Build, buy, or partner: A decision framework for high-stakes technology initiatives

S&P Global Market Intelligence surveyed 1,006 IT and business leaders across North America and Europe and found that the share of companies abandoning most of their AI initiatives — specifically in the gap between a working proof of concept and production — rose from 17% at the end of 2023 to 42% at the end of 2024. On average, organizations killed roughly 46% of their AI projects somewhere in that gap.

Read that again. The proof of concept worked. What failed was the decision — the moment, months before launch, when someone chose a path without a framework and committed budget, headcount, and political capital to it.

Build, buy, or partner is the oldest question in enterprise technology. It’s also the one most leadership teams still get wrong, because they’re answering it with instincts calibrated for a world that no longer exists.

Why your instincts are lying to you

For fifteen years the answer to almost everything was “buy.” SaaS was cheap, fast to deploy, and someone else’s problem to maintain. Building meant six-figure budgets, multi-month timelines, and a maintenance tail you’d drag around for years. Three things broke that math.

1. AI collapsed the cost of building. Experienced engineers paired with agentic coding tools have compressed custom build timelines enough to make most pre-2024 estimates irrelevant. Retool’s 2026 survey found 93% of builders using LLMs to code or automate at work, and 35% who have already replaced at least one SaaS tool with something they built. Read that with the sample in mind — 817 of Retool’s own customers, roughly a third software engineers, which is close to a best case. But the direction is unambiguous: the entry cost of finding out whether building makes sense has collapsed.

2. Buying got more expensive and less safe. Per-seat costs are climbing, AI-feature surcharges arrive at every renewal, consumption pricing gets layered onto subscriptions, and lock-in deepens. Meanwhile platforms reprice mid-contract, roadmaps pivot away from your use case, and capabilities you depend on get deprecated.

3. Partnering matured into a genuine third path. Modern partnership isn’t throwing a spec over a wall. It’s scoped, time-boxed engagements — a proof of concept, a feasibility assessment, a roadmap — where you own the code and the decision at the end. BMW and Bosch have formalized this as “venture clienting.”

A question that used to have a default answer now has three legitimate ones. Which means you need a way to decide.

The CORE test: Four questions before you commit

Forget the 40-row scoring matrix. Four questions: Competitive differentiation, Ownership appetite, Rate of change, Execution capacity.

C — Does this create Competitive differentiation?

If a capability isn’t your core IP or something customers notice, buy it. Payroll, ERP, financial ledgers, HR systems of record — rational purchases, and AI hasn’t changed that. Nobody won a customer with a homegrown expense-report tool.

Be honest in the other direction too. If the technology is how you compete — your customer experience, your pricing engine, your data advantage — renting it means renting your differentiation, and your competitors can rent the same thing.

“This is strategic” is the most abused phrase in technology planning. If you can’t say how a capability wins you a customer or a margin point, it’s not differentiation. It’s plumbing.

O — What’s your Ownership appetite?

Because building is now possible for almost everyone, the temptation to build everything is stronger than ever. But building isn’t a project; it’s a commitment to security patching, compliance, monitoring, and evolution — forever, or until you sunset it.

Klarna is the case study everyone cites, and it’s worth getting right, because the circulating version is wrong. What happened: in February 2024 Klarna announced an OpenAI-powered assistant doing the work of ~700 support agents, cutting resolution time from 11 minutes to under two. By May 2025 CEO Sebastian Siemiatkowski said the company had pushed too far on cost, that quality had suffered, and Klarna began rehiring humans.

The version repeated in a lot of build-vs-buy content — that Klarna ripped out Salesforce, replaced it with internally built AI, and reversed when satisfaction dropped — mashes together two unrelated stories. Klarna did shut down Salesforce and Workday in 2024, but later clarified it consolidated onto alternative SaaS tools with AI layered over them. Siemiatkowski has said he was embarrassed by how that got reported.

The real lesson is sharper than the myth: owning something end-to-end means owning its failure modes. Klarna wasn’t burned by bad technology. It underestimated what it had taken responsibility for.

R — What’s the Rate of change?

• Stable domain, stable requirements → Buy. Mature vendors have absorbed twenty years of edge cases you haven’t thought of.
• Fast-moving domain (anything AI-adjacent) → Avoid long commitments in either direction. A three-year vendor contract in a market that reinvents itself every six months is as risky as a custom build on a framework that may be obsolete at launch.
• Unique, fast-changing requirements → Build or partner. A vendor roadmap will never keep pace with your specifics.

E — Do you have the Execution capacity?

The least glamorous question, and the one that quietly kills the most initiatives.

Prototypes are cheap now; production systems still demand architecture, testing, security review, and observability. The evidence on how much AI speeds that up is messier than either camp admits. METR’s randomized controlled trial (February–June 2025, 16 experienced developers, 246 real tasks in repositories they’d worked in for years) found developers took 19% longer with AI while believing they’d been sped up by 20%. That result is now widely misquoted as proof AI doesn’t help — but METR published an update in February 2026 finding evidence of speedup in a larger follow-up, and stated it believes developers are likely faster with early-2026 tools.

The honest read: gains are real and probably growing, but the gap between perceived and measured productivity is real too. Don’t budget a build on how fast the prototype felt. If your team is at capacity, “we’ll build it” means “we’ll build it instead of something else.”

