prediction on rehiring by 2027](https://www.gartner.com/en/newsroom/press-releases/2026-02-03-gartner-predicts-half-of-companies-that-cut-customer-service-staff-due-to-ai-will-rehire-by-2027)

---

Cloudfactory S.r.l., Rome. Contact: hello@cloudfactory.it