By Pat Conroy, Head of Strategy, Product & Innovation, STP Investment Services.
An operations leader at a mid-market investment manager is sitting across from a vendor demo. The pitch is compelling, but in the back of their mind, one question keeps surfacing: should we just build this ourselves?
A year later, that same firm is twelve months into a build, three engineers deep, still not in production, and the vendor they passed on has shipped two new iterations. The scenario is not unusual. It is, increasingly, the default outcome for firms that apply a 2016 framework to a 2026 decision.
The data makes the stakes plain. According to S&P Global’s 2025 enterprise survey, 42% of firms abandoned most of their AI initiatives last year, up from 17% the year prior, with the average organization scrapping nearly half of all proofs of concept before reaching production. The average sunk cost per abandoned initiative was $7.2 million. This is happening inside firms that can absorb it.
Consider who actually makes up the investment management industry: according to the IAA Investment Adviser Industry Snapshot 2026, 92.8% of SEC-registered advisers employ 100 or fewer people, with a median firm size of just eight employees and median assets under management of $446.9 million. For most of the industry, a failed build cycle is not a setback. It is an existential event.
The buy vs. build question hasn’t gotten harder because the technology has changed. It has gotten harder because the commitment it demands has fundamentally shifted. AI has split the decision into two.
The firms getting hurt are not the ones who chose wrong. They are the ones who answered a two-tier question as if it only had one answer. There are several dimensions in which AI has changed decision-making: speed, cost, timing, and due diligence. Understanding what each means by tier is how firms can stop making that mistake.
Speed Is No Longer Time to Deployment, It’s Iteration Velocity
For enablement and efficiency solutions – document summarization, workflow automation, internal knowledge bases, departmental productivity tools – mid-market teams can and should build. Modern foundation models plus a small technical team can stand up internal tools in weeks. The iteration loop with users is tight, the blast radius is bounded, the organizational commitment is manageable, and the speed advantage of building shows up clearly.
For mission-critical solutions – anything that touches investment decisions, client-facing outputs, records subject to examination, or systems embedded in trading, custody, or accounting – the calculus inverts entirely. Buying embedded AI in those use cases means buying 18 to 24 months of testing, integration, and refinement that the vendor has already absorbed. A team starting today restarts that clock with no peer reference points and limited internal capacity to scale legal, compliance, and risk review on demand.
The deeper shift is in how speed itself is measured. It is no longer time to deployment. It is iteration velocity. A vendor running across 50 firms iterates exponentially more than a team building inside one. Pilot purgatory, which S&P data suggests is the norm, is a failure to ever cross the production line.
The Hidden Cost Column: What “Build” Really Costs Across Cycles
The upfront cost of buying is visible on a spreadsheet. The cost of building is visible only in hindsight.
For enablement and efficiency solutions, the total cost of ownership equation is manageable. Bounded scope, limited integration, a small technical team, predictable maintenance – mid-market firms can absorb this cost and often should.
For mission-critical solutions, the build column on every legacy decision matrix systematically underprices three costs.
- The first is the permanent department. Building mission-critical AI is not a project. It requires sustained AI engineering, model governance, data engineering, and platform operations roles that do not disband after launch. These roles compete directly with frontier model labs and Big Tech for the same talent, at base salaries that now routinely clear $200,000 for senior specialists, before equity or retention.
- The second is the maintenance tax. AI systems are not static. Foundation models evolve, training data drifts, integration surface area expands, security postures shift, and regulatory cycles add their own maintenance events. Vendors amortize this cost across their customer base. Builders absorb it individually, every cycle.
- The third is the Headcount Trap. When the build slips, firms hire to close the gap. But those engineers spend most of their time on platform maintenance rather than on differentiation. The firm bought the bench. It did not buy the leverage.
The Window Question: Is It Too Late to Build?
For enablement and efficiency solutions, the window remains open. Modern foundation models, mature tooling, and shrinking talent requirements mean a small team can stand up internal tools at almost any point in the cycle.
For mission-critical solutions in investment operations, the window is effectively closing, and not because the technology moved. It is closing because the cost of falling behind on organizational AI capability is now higher than the cost of sourcing that capability from a partner that has already crossed the curve.
The threshold is an organizational learning curve, not a date. By 2027, firms that moved decisively in 2026 will have completed multiple model iterations, identified which applications actually drive outcomes, and built the internal muscle to sustain the platform. Firms still evaluating in late 2026 are not 18 months behind on technology. They are 18 months behind on organizational capability, a harder gap to close.
The build case holds for a narrow cohort: firms with genuine proprietary investment logic, a sustained AI engineering bench, the data architecture to support it, and the operational depth to defend the system across multiple cycles. This does not describe most firms.
What AI Diligence Actually Looks Like Now and What It Means for Your Competitive Position
Allocator due diligence on AI has expanded substantially over the last 12 months, and the velocity is accelerating faster than most firms have priced in.
Previously, it focused on vendor oversight and basic policy attestation. It now extends to proprietary AI tools, model governance, data lineage, human oversight protocols, transcription governance, and valuation controls. The ODD function is not just asking whether you use AI. It is asking how you govern it, what it touches, and what would happen if it failed.
I think of this as the Evidence Economy. Allocators are asking for the artifact, not the policy. AI governance is on the same arc cybersecurity and cloud posture walked before it: from optional differentiator, to expected baseline, to mandate gate.
There is also a commercial dimension that often gets overlooked. Allocators are asking about AI capability, not just AI risk. Firms with credible AI deployed in workflows that touch clients are differentiating themselves on service quality, response speed, and analytical depth. Firms without are increasingly invisible to their prospects.
The asymmetry hits mid-market firms hardest. The governance and capability questions a $500 billion manager faces are now the same as those a $5 billion manager faces, without the same budget. AI readiness has become a proxy for operational maturity. Firms that fail this section of the DDQ often do not make it to the investment thesis discussion.
What Are You Willing to Permanently Own?
The buy vs. build question has always been framed as a technology preference. AI has made it an organizational commitment question.
Before answering build or buy, every firm needs to get clear on one thing: what are you willing to permanently own? Ownership in 2026 spans at least six dimensions: talent and key person dependency, technical maintenance burden, data architecture, roadmap velocity, regulatory exposure, and cost structure across multiple cycles.
Most firms have been answering buy vs. build as if it were a technology preference. It is not. It is an organizational commitment across all six of those dimensions at once.
For most mid-market investment operations teams, honest accounting answers the mission-critical question before they ever open a vendor RFP or stand up a GitHub repo. The internal build is not the wrong answer. It is the wrong answer for the wrong tier.
For enablement tools, the math looks different, and the internal build case is real. A small team, a bounded problem, a tight feedback loop – that is a build worth making.
The firms that get this right in 2026 will not have been the fastest to decide. They will have been the most honest about what they were actually deciding.
Subscribe to our newsletter
