Skip to main content

OneValley

Build vs. Buy in AI: Comparing custom tools and AI-native platforms.

TL;DR: Buy versus build isn’t about how important a tool feels or how fast AI can write it. It comes down to what the software has to reflect: your own operations, a sales motion most companies share, or an external standard set by a regulator. Where that answer lands tells you where to spend your engineering time and where to spend your budget.

For the past year, founders have been telling each other some version of the same story: AI made building software so cheap that buying it barely makes sense anymore. A team ships a support bot in a weekend. Someone cancels a dashboard subscription after a few hours in Cursor. But how true is that story? 

While Wall Street’s concern around the SaaSpocalypse has started to fade, the data suggests there is a real behavioral change happening, at least at startups.. Retool’s 2026 Build vs. Buy survey of 817 builders found that 35% of organizations had already replaced at least one major SaaS tool with something they built themselves, and 78% expect to build more this year. Harmonic, a startup discovery platform, hit a breaking point with a $20,000-a-year vendor whose support was so slow that rebuilding the tool in-house was faster than waiting for a ticket response. ClickUp evaluated a wave of AI vendors for its go-to-market operations, found none of them fit, and built six internal tools instead, cutting $200,000 a year in software spend.

But look closer and you’ll find that the build vs buy calculation looks very different depending on which kind of software you’re talking about. Founders aren’t ditching their CRM to build one from scratch. They’re leaving Salesforce for AI native platforms instead, like Attio, a three-year-old company serving 5,000 paying customers and on pace to quadruple its revenue this year, and Clarify, which offers autonomous, zero-admin execution. Not building, just buying from someone new. 

At the same time, no amount of AI progress has shortened a SOC 2 audit or made a home-grown authentication system pass certification at the rate a purpose-built one does. In some categories, AI hasn’t actually changed anything at all. 

We’re going to walk through these three patterns and see where building now actually wins, where the vendor landscape has flipped under everyone’s feet, and where AI progress is still yet to make its mark. 

It’s also important to note here that we’re talking about mostly operational infrastructure, not the product itself. Features that make your product stand out should always be built in-house, especially if it’s one of the key reasons a customer picks you over a competitor. 

Internal Tools and Workflow Automation: Where Building Actually Wins

Start with what a piece of internal software has to reflect. A CRM has to reflect a sales motion. A compliance system has to reflect an external regulation. An internal tool has to reflect your own operations: your team’s actual process, your specific data model, the particular way your company happens to work. But, most software is built to serve a vendor’s median customer, making it an awkward fit when you need something actually custom tailored. This is the space where AI has made the biggest impact. 

Retool’s 2026 survey of 817 builders, spanning startups to Fortune 500 companies, put a number on how far this has gone. Thirty-five percent of organizations had already replaced at least one major SaaS tool with a custom build in the past year, and 78% expect to build more in 2026. 

Two named examples from the report show what that looks like in practice. Harmonic, a startup discovery platform, was paying $20,000 a year for a third-party tool with support so slow that its automation lead found it faster to rebuild the product inside Retool than wait for a response. That rebuild became a habit. Harmonic now runs 33 internal apps connected to Salesforce, Gong, Slack, and its own internal APIs. 

ClickUp went through a similar arc from the other direction: its GTM team evaluated a wave of AI vendors, found none had the integrations or staying power it needed, and built six internal tools connected to Salesforce, Zendesk, and Snowflake instead. That decision cut $200,000 a year in automation software spend on top of the hundreds of engineering hours it saved weekly.

The same survey found that internal tools, automated workflows, and custom dashboards are what builders reach for first, and for a specific reason: these are exactly the categories that always varied the most from company to company. A workflow automation tool sold to thousands of customers has to optimize for the average case, and the average case matches nobody’s actual process.

However, building fast with AI carries some risks. The same research found that 60% of these builds happened outside official IT oversight, and internal teams have to build in the auditability and access control that a mature vendor would have included by default. Speed to build doesn’t remove the need for someone to own maintenance once the tool is running.

For most early-stage startups, if you’re maintaining a workaround, a spreadsheet, or heavy customization on top of a generic tool just to make it fit how your team actually works, that’s usually a signal you’re already past the point where building makes sense.

The flip: if nobody on your team can commit to owning the tool after it ships, or if it needs to touch sensitive customer data without a security review process behind it, the fast build isn’t actually cheap. It’s a liability with a due date. Be realistic about building what you’re actually able to maintain.

 

CRM: The Vendor Landscape Flipped, Buying Didn’t Stop

A CRM has to reflect a sales motion, and most sales motions share enough structure that a vendor can serve them well. That’s precisely why this category didn’t move toward building. It moved toward a new generation of vendors building the same idea on a different foundation.