This is where partnering earns its seat. The question a partnership answers is: are you solving a problem, or building an internal capability? If it’s the former, a scoped engagement with domain expertise compresses your timeline, bounds your risk, and leaves you owning the result.

The decision at a glance

Signal	Build	Buy	Partner
Core differentiator?	✅ Yes	❌ No	✅ Yes, but capacity-constrained
Ownership appetite	High — you want the asset	Low — you want the outcome	Medium — the asset without staffing the build
Rate of change	Requirements outpace vendors	Domain is mature and stable	Domain is complex; you need expertise fast
Execution capacity	Strong team with bandwidth	Thin team, generic need	Strong vision, stretched team
Best when	The tech is the business	The tech supports the business	The tech is strategic but the clock is running

The part most frameworks skip: It’s rarely one answer

Mature organizations don’t choose one path — they choose per initiative, and increasingly they blend: buy systems of record, build the workflow-specific tools where your process is your advantage, partner where stakes are high, the domain is specialized, and the timeline matters. UC Berkeley researchers have formally extended the classic triangle to include “blend,” because hybrid is now the norm.

Whatever mix you choose, protect your exit ramps. Modular, API-first architecture and explicit data-extraction terms aren’t paranoia — they’re increasingly what regulators and the market expect.

The broadest signal is the EU Data Act, applicable since September 2025. Unlike sector-specific rules, its switching provisions cover data processing services generally — IaaS, PaaS, and SaaS — giving customers a statutory right to switch, requiring portability terms in the contract, and banning switching charges entirely from January 2027. Watch what the market did rather than what it said: AWS, Microsoft, and Google all launched free-egress-on-exit programs ahead of those provisions. That’s not generosity. That’s pricing in a regulation everyone saw coming.

Sector rules go further where stakes are highest. EU financial entities under DORA must maintain tested exit strategies for ICT services supporting critical functions, including transition periods long enough to migrate elsewhere or bring the function in-house. US banking organizations work under 2023 Interagency Guidance treating termination as a formal lifecycle stage. In healthcare, HIPAA business associate agreements have long required defined terms for returning or destroying PHI.

Not all of these apply to you. The pattern should anyway: your data and workflows stay yours regardless of how a relationship evolves. The most expensive decisions aren’t the wrong ones — they’re the irreversible ones.

Three failure modes

1. Building for ego. “We’re an engineering-driven company” is a culture statement, not a business case.
2. Buying for comfort. A vendor logo feels safe until you’re paying for 60% fit and fighting the other 40% with workarounds and shadow spreadsheets.
3. Partnering without ownership. Open-ended retainers with fuzzy deliverables are how partnerships get a bad name. Insist on scoped work, tangible deliverables, and code you own.

The bottom line

The organizations winning with technology aren’t the ones that picked the “right” path. They picked deliberately — per initiative, with a framework, and with the humility to revisit as things change. Build, buy, and partner aren’t ideologies. They’re tools. The failure isn’t choosing wrong; it’s choosing on autopilot and finding out twelve months into a program you can’t unwind.

Before your next high-stakes initiative, run the CORE Test. Four honest questions, thirty minutes with your leadership team, and a decision you can defend — to your board, your engineers, and your future self.

---

Facing a build-buy-partner decision right now? Creed Interactive helps organizations pressure-test these choices every week — from scoped proofs of concept to full custom builds you own outright. Let’s talk before you commit the budget.

Sources

• AI project abandonment (17% → 42%): S&P Global Market Intelligence / 451 Research, Voice of the Enterprise: AI & Machine Learning, Use Cases 2025. Survey of 1,006 midlevel and senior IT and line-of-business professionals in North America and Europe, fielded October 21 – November 25, 2024. Figures compare Q4 2023 to Q4 2024. Overview · Related analysis
• 93% LLM usage / 35% SaaS replacement: Retool, The Build vs. Buy Shift: AI, Shadow IT, and the SaaS Replacement Era (2026), n=817 Retool customers and builders. Report
• Developer productivity RCT: Becker, Rush, Barnes & Rein, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, METR, July 10, 2025. Blog · Paper (arXiv 2507.09089)
• METR follow-up: Becker, Rush, Cunningham, Rein & Mahamud, We are Changing our Developer Productivity Experiment Design, METR, February 24, 2026. Update
• Klarna customer service reversal: Reported from the February 2024 announcement through Siemiatkowski’s May 2025 comments to Bloomberg. Fast Company summary
• Klarna SaaS clarification: Klarna’s correction that it moved to alternative SaaS tools rather than internally built AI replacements. CX Today · Siemiatkowski’s account
• EU Data Act cloud switching: Regulation (EU) 2023/2854, Chapter VI (Articles 23–31). Applicable September 12, 2025; switching charges prohibited from January 12, 2027.
• DORA exit strategies: Regulation (EU) 2022/2554, Articles 28(8) and 30(2)(f). Applicable January 17, 2025, to EU financial entities and their ICT providers. ESMA overview
• US third-party risk: Interagency Guidance on Third-Party Relationships: Risk Management, Federal Reserve / FDIC / OCC, final June 6, 2023. OCC announcement