You Can Build It. Can You Be the Vendor?

Is it really happening? Yes — at one specific layer

This is not vendor-fear content. The trend is real and measurable.

Gartner estimates that by 2028, roughly 40 percent of all new enterprise software will be assembled using vibe coding techniques. In July 2026, Gartner put up to $234 billion of enterprise SaaS spend — about a fifth of the total — at risk from agentic arbitrage by 2030.

And the replacements are shipping. Companies are building customer support workflows and legal tooling internally with AI agents, and replacing third-party analytics tools with internal builds. SaaStr's founder has shipped more than ten production applications since 2025; his team's internal tool now runs marketing operations outright rather than assisting them.

Notice what all of these have in common. Survey tools. Pricing calculators. Analytics dashboards. Marketing ops. Reporting layers.

These are workflow surfaces — thin applications over data that lives somewhere else. That is exactly where the economics now favour building, and any CFO still paying per-seat for a glorified form should be annoyed about it.

Is it that simple? No — and the case everyone cites proves the opposite

Klarna is the exhibit.

In 2024 its CEO told investors the company had shut down Salesforce and would shut down Workday within weeks, as part of internal initiatives combining AI, standardisation and simplification. The headline wrote itself: AI killed SaaS.

Then read what actually happened. Klarna confirmed it stopped using Workday and Salesforce's CRM — and replaced them with Deel for HR, plus a set of other SaaS tools for CRM functionality. It kept Slack. The CEO's own description was blunter than the coverage: they did not replace SaaS with an LLM; they built an internal stack on a graph database to bring data together, and the consequence was the liquidation of a lot of SaaS — not all of it.

That is a consolidation story, executed by a company that was running roughly 1,200 SaaS applications, with a large engineering organisation and a specific data-unification thesis behind it.

It is not a build-your-own-ERP story. It has been quoted as one for two years, and I have sat in meetings where it was the entire business case.

One more data point, which I find more persuasive than any analyst forecast: Anthropic — arguably the company on earth best positioned to build its own CRM with AI — is hiring Salesforce developers.

The reframe: you are not buying software, you are buying a vendor

Here is the line item that never makes it into the business case.

When you cancel a subscription and build the replacement, you inherit the vendor's entire job. The roadmap. Security patching. Dependency upgrades. Uptime. Support. Compliance evidence at renewal and at audit. Onboarding documentation. Training for every new hire. And the one that actually kills programmes: continuity when the person who built it leaves.

That work does not appear on a budget line. It appears as a slow tax on your best engineers, and it compounds.

The outcome data is consistent with the maintenance math above: it most often fails for unclear ROI, data quality problems, and underestimated maintenance cost.

Meanwhile, SaaS spend has not fallen. The mix is changing — heavyweight systems stay SaaS while lighter tools get absorbed — but the aggregate is rising.

Both things are true at once. That is precisely what makes this a sequencing decision rather than an ideology.

Four questions before you cancel anything

I use these in the first hour of a diagnostic. They are ordered deliberately — a no on any one stops the exercise.

• Is this a system of record, or a surface over one? If losing this database means losing history you cannot reconstruct, you are not building an app. You are becoming a custodian.
• Who is on call in eighteen months? Name the person. If the answer is the person who built it, and they are a revenue operations manager, you have a single point of failure with no backup and no handover document.
• What happens when your largest customer's security team asks for evidence? A vendor hands over a report. You will hand over your own architecture, and it will be reviewed by someone whose job is to find the gap.
• What is the cost of being wrong? Not the build cost. The cost of reconstructing the data, re-licensing at list price without your negotiated discount, and running both systems in parallel through a migration you did not budget for.

Where the line sits today

The line moves. It has moved considerably in two years and it will keep moving. It has not moved far enough to put your general ledger on the wrong side of it.

• Character — Build it: workflow surface, reporting layer, internal tool, integration glue, calculator. Buy it: system of record — CRM, ERP, HR, finance, billing.
• Data — Build it: derived, reconstructible. Buy it: authoritative, irreplaceable.
• Regulatory load — Build it: low. Buy it: audit trail, retention, segregation of duties.
• Users — Build it: one team. Buy it: cross-functional, with a permissions model.
• Real cost driver — Build it: the build itself. Buy it: ownership after go-live.

What I would actually do

Stop framing this as build versus buy. Frame it as a portfolio decision with a baseline.

Inventory what you pay for, then score each subscription against the four questions above. You will typically find that 15–30% of line items are surfaces you are renting at per-seat prices. Cancel those, build the replacements, and the savings are real and fast.

You will also find that the two or three systems generating your largest invoices are systems of record — and those invoices are buying you something the build case never prices.

So build the things you can own. Just be honest that "own" is a verb with a bill attached, and that the subscription you are cancelling was never really for the software.

My co-founder, Shampa Bagchi, takes the architecture side of this question — where the technical boundary actually sits, and what the security research says about AI-generated code.