For twenty years the standard advice was easy: buy for commodity, build only for differentiation, and remember that building always costs three times the estimate. AI-assisted development has genuinely disturbed that equation, but not in the way the demos suggest.

What actually got cheaper

Be precise about it. AI tooling has collapsed the cost of the first 70 percent of a custom app: scaffolding, CRUD screens, integrations with well-documented APIs, test boilerplate, admin panels. Work that took a small team a quarter can now take one capable developer a few weeks.

What did not get cheaper:

  • Deciding what the software should do, and getting stakeholders to agree.
  • The last 30 percent: edge cases, performance under load, accessibility, the weird invoice format your biggest client insists on.
  • Ownership: hosting, monitoring, security patches, dependency upgrades, and someone answering when it breaks during a cyclone weekend.

So the honest summary for 2026 is: building is much cheaper, owning costs roughly what it always did. Any build vs buy analysis that compares AI-speed construction against a SaaS subscription, while ignoring the ownership line, will systematically choose wrong.

The case for buying got sharper too

SaaS vendors have the same AI tools you do, plus your competitors' feature requests. The gap between a mature product and your v1 is wider than it looks from a demo, because the vendor has already absorbed a decade of edge cases.

Buying still wins clearly when:

  • The domain is regulated or unforgiving: payroll, accounting, payments.
  • The product is the moat of a company whose whole business is that product.
  • Your need is genuinely standard and your team is small.

For a Mauritian SME, there is an extra wrinkle: many global SaaS tools handle local realities poorly, MRA compliance, MUR invoicing quirks, or pricing set in dollars that stings at local margins. "Buy" sometimes means "buy and tolerate a permanent mismatch", and that tolerance has a cost worth writing down.

The new middle path: buy the core, build the edges

The most interesting shift is that the middle option became viable. Because integration code is exactly the kind of code AI writes well, it is now realistic for a small team to buy a solid core system and build thin custom layers around it: a customer portal over the accounting package, an automation that turns WhatsApp orders into ERP entries, a reporting layer the vendor never shipped.

This pattern gets you vendor-grade reliability where it matters and custom fit where you are actually different. It also keeps your custom surface small enough that one developer, or an outside partner, can maintain it without heroics.

A framework that survives contact with reality

Work through these in order, and stop at the first decisive answer:

  1. Is this capability part of why customers choose you? If no, default to buy.
  2. Does a product exist that covers 80 percent of the need without contortions? If yes, buy it and script the gap.
  3. Can you name the person who will own the custom system in year three? Not build it, own it. If no name comes to mind, do not build.
  4. Is your requirement stable? Building onto shifting requirements multiplies cost; buying flexible tooling absorbs it better.
  5. Finally, and only now, compare costs over three years: subscription and workaround costs on one side, build plus hosting plus maintenance at roughly 20 percent of build cost per year on the other.

Most organisations that run this honestly end up with a portfolio: bought core systems, a handful of built edges, and one or two genuinely custom products where they compete.

The trap to avoid this year

The 2026-specific failure mode is seductive: a manager watches an AI tool generate a working prototype in an afternoon and concludes that software is now free. Prototypes are now nearly free. Products are not. The distance between them is testing, security, data migration, training, and years of unglamorous upkeep.

Use the cheap prototype for what it is genuinely good for: de-risking the decision. Build the throwaway version, put it in front of five users, and let what you learn inform a calmer choice. That might be the best return on AI tooling available to any business today: not free software, but cheap certainty.