In April 2024, we published The Compound Advantage, our thesis that enterprise AI investments should compound, not accumulate. One year later, we have the data to prove it. The organisations that built foundations are deploying their third and fourth capabilities at 40% the cost of their first. The organisations that didn't are still restarting from scratch every time.
What You Need to Know
- The compound thesis is proven. Across our enterprise engagements, foundation-first organisations are delivering their third AI capability at 40% the cost and 35% the time of their first. Project-mode organisations show no improvement between capability #1 and capability #3.
- The gap is accelerating, not linear. Foundation builders aren't just ahead. They're pulling away faster with each capability. The compounding effect is more pronounced than we originally modelled.
- Data infrastructure is the biggest compounding asset. Organisations that invested in shared data pipelines and knowledge bases saw the largest returns. Those that built isolated data stores for each capability captured zero compound value.
- Governance compounds too. The most overlooked finding: governance frameworks that were expensive to establish for capability #1 became nearly free for capabilities #2-4. Organisations without governance are still paying the full approval cost every time.
- The switching cost is real but manageable. Three organisations in our data set transitioned from project mode to foundation mode mid-programme. The retrofit cost 3-4 months, painful but far less expensive than continuing the project-by-project approach.
40%
cost reduction by capability #3 for foundation-first organisations
Source: RIVER Group, enterprise engagement data, 2024-2025
0%
cost improvement by capability #3 for project-mode organisations
Source: RIVER Group, enterprise engagement data, 2024-2025
What We Predicted vs What Happened
When we published the original thesis, we made three claims. Here's how they held up.
Claim 1: "Capability #4 at 4x faster delivery"
Predicted: Fourth capability delivered in roughly 25% of the time of the first.
Actual: Closer to 30% of the time. Still dramatic, but slightly less than modelled. The reason: integration complexity doesn't reduce as linearly as infrastructure reuse. Even with shared pipelines, connecting to a new business system takes time.
The more interesting finding: quality improved faster than speed. Capability #4 wasn't just faster. It was more accurate, better governed, and had higher user adoption on day one. The team knew the patterns. The users knew the interfaces. The governance was established.
Claim 2: "50% total investment savings by capability #4"
Predicted: Cumulative 50% savings across four capabilities compared to standalone approach.
Actual: 47% savings on average, with significant variance. Organisations with strong data infrastructure saw 55-60% savings. Those with weaker data foundations saw 35-40%.
The data infrastructure finding was the strongest signal in our analysis. Shared data pipelines, unified knowledge bases, and consistent embedding strategies accounted for more compound value than any other architectural decision.
Claim 3: "The breakeven point is between capabilities #2 and #3"
Predicted: Foundation approach becomes cheaper than project approach somewhere during the second capability.
Actual: Confirmed. Every foundation-first organisation in our data set achieved cost parity by the midpoint of capability #2. By the end of capability #2, they were ahead. By capability #3, the gap was substantial.
47%
average total investment savings by capability #4 on foundation approach
Source: RIVER Group, enterprise engagement data, 2024-2025
The Surprises
Three findings we didn't anticipate:
1. Organisational Learning Compounds Faster Than Technology
We focused our original thesis on infrastructure reuse: shared pipelines, knowledge bases, integration patterns. What we underestimated was how much organisational capability compounds.
By capability #3, teams knew how to scope AI work. Product owners understood what AI could and couldn't do. Change management was faster because people had been through it before. Governance reviews took days instead of weeks because the framework was established and the reviewers had pattern recognition.
This organisational compound effect was as valuable as the technical one, and harder to replicate. You can copy someone's architecture. You can't copy their team's experience.
2. The "Dead Zone" Between Capabilities #1 and #2
Several organisations experienced a momentum gap after their first capability went live. The first deployment consumed significant executive attention and organisational energy. After launch, there was a natural exhale, and in some cases a 2-3 month pause before capability #2 began.
This dead zone is dangerous. It's where foundation investments feel expensive (you've paid the premium but haven't seen the compound return) and where project-mode advocates argue for a simpler approach next time.
The organisations that avoided the dead zone had one thing in common: they scoped capability #2 during the build phase of capability #1, not after. By the time capability #1 launched, the team transitioned immediately.
3. Governance Is the Highest-Return Compound Asset
We listed governance as a compound benefit in the original thesis, but we underestimated its impact. The governance framework (use case approval, risk classification, monitoring protocols, incident response) is expensive to create for the first capability. For capabilities #2-4, it's essentially free.
More importantly, established governance accelerates approval. A new AI capability that fits within an existing governance framework gets approved in days. One that requires building governance from scratch takes weeks to months. In one organisation, the governance time savings alone accounted for 30% of the total time reduction between capability #1 and capability #3.
The Foundation Builders vs The Project Runners
Here's what the two approaches look like at the one-year mark, drawn from real engagements:
Foundation Builder A (insurance sector): Started with claims intelligence on a shared platform. Built fraud detection in 60% of the time. Added customer communication in 45% of the time. Knowledge base now contains 15 years of claims data, policy documents, and compliance frameworks. Every new capability inherits this knowledge.
Project Runner B (financial services): Started with a customer sentiment tool. Built a separate document classification system. Started a third initiative for compliance monitoring. Each has its own data pipeline, its own vendor, its own governance approval process. Total investment: 2.3x what Foundation Builder A spent for comparable functionality.
The most telling metric: Foundation Builder A's team talks about "the platform." Project Runner B's team talks about "the projects." Language reflects reality.
What This Means for 2026
The compound advantage is no longer a thesis. It's an observable pattern with a year of evidence behind it. For organisations planning their 2026 AI strategy, the implications are clear:
If you haven't started: The cost of delay is compounding against you. Every quarter that foundation builders add capabilities, the gap widens. Start with a structured discovery and build your first capability on shared infrastructure.
If you're in project mode: The retrofit is worth it. Yes, it costs 3-4 months to restructure around a foundation. But that investment pays back within one capability cycle. The alternative (continuing to build isolated projects) guarantees you'll never see compound returns.
If you're already building on a foundation: Push through the dead zone. Scope capability #2 before capability #1 launches. Invest in your data infrastructure. It's the single highest-return compound asset. And measure the compound metrics: time-to-capability, foundation reuse percentage, governance approval time.
The compound advantage isn't a theory any more. It's a measurement. The organisations that understood this a year ago are now pulling away from those that didn't. The good news: it's not too late to switch. The bad news: every quarter you wait makes the switch more expensive and the gap harder to close.
- Is the compound effect specific to certain industries?
- No. We've observed it across insurance, financial services, government, and professional services. The magnitude varies (industries with richer data estates see larger compound effects) but the pattern is consistent. Any organisation building more than two AI capabilities benefits from the foundation approach.
- Can you retrofit a foundation under existing AI projects?
- Yes, and three organisations in our data set did exactly this. The retrofit takes 3-4 months and involves consolidating data pipelines, unifying knowledge bases, and establishing shared governance. It's more expensive than building the foundation from the start, but significantly cheaper than continuing the project-by-project approach.
- What if our first AI capability failed? Does the compound thesis still apply?
- If it failed because the use case was wrong, the compound thesis still applies. You just need a better first use case. If it failed because the foundation was poorly built, that's an architecture problem, not a thesis problem. In either case, the answer is the same: invest in getting the foundation right, then compound from there.