Abner Ballardo | Founder, Aule Tech Foundry

Former CIO, Scotiabank Peru | Technology, business & architecture | Educator

What AI Changes in the Build or Buy Decision

AI can make software cheaper to build. Before you decide, estimate the cost of maintaining it and name the work your team would give up.
Two pale limestone blocks joined by a brushed bronze connector against a charcoal background.

Imagine a company in Peru is about to renew its payroll software contract. In the budget review, someone asks a fair question: now that coding assistants can write so much of the code, why not build it ourselves?

The work looks well defined. For each employee, the software has to take salary, leave, and approved bonuses from the company's own systems, apply local tax and contribution rules, and calculate take-home pay.

If AI makes a first version much cheaper, should the company build its own payroll software?

I've argued before that companies should buy commodities and build what differentiates them. In a later article, I used Geoffrey Moore's distinction between core and context and put payroll on the context side: essential to running the business, but owning the software gives the company little advantage.

As a CTO and CIO, I had to weigh what my teams could build against what they could afford to own. I suggested revisiting those decisions every 12 to 18 months, and I'm taking my own advice now that I'm building AI-native startups. With coding assistants, some of the work I would have bought or put off is now worth considering as an internal build.

I still lean toward buying commodities like payroll. AI can make smaller integrations around purchased software worth building. It can also give teams room to build more where their own software gives the business an advantage.

Who owns the next payroll update?

Suppose the company had asked that question a year ago, said yes, and built it. Through Supreme Decree 015-2026-TR, Peru raised the private-sector minimum wage to S/1,230 on 1 October 2026. Before October's pay run, that software would have needed an update for employees paid the minimum wage.

A second step is still to come. At the signing, the president said it would follow once El Niño has passed. The decree itself leaves it to a second decree, to be issued in the first half of 2027.

Whether the company buys or builds, its payroll team has to check the first pay run affected by each step. The change may be as simple as updating an effective-dated setting.

If the company owns the software, it also owns implementing the change and validating the result, with engineering support when code changes are needed. The second decree is due in the first half of 2027, but the company doesn't yet know the date or how much time it will have to act.

What AI makes cheaper, and what it doesn't

A coding assistant can help write the calculation, connect to internal systems, and generate tests for cases involving leave and bonuses. It can help write the update when a rule changes.

The team still has to interpret the rule and know what the right result looks like. It also has to protect employee data and keep enough history to explain a calculation months later.

I'd expect AI to cut the cost of getting to the first pay run more than the cost of keeping payroll correct month after month.

The vendor gets AI too. Its engineers can use the same tools to absorb rule changes faster, and the vendor spreads that work across all its customers. If AI lowers the company's cost of building payroll, it likely lowers the vendor's cost of maintaining it as well.

A company can benefit from AI through purchased software, too. I'd look for that benefit in the vendor's actual price, service, or product improvements. Buying still leaves the company responsible for its integrations and for checking that payroll is correct.

Payday doesn't wait for the roadmap

The cost that rarely shows up in the estimate is what owning payroll takes away. When a payroll change requires engineering, it joins the same roadmap as product commitments. If it must be ready by the next payday, it can push back a customer-facing release the same engineers were about to ship.

The company doesn't set that deadline and can't move it. When payroll is wrong, employees get paid the wrong amount and the company may face penalties.

That's why I still lean toward buying commodities. If AI frees engineering capacity, payroll is rarely where I'd spend it.

The same engineers could use that capacity to build what customers actually see. With AI helping with implementation, a team may have room to try more versions of a customer experience before choosing one.

For a bank, I'd apply that distinction to buying the core system and building mobile and web apps. At one bank, I led technology for the first version of a mobile payments app the bank built itself and delivered in 16 weeks. That's where a bank can design its own experience and create value.

That preference has limits. I'd consider building if payroll were part of what the company sells, or if its pay rules were unusual enough that no vendor could handle them without heavy customization. In both cases the software stops being a commodity for that company.

Where a smaller build could earn its place

Buying payroll doesn't mean building nothing around it. Suppose the company renews. The vendor handles the pay calculation and local rules well, but every month someone has to copy approved bonuses from an internal system and check them against the payroll import. I'd first test whether the vendor can take those records through an existing import or configuration.

If the gap remains, a small integration could prepare the approved records and flag entries that need review. Someone would still check those entries before payroll closes. Many integrations like this never made the cut because they cost too much to build for the time they saved.

With AI helping to write and test the code, I'd compare the new estimate with what the manual work costs each month. Some of those builds are now worth doing.

I've made that call on a small scale myself. To track my own publications, I used to gather the figures for each article and LinkedIn post by hand. I didn't want to buy a professional content analytics tool: a complete product isn't a priority for me right now, though it may be later. And before coding assistants, I wouldn't have written a tool for it either.

In October, working with a coding assistant, I built a small command-line tool that brings those figures into one catalog from my blog and my LinkedIn exports. It does just enough for the work I do today.

Using AI inside the monthly payroll process is a different decision. If some of the company's inputs still arrive as documents, such as bonus letters, a model could extract the amounts for someone to verify before importing them. Then the model becomes part of running payroll.

A coding assistant contributes to the cost of writing and maintaining the integration. The integration itself can run on approved data without calling a model each month.

A model that reads documents every pay run adds recurring costs: the model calls, the time spent checking what it extracted, and a plan for documents it can't read. I'd price those separately.

Compare the costs over the same period

If someone in your company is asking whether to build what you used to buy, I'd start by putting the vendor quote and the internal estimate side by side over the same period.

For each option, separate the cost of getting to the first working version from the cost of keeping it correct afterward. Ask who has priced the second.

If the case for building still holds, name the team that will own the updates and reserve time on its roadmap for deadlines you don't control. Then write down what that team won't deliver because of it.

AI changes the economics of building. It doesn't remove the need to ask where owning software gives the company an advantage.

If you decide to own this software, what other work are you prepared to give up to keep it working?

Subscribe

No spam, no sharing to third party. Only you and me.

Member discussion