Build vs. Buy AI for Private Capital Firms: The Decision That Costs More Than You Think
Last updated 5 June 2026
Internal AI builds at private capital firms fail roughly twice as often as purchased solutions. MIT's Project NANDA report, The GenAI Divide: State of AI in Business 2025, analysed more than 300 AI deployments and found internal builds succeed in approximately one-third of cases. Bought solutions from specialist vendors succeed in two-thirds.
If someone is recommending you build your own AI tooling rather than buy a specialist platform, this article gives you the evidence to pressure-test that recommendation.
Key takeaways
- Internal AI builds succeed approximately 33% of the time, versus 67% for bought solutions from specialist vendors (MIT Project NANDA, 2025).
- Custom builds typically take 6 to 18 months before delivering value. SaaS platforms deliver in days to weeks.
- Annual software maintenance costs 15 to 20% of the original build, every year, usually falling to whoever remains when the consultants exit.
- Under DORA, the EU AI Act, and AIFMD II, buying from a compliant vendor shifts the bulk of the provider-side compliance obligations off your firm.
Why internal AI builds fail at twice the rate of ordinary IT projects
The RAND Corporation's "The Root Causes of Failure for Artificial Intelligence Projects" (2024), based on interviews with 65 data scientists and engineers across organisations, found that more than 80% of AI projects fail. That is roughly double the failure rate of non-AI IT projects. MIT's NANDA report adds the scale: despite $30 to 40 billion in enterprise investment in generative AI, 95% of organisations are generating zero measurable return on that investment. Full report coverage at https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/.
The failure reasons are consistent across both studies: poor data readiness, inadequate domain fit, and the inability to sustain the engineering capacity needed to maintain and evolve the tool after the initial build. That last failure mode is the most relevant to a 20 to 150-person investment firm with no dedicated AI team. The consultants who built the tool leave. The roadmap stops. The firm is left maintaining code it did not write.
MIT's conclusion, based on 52 structured interviews and 153 survey responses: "Build less, buy smart."
The real cost of building AI for an investment firm
The upfront cost estimate for an internal build is almost always wrong in the same direction. McKinsey and the University of Oxford found that large IT projects run 45% over budget on average and deliver 56% less value than predicted. That is before accounting for the hidden cost tail.
Machine learning engineers in Western Europe command salaries of roughly €70,000 to €100,000 annually (DigitalDefynd, 2026 — https://digitaldefynd.com/IQ/ai-salaries-in-europe/). Specialist AI consultant day rates from established firms reach €140 to €190 per hour, with large consulting houses charging $100 to $300 per hour and enterprise-tier firms reaching $400 to $900 per hour. A credible internal AI build requires hiring or contracting a category of talent that competes directly with Big Tech salaries, in a market where that talent is scarce.
Then there is the ongoing cost. Annual software maintenance runs 15 to 20% of the original build cost, according to IEEE research showing roughly 60% of total software lifetime cost occurs after launch. A €500,000 build does not cost €500,000. It costs €500,000 plus €75,000 to €100,000 every year after that, before any new use case or improvement is added. Over five years, the total is rarely competitive with a specialist platform subscription.
The maintenance trap: what happens after the consultants leave
The table below captures the build vs. buy decision across the dimensions that matter most to a regulated investment firm.
| Factor | Build in-house | Buy a purpose-built platform |
|---|---|---|
| Time to first value | 6 to 18 months | Days to weeks |
| Annual maintenance cost | 15–20% of build cost, indefinitely | Included in subscription |
| Domain expertise | Only if you hire for it specifically | Core product requirement |
| Model updates as AI advances | Full re-architecture at your cost | Vendor responsibility |
| DORA-ready contracts and audit rights | Your firm is the provider | Available from day one |
| EU data residency | Depends on your infrastructure choices | Contractually enforceable |
| Output traceability and audit trails | Must be engineered from scratch | Built-in governance feature |
| Risk if key builder leaves | Total knowledge loss | No single-person dependency |
Consultant-built tools are among the most brittle in any organisation because institutional knowledge sits in people, not documentation. When the project team moves on, the firm is left maintaining code it did not write, without the context to evolve it safely. This is the maintenance trap: an internal tool that was cutting-edge at launch, static six months later, and a liability by year two.
The pace of AI change makes custom builds obsolete faster than you expect
AI capability is not advancing at a rate that favours locking into a custom-built solution. METR's research, published on arXiv in March 2025 (https://arxiv.org/abs/2503.14499), found that the length of tasks AI agents can complete autonomously with 50% reliability has been doubling approximately every seven months for the last six years. Frontier models from Anthropic, OpenAI, and Google now ship major capability updates every few months, often at dramatically lower cost per task.
Bloomberg built a genuine 50-billion-parameter proprietary financial language model, BloombergGPT, using 1.3 million GPU hours across 512 NVIDIA A100s (https://www.hpcwire.com/2023/04/06/bloomberg-uses-1-3-million-hours-of-gpu-time-for-homegrown-large-language-model/). Within a year, general-purpose GPT-4 outperformed it on core financial tasks. BloombergGPT is now largely static, because retraining at that scale is not a recurring option for most organisations. A bespoke model built at enormous cost was made obsolete by cheaper, faster-moving off-the-shelf alternatives.
