The leverage isn't in the model layer

Most fund AI investment is going to the wrong layer.

Over the last two months we've sat in conversations with more than 15 investment teams across Europe and the US. Mid-cap PE specialists, climate-focused minority investors, special-situations debt-equity hybrids, fintech VCs at fundraising stage, family-office-backed long-hold funds. Different sizes, different mandates, different speeds of conviction. The same shape shows up almost everywhere.

Every fund has a Claude or ChatGPT subscription. Most have a champion building shadow IT on their personal API key. Some have piloted a horizontal platform (Hebbia, Model ML, Naviia, Rogo, Dust) and quietly let it lapse. A handful are deep into custom builds. The aggregate investment in fund AI is real. The compounding leverage is patchy.

The question we keep coming back to is not which model to use, or which platform to subscribe to. Almost nobody is asking what's actually generating their alpha and how AI should encode it. The leverage isn't in the model layer. It's one layer up.

What 15+ funds told us

Five patterns showed up across nearly every conversation, regardless of fund size, strategy, or geography. Quotes are anonymised by role, fund type, and region.

The platforms-got-worse-as-they-grew pattern is naming itself. Funds that adopted an AI platform early are starting to articulate the same complaint. The platform was responsive when the fund was an important customer. Then the platform's client mix shifted upmarket, feedback loops elongated, and the fund quietly stopped being represented on the roadmap. It is a predictable shape, not a story about one bad vendor.

"As they got more clients with bigger check sizes, we were deprioritised. That's a point we'd like to understand how it's mitigated in the long run."

Chief of Staff, fintech-focused VC (Europe)

Compliance is the silent bottleneck on adoption. Almost every fund we spoke to has a frustrated investment team and a cautious compliance function. The drives, CRMs, emails, and board reporting platforms that would let AI actually compound sit behind data-residency, PII, and connector restrictions that haven't been resolved. Skills, agents, and Cowork-style integrations are blocked at most funds for exactly this reason. The fix is procurement and architecture work, not better models. Funds that have solved it move noticeably faster.

"The investment team is dying to upgrade. Compliance is the only blocker."

Senior Associate, fintech-focused VC (Europe)

Every fund has one AI champion. Most of them have no leverage, and they're flight risks. The pattern is predictable enough to be a tell. One analyst, associate, or rare partner is shipping Python orchestrators on a personal API key, custom front ends on Vercel, Claude skills, or artifact-based NDA reviewers that have quietly displaced hundreds of thousands in platform licenses. They are usually the only person in the firm who could explain what a skill is. Partners get excited when they see the work, but the leverage almost never reaches the rest of the team. And the strongest champions are openly considering leaving to build full-time.

"I built it because the partners get excited when they see it. It's cheap, it costs almost nothing. But people here still have no clue what Claude is. They don't even use the plugin."

Analyst, large-cap PE (London)

The real pain is program-level, not workflow-level. When funds describe what they actually want built, two workflows dominate: portfolio reporting (turning non-standardised board packs into structured KPIs and LP-ready insights) and sourcing. Sourcing is the emotionally-charged one. Competition for deals is up, the obvious signals are crowded, and the question is increasingly not "can AI save me time" but "can AI help me see signals others can't." Deeper than the workflow pain, the felt problem is that nobody owns AI internally. The investment professionals doing the building are out of their depth on system design. There is no LP narrative ready for the next fundraise. The pain that actually moves senior partners is "what do I tell LPs we're doing about AI" and "we're falling behind," not "this one workflow takes too long."

"The key bottleneck is identifying good targets. Once you see a company with great profitability, you can be sure they've had 100 people approaching them already. How do you infer signals from different data points?"

Principal, mid-market PE (Continental Europe)

Most funds don't realise they're making a build-vs-buy choice. Funds rarely think about AI as a deliberate build-vs-buy decision. They think in execution terms ("how do we accomplish X"). The funds we see getting traction with AI internally have made the choice explicit. They've invested in the AI-readiness of their own knowledge base and data layer, so agents can compound on top of it, rather than plugging LLMs into N8N-style workflow tools and rebuilding every six months.

"Are tools like N8N the way to go, or should we focus on the AI readiness of the entire knowledge base? Which approach is sustainable without rethinking the tech stack every six months?"

COO, mid-cap PE (Germany)

The diagnosis: most investment is going to the wrong layer

Read the five patterns together and the same diagnosis surfaces in each one. Funds are placing their AI investment where the leverage doesn't compound.

