Beyond Go-Live: The AI Operating Model That Compounds
Most organisations procure AI the way they procured software in 2012 — as a bounded project with a delivery date, a handover, and a support arrangement. That model was designed for stable technology. AI is not stable. Here is the operating model that actually compounds.
73% of AI deployments fail within six months. Not because the technology fails — but because organisations treat deployment as the finish line.
This finding, from MIT Sloan's 2024 research on enterprise AI, captures the central problem: the commercial and operational model that most organisations use for AI was designed for a different kind of technology. It was designed for software that, once deployed, is essentially stable. AI is not stable. It is never finished.
The organisations capturing compounding AI value — not just first-year returns, but sustained improvement over years — have a fundamentally different operating model. They are not running projects. They are running partnerships designed for continuous evolution.
The Model That Doesn't Work
Most enterprise AI is still procured the same way software was procured in 2012:
- Define requirements
- Select a vendor or delivery partner
- Build the solution
- Handover to operations
- Maintain via support tickets
This model works when the technology is stable. A CRM system installed in 2018 was largely the same system in 2022. The "done" state was real. Maintenance meant fixing bugs and applying security patches — not fundamentally evolving what the system does.
AI breaks every assumption in this model:
Models are updated or deprecated continuously. GPT-4, GPT-4o, GPT-4.1 — each version changes performance characteristics, cost profiles, and behaviours, often with limited migration runway. A system built for one model version requires active stewardship to remain accurate on the next.
Knowledge infrastructure drifts. RAG pipelines retrieve from documents. Documents go stale. Internal processes change. Nobody built the process for keeping context current. The AI confidently returns answers from a world that no longer exists.
The capability landscape evolves faster than handover cycles. Something that wasn't possible when the project was scoped becomes possible six weeks after delivery. Teams that were handed a fixed deliverable cannot absorb it.
Value compounds through iteration, not installation. IBM's internal AI transformation — delivering $3.5 billion in productivity savings since 2023 — took six years of continuous iteration to reach its current state. AskHR didn't achieve 94% automation in a project cycle. It got there through measurement, calibration, and expansion over years.
When you apply a project model to technology that behaves this way, you capture a fraction of available value — and watch that fraction erode as the system drifts out of alignment with reality.
The Operating Model That Compounds
The alternative is an engagement model designed for ongoing evolution, not for completion. It differs from the project model in six structural dimensions.
Model Comparison
The Project Model vs The Compounding Model
Six dimensions that define how AI value is captured — or lost
Select a dimension to explore how the two models differ — and what the research says
Each dimension above is interactive — select any one to see how the two models differ in practice, what the consequences are, and what the research says.
The pattern across all six dimensions is consistent: the project model optimises for a successful delivery milestone; the compounding model optimises for sustained value creation over time. These are genuinely different objectives, and they produce genuinely different commercial and operational structures.
The Compounding Cycle
AI value compounds through continuous cycles — not through a single deployment event. The organisations that capture the most value run five stages on a continuous rhythm.
The Compounding Cycle
Deploy → Measure → Calibrate → Expand → Evolve
AI value compounds when the cycle runs continuously — and decays when it stops
Select any stage to explore what it involves — and where organisations typically get stuck
The critical insight is that this cycle never concludes. Stage 5 (Evolve) feeds back into Stage 1 (Deploy) — at a higher baseline, with more capability, with more institutional knowledge about what works. Each complete cycle makes the next more effective.
Organisations that stop the cycle — that treat deployment as the finish line and move the team to the next initiative — capture whatever value was available at that moment and watch it decay. The MIT Sloan research on AI compounding found that organisations with systematic feedback loops between humans and AI are six times more likely to derive substantial financial benefits. The cycle is the mechanism through which that advantage is realised.
Why Most Organisations Get Stuck
BCG's research on AI transformation offers a useful diagnostic: the 10/20/70 rule. Algorithms account for 10% of AI transformation. The technology backbone accounts for 20%. The remaining 70% is people and processes — change management, role evolution, workflow redesign, measurement discipline.
The project model handles the 30% well. It scopes requirements, builds systems, and delivers technology. What it does not handle is the 70% — and the 70% is what determines whether value compounds or decays.
This is where the failure rate comes from. It is not that organisations deploy bad AI. It is that they deploy good AI with no operating model for the ongoing human and process work that AI requires. The technology works. The people and workflows around it do not evolve to match. Adoption decays. The AI is quietly bypassed. The initiative is declared a partial success and the organisation moves on.
What Compounding Looks Like in Practice
EY's AI Pulse Survey found that among companies seeing significant productivity improvements from AI, nearly all are reinvesting those gains — with large portions flowing directly back into AI adoption initiatives and R&D. This creates a reinforcing flywheel:
- Early AI deployments generate productivity and capability gains
- Those gains fund more ambitious AI programmes
- More ambitious programmes require better data, tools, and operating models
- Better operating models support more impactful AI systems
- More impactful systems generate more gains to reinvest
This is not a theoretical pattern. It is the observable behaviour of organisations at the front of enterprise AI adoption. They are not treating AI as a discrete investment with a discrete return. They are treating it as compounding capital — and the organisations that start early build advantages that late entrants cannot easily close.
The research is consistent on this point. A 2025 study of 1,250 organisations found that only 5% achieve substantial AI value at scale. That 5% — the organisations that measure outcomes and reinvest them into better data and workflows — see 1.7× greater revenue growth and 3.6× higher shareholder returns over three years. The gap is not closing. It is widening.
The Commercial Implication
If AI value requires a compounding operating model, then the commercial structure of AI engagements must support it. Project-based pricing — fixed scope, fixed timeline, fixed fee, handover at the end — is structurally misaligned with compounding value.
The alternatives are emerging:
Managed service models — ongoing accountability for system performance, with pricing tied to value delivered rather than hours worked. The partner remains responsible for model stewardship, knowledge maintenance, and continuous calibration.
Outcome-based pricing — commercial alignment around business outcomes rather than project milestones. The partner's incentives are aligned with compounding value, not project completion.
Hybrid models — initial project delivery followed by ongoing operational partnership, with commercial terms that reflect both phases and the transition between them.
The right structure depends on the organisation, the use case, and the partner. What matters is that the commercial model explicitly supports ongoing evolution — that it does not create incentives for handover and departure at the moment when ongoing stewardship becomes most important.
What This Means for Partner Selection
If the operating model is the primary determinant of AI value — and the research consistently suggests it is — then partner selection criteria need to reflect it.
The questions that matter are not just about capability:
- Can they build what we need?
- What is the delivery timeline?
- What is the project cost?
The questions that matter are about partnership:
- What does the relationship look like two years after go-live?
- What is the total cost of the relationship over three years?
- What is their methodology for continuous calibration?
- Who is accountable for model performance after the initial deployment stabilises?
We explore this in depth in Evaluating AI Partners, Not AI Vendors — the companion piece to this article. The short version: vendor evaluation criteria designed for software procurement do not work for AI. The relationship you need is a partnership designed for ongoing evolution, not a vendor relationship designed for delivery and handover.
The Decision in Front of You
The question is not whether to adopt AI — that decision is already made for most organisations. The question is whether to adopt it with an operating model designed for compounding value, or with a model designed for project delivery.
The project model is familiar. It fits procurement processes. It aligns with how organisations have bought technology for decades. It is also structurally incapable of capturing the value that AI makes available.
The compounding model requires a different relationship — with partners, with internal teams, with the timeline of value creation. It is unfamiliar. It does not fit neatly into existing procurement categories. It is also the model that the organisations at the front of enterprise AI are using to build advantages