Even the firms with the deepest AI resources reached the same conclusion. JPMorgan, with more than 2,000 AI and machine learning staff and approximately $2 billion in annual AI spend, built a thin integration layer around OpenAI, Anthropic, and Google models rather than a proprietary model. Citadel's CEO Ken Griffin told Bloomberg in March 2023 that the firm was negotiating an enterprise-wide ChatGPT licence (https://www.bloomberg.com/news/articles/2023-03-07/griffin-says-trying-to-negotiate-enterprise-wide-chatgpt-license) rather than building. If organisations with effectively unlimited AI budgets concluded that building the underlying intelligence is uneconomic, the calculation for a 50-person fund is clear.
EU compliance shifts the calculation further toward buying
DORA has been enforceable since 17 January 2025. Under Regulation (EU) 2022/2554, it applies to approximately 22,000 EU financial entities, including investment firms and AIFMs. Article 28 requires each firm to manage ICT third-party risk as an integrated part of its risk framework: pre-contractual due diligence, a register of all ICT arrangements, and contractual terms covering audit rights, SLAs, sub-contractor transparency, and exit strategies.
If you build your own AI tool, your firm becomes the ICT provider. You must impose DORA's resilience, incident-reporting, and penetration-testing obligations on yourself. A specialist vendor arrives with DORA-ready contracts and resilience documentation already prepared, ready to sign.
The EU AI Act adds further weight to this calculation. High-risk obligations apply from 2 August 2026, covering risk management systems, data governance, technical documentation, human-oversight controls, conformity assessment, and post-market monitoring. Many investment-firm AI uses can fall into the high-risk category, with credit and eligibility assessment being the clearest example. When you build your own high-risk AI system, your firm becomes the "provider" under the Act and carries the full provider obligation stack. Penalties reach up to €35 million or 7% of global turnover for the most serious breaches. AIFMD II, with national transposition required by 16 April 2026, tightens delegation and outsourcing oversight further and requires clear documentation of any ICT arrangements that touch fund management functions.
None of these obligations disappear when you build internally. They move from the vendor column to yours.
What "buy" should actually mean: domain expertise, not just a chatbot
Not all platforms are equal, and buying a generic AI subscription is not the same as buying a purpose-built platform. Research from Stanford's RegLab, published in the Journal of Legal Analysis in 2024 (https://academic.oup.com/jla/article/16/1/64/7699227), found that general-purpose LLMs hallucinate between 58% and 88% of the time on specific, verifiable factual questions. For investment workflows where an error in a lease validation, a payment-flow check, or a fund report can be material, that error rate is not workable.
A purpose-built platform differs from a generic AI subscription in three concrete ways. It has domain-validated logic baked into the workflow, not applied as an afterthought by a prompt engineer. It provides source-linked, traceable outputs that meet the audit-trail expectations of DORA and the EU AI Act. And it handles the specific document types your firm processes: lease agreements, fund reports, payment instructions, insurance certificates, data room materials, not a generalised approximation of them.
Generic AI tools are right for generalist tasks: drafting, translation, summarising public information, meeting notes. For regulated, document-intensive investment workflows, they are the wrong tool, whether you access them through a subscription or wrap them yourself in an internal build.
How Xlagent combines domain expertise with a tailored approach
Xlagent was built specifically for private capital document workflows. Lease validation, payment-flow checks, insurance certificate review, due diligence document processing: each of these involves structured rules, cross-document logic, and exception handling that generic AI tools consistently fail at. That domain expertise is built into the platform from the ground up, not added by a consultant working from a general-purpose model.
The platform is also tailored to each client's specific processes. Most investment firms do not run identical workflows. The way one firm handles its lease review cycle differs from another's in ways that matter for accuracy and accountability. Xlagent's implementation approach takes your actual process as the starting point, not a generic template. That tailoring happens within a scalable architecture, so workflows designed for one use case extend to the next without rebuilding from scratch.
Critically, the platform keeps evolving. When Anthropic ships a new Claude model, when a new regulatory requirement changes what an audit trail must include, when a client's workflow needs to expand to cover a new asset class, that engineering work falls on Xlagent. Your firm captures the improvement without carrying the cost of re-architecting every 12 to 18 months or explaining to a regulator why your AI infrastructure has not kept pace with the standard.
The ongoing cost of a specialist platform is a fraction of what firms spend when they attempt to build equivalent capability with external consultants at €140 to €190 per hour, then maintain it without the original team. And unlike a consultant-built tool, the platform does not stop improving when the project closes.
The build vs. buy decision for EU private capital firms comes down to a practical question: is building this a core part of how your firm wins, or is it the infrastructure that lets your team do the work that actually wins? For document-heavy, regulated investment workflows, the answer shapes everything that follows. You can find out more about how Xlagent works with private capital firms at https://xlagent.ai.