Traditional CRMs force customer relationships into fixed categories: a contact, a company, a deal. That structure works for a conventional sales funnel, but it breaks down for relationships that don’t fit it well: a venture fund tracking a portfolio, a product-led company handing self-serve leads into an enterprise sales process, a partnerships team managing relationships that are neither pure sales nor pure support.

Attio built its CRM around a different premise. Rather than adding AI features onto a data model designed years before generative AI existed, it treated AI-native architecture as the foundation and let customers define their own objects. That bet has moved quickly. According to the company’s own funding announcement, Attio raised a $52 million Series B led by GV in August 2025, bringing total funding to $116 million just two years after launch. The same announcement lists AI companies including Modal and Replicate as customers and says the company is on pace to quadruple its annual recurring revenue this year.

None of that required a single customer to write code. They switched vendors. Salesforce’s Einstein and HubSpot’s Breeze are AI features, but they run inside data structures that predate the technology by years. There’s now a growing distinction between companies where AI shaped the foundation and where it’s just been added on top of existing foundations. 

If you’re setting up a CRM for the first time, consider starting with an AI-native option instead of defaulting to whatever your last company used. There’s no reason to inherit constraints from an architecture that predates the tools you’ll actually be using.

But, if you already have meaningful pipeline history in a legacy CRM and a technical hire who can own a migration, comparing your setup against a newer alternative is worth an afternoon. Otherwise, the switching cost usually isn’t worth your time yet.

Compliance and Authentication: Where Nothing Changed

Authentication and compliance have to reflect something neither your team nor any vendor gets to define: an external standard set by a regulator or an industry body. That’s the reason this category is where the “just build it” instinct causes the most damage. The bottleneck here was never how fast anyone could write code.

Take passwordless authentication. The FIDO Alliance, the industry body behind the passkey standard, published a Passkey Index in 2025 drawing on data from companies including Amazon, Google, Microsoft, PayPal, and Target. Passkey sign-ins succeed 93% of the time, compared to 63% for other authentication methods. Sign-in takes 73% less time, and login-related help desk incidents dropped 81%. Those numbers describe what happens when authentication is implemented correctly against a mature standard. Getting there depends on correctly following a specification that took years to develop and requires ongoing interoperability certification, work that has nothing to do with how fast an engineering team can ship.

Compliance tells the same story with harder numbers attached. SOC 2 requires an independent CPA firm to verify your security controls actually operate as described, not just that your code looks secure. Companies are generally looking at a first-year cost between $25,000 and $60,000, and a timeline running six months to over a year. Most of that timeline comes from an observation period during which auditors need to watch your controls operate over time. A faster engineering team doesn’t shorten a window that exists specifically because a single snapshot isn’t sufficient evidence.

Customer support makes an unexpected control case here, because most founders assume AI has already transformed it completely. Gartner‘s research projects that by 2026, conversational AI will reduce contact center labor costs by $80 billion. The detail that gets lost in that headline is that the same research projects only one in ten customer service interactions will be fully automated by 2026, up from roughly 1.6% today. Most of the savings come from making human agents faster, not from removing them. If a category associated with AI disruption still routes nine of every ten interactions to a person, a category like compliance, where the constraint is regulatory rather than technical, was never going to move faster.

When it comes to compliance, it’s clear that the decision is buy, without exception, regardless of team size or how comfortable your team is with AI coding tools. 

One Audit, Not a Checklist

If you line up all three categories, you’ll see the rule: buy vs. build isn’t about how important the software feels or how fast AI can write it. It’s about what the software has to reflect. 

Internal tools reflect your own operations, so generic SaaS was always a bad fit, and now building one is cheap enough to close that gap. CRM reflects a sales motion, and most sales motions are similar enough that a vendor can serve them well; what changed was who builds the best version, not whether you need one. Compliance and authentication reflect an external standard set by a regulator or an industry body. No amount of specificity or coding speed changes what that standard requires or how long an auditor needs to verify it.

As you think about your own operations, resist optimizing for what feels cheapest this quarter. A weekend build that saves a subscription fee isn’t a win if it becomes an unmaintained liability in a year, and a vendor switch isn’t worth the effort if it doesn’t change what your team can actually do. Ask what each tool has to reflect, then put your engineering time where it compounds: building where your operations are genuinely yours to own, buying where a vendor already solved the hard part, and treating every choice as one you’ll have to live with past this sprint.

Discover More blogs for Growing Innovation Ecosystems

Building the future starts with the right connections!