The model layer is converging. The gap between Claude, GPT, and Gemini on the workflows investment teams actually run is now small and shrinking. Most teams could swap their default model tomorrow and lose nothing operationally. Picking a frontier model is no longer a differentiating decision, and any platform whose value proposition is "we picked the right model for you" has been quietly commoditised.

The platform layer is more complicated. A good horizontal platform can give a fund 80% of what custom would deliver, in days rather than months. That is a real proposition. But it is structurally fragile in a way most fund teams don't price in at the point of subscribing. Multi-tenant platforms maximise revenue at the margin by serving the larger customer. As their client mix shifts, smaller funds get predictably deprioritised. The earliest adopters are the most exposed. The pain isn't malice. It's the unit economics doing exactly what unit economics do. It just happens to look indistinguishable from a vendor that stopped caring.

The workflow layer is where most automation effort goes today. Plug an LLM into N8N or a chain of skills, automate the IC memo first-draft, automate the screening checklist, automate the NDA redline. These are real wins. They also don't compound. Each automation is an isolated instance with no shared substrate. The context doesn't carry across, the work doesn't accumulate, and a new model release usually means revisiting each one to make sure the prompts still hold. Funds that build at this layer alone find that after eighteen months they have a dozen automations, none of which feel like AI is making the firm meaningfully different from peers.

What's missing in each case is the layer above. The methodology layer. The structural argument that explains why this particular fund makes money. The patterns the senior partners spot that don't fit a column in your CRM. The thesis development that goes into a sector before the first call gets booked. The judgement that says one deal is interesting and the next twelve aren't. That work doesn't live in a platform. It lives in the heads of your partners, in IC memo prose, in the back-and-forth on portfolio reviews, in the emails that never make it into Affinity. That's where the alpha is. That's the layer AI needs to encode.

What's actually compounding

Across the funds where AI is delivering real leverage rather than isolated productivity gains, the architecture looks consistent. Three layers, in order.

First, a data layer built for agentic access. Not a SharePoint migration with AI bolted on as an afterthought. A structured representation of the firm's deal universe (companies, deals, contacts, documents, metrics, decisions) with provenance, low-confidence flagging, and human review surfaces built in. The data layer is the foundation everything else compounds on. Most funds don't have it yet. The ones that do treat it as infrastructure, not a project.

Second, encoded methodology. Skills, agents, and tools with real structural constraints, not just well-tuned prompts. An extraction agent that cannot return a value without provenance. A screening agent whose output is bound by the firm's actual investment criteria, not generic LLM judgement. An IC memo workflow that pulls from real fund history rather than re-inferring it each time. The encoded version of what your senior partners would do, made systematic.

Third, a human-in-the-loop oversight surface. The work the agents do still needs to be reviewable, correctable, and traceable back to source. The funds that are getting compounding leverage treat AI output as draft material that earns its place in the workflow only when a partner can audit the reasoning behind it. That's not a constraint on AI. It's the reason the AI work survives contact with the next deal.

This architecture is slower to build than subscribing to a platform. It is more expensive than wiring up automations. It is harder to defend to a CFO with a quarterly horizon. But it is the only configuration we see compounding across the funds we've spoken to. The investment goes up the stack, not into the stack.

The road ahead

The funds we're talking to split into roughly three camps. A small group has made the build-vs-buy decision explicit and is investing in their own data layer and methodology layer. A larger group has piloted a platform or two, is feeling the deprioritisation pattern, and is starting to ask whether building is the alternative. The largest group is still figuring out what Cowork is.

In twelve months we expect the middle group to bifurcate. The funds that move up the stack will start to look meaningfully different on the dimensions partners care about: sharper conviction on screened deals, faster turnaround on diligence, more legible LP narratives on portfolio performance. The funds that stay at the model and platform layer will keep paying for tools whose marginal value is shrinking, and will be telling LPs the same "we're exploring AI" story they were telling this year.

Picking the right model never mattered very much. Picking the right platform mattered briefly and matters less every quarter. Picking which layer of your fund's actual workflow you want to encode in software you own. That is the decision that compounds.

---

Roni Schickendantz is co-founder of Routine Labs. Routine Labs builds custom AI-native systems for investment teams: encoded methodology, the data layer underneath, and the human-in-the-loop surfaces that make AI work survive contact with the next deal. If any of this resonates and you'd like to compare notes on what other funds in similar spots are doing, get in touch.