A senior ops manager at a 200-person company replaced her team's workflow automation SaaS on a Friday afternoon. Claude Code, a weekend, a few hundred dollars in API credits. By Monday, it was in production. The SaaS subscription - $14,000 a year - got cancelled that week.
That story is real, and it's happening at scale. Retool surveyed 817 builders across engineering, operations, and IT, and 35% of teams have already replaced at least one SaaS tool with a custom-built solution, and 78% plan to build more custom tools in 2026. The build side of the classic build-vs-buy ledger just got a lot cheaper. That part is true.
What isn't true is that the cost went away. It moved - mostly into a line item nobody owns yet.
Why "build" finally pencils out for internal tools
For most of the SaaS era, the build-vs-buy default was obvious: the old logic was that building custom software was expensive and slow, so you bought SaaS - Salesforce, Jira, ServiceNow - paid the subscription, ate the customization costs, and accepted the vendor's workflow because the alternative was worse.
Two things broke that logic at once. AI-assisted development cut custom build timelines by 30 to 50%, while SaaS pricing rose 11.4% on average in 2025. Those two curves crossing is the entire story.
Every SaaS category is now under replacement pressure. Workflow automations (35%) and internal admin tools (33%) top the list - areas where the gap between what a SaaS tool provides and what an organization actually needs is often widest. BI tools, CRMs, and project management software all show meaningful replacement rates too.
The pattern inside the shadow IT numbers is more telling than the headline figure. 31% of builders said they could build faster than IT could provision tools, 25% said existing SaaS tools couldn't do what they needed, and 18% said IT's process was too slow - and 64% of these shadow builders were senior managers and above, experienced operators choosing "build it myself" over waiting for procurement.
Where the cost actually went
Here is the part most "build vs buy just flipped" coverage skips.
Plans start at $20/month for Claude Code or Codex Plus, but agents bill by the token, and one task pushes 400K to 2M cumulative input tokens through the API, so heavy automation reaches $500 to $2,000 per engineer per month.
Anthropic's cost documentation reports that across enterprise deployments, average spend is around $13 per developer per active day and $150-$250 per developer per month, with 90% of users staying below $30 per active day. That's the median. The distribution has a long right tail.
Run the math at team scale: for a team of 100 engineers with a realistic usage mix - 70 at $200/month, 25 at $600, and 5 at $1,800 - total monthly API spend reaches roughly $38,000.
That's not an argument against agentic coding. A $38,000 monthly tool bill across 100 engineers is $456/year per engineer, which is almost certainly less than those engineers' time cost on the manual work being replaced. But it is a real number, and it's a number that currently has no owner in most organizations.
The governance cost compounds this. Wiz's 2026 State of AI in the Cloud report found that at least 80% of organizations have AI IDE extensions in their environments and 71% have at least one AI coding assistant, with much of that adoption happening from the bottom up rather than through centrally managed programs. And the risk extends to how software gets built: Wiz's research found that roughly one in five organizations using AI-powered vibe-coding platforms had applications affected by systemic security weaknesses.
We're entering the era of shadow operations: the uncontrolled deployment of autonomous agents that execute logic, integrate with systems by calling APIs, and modify states without formal security oversight. A shadow build that replaces a $14,000 SaaS subscription but ships without code review, a dependency audit, or any access-control documentation hasn't saved $14,000. It's deferred a larger bill.
The steelman for "buy" is still strong in three places
It's worth being honest about where the old calculus still holds.
Compliance-heavy categories. Payroll, HIPAA-adjacent data handling, SOC 2 audit trails - building custom here means owning the audit surface, not just the feature surface. Building requires dedicated ML engineering talent, ongoing model maintenance, data pipeline management, and continuous optimization. For regulated workflows, that cost is real and recurring.
Categories moving fast. When a category is evolving rapidly - new features shipping monthly, competitive dynamics shifting - vendors dedicated to that category will outpace what your team can build and maintain. A team building its own video conferencing tool in 2019 would have lost badly to Zoom by 2021. The same logic applies to any category with genuine innovation velocity.
Opportunity cost of engineering time. Every hour your engineering team spends on internal AI tooling is an hour not spent on your core product. For most companies, the opportunity cost is substantial. Agentic coding compresses build time but doesn't zero it out - and the maintenance tail is real.
What the 35% of companies in the Retool report have figured out is that they're not abandoning SaaS entirely - they're replacing the tools that were costing them the most in workarounds and limitations, and keeping the ones that genuinely serve them well.
What "build" actually requires to work
The teams getting this right share three practices that the weekend-build narrative glosses over.
A human signs off before it ships. Agentic coding tools have made the writing fast. Review, documentation, and incident ownership still require a decision. A teammate like Beagle can assemble context and surface a PR summary into Slack for a quick nod - but the nod still has to happen.
The build decision is scoped, not open-ended. Agentic coding harnesses cut the engineering hours needed to scaffold, integrate, test, and ship custom software - lowering the upfront barrier that historically pushed teams to rent almost everything below their core product. That's true. But "lower barrier" does not mean "no barrier." The right unit of comparison is a specific workflow, not "our entire SaaS stack."
Token spend has an owner. Agent teams use roughly 7x the tokens of a standard session when teammates run in plan mode, because each teammate keeps its own context window. Without someone watching that number, the savings from canceling the SaaS subscription quietly evaporate into the API bill.
Beagle in action#engineering, Thursday 3:40pmThe build-vs-buy decision didn't get simpler. It got faster to get wrong.
Agentic coding tools and build vs buy: common questions
Does agentic coding actually make internal tools cheaper to build?
Yes, materially. AI-assisted development cut custom build timelines by 30 to 50% , and 51% of builders have already shipped production software using AI, with about half of them reporting saving six or more hours per week. The build cost is genuinely lower - the governance and maintenance costs are not.
How much does Claude Code cost for an engineering team?
Across enterprise deployments, Anthropic reports average spend of around $13 per developer per active day and $150-$250 per developer per month, with 90% of users staying below $30 per active day.
At that median, a team of 50 spends $7,500-$12,500 per month; add 5-10 heavy users at $500-$2,000 each and the total can push past $20,000.
Which SaaS categories are most at risk from custom internal builds?
Every SaaS category is now under replacement pressure, with workflow automations (35%) and internal admin tools (33%) leading the list - areas where the gap between what a SaaS tool provides and what an organization actually needs is often widest. BI tools (29%), CRMs and form builders (25%), and project management (23%) follow.
What's the real risk of building internal tools outside IT oversight?
Shadow AI risks are higher than traditional shadow IT: unlike unauthorized SaaS applications, AI tools can ingest, retain, and potentially train on sensitive corporate data, creating irreversible exposure risks.
Wiz's research found that roughly one in five organizations using AI-powered vibe-coding platforms had applications affected by systemic security weaknesses.
When should a team still buy instead of build?
Buy when the category is compliance-heavy, when vendor innovation velocity outpaces what your team can maintain, or when the integration complexity is in a domain where your engineers have no existing expertise. If AI is what makes your product unique, building gives you control over the intellectual property; if AI is a supporting function like internal document search or customer support automation, buying makes more sense because you are not competing on that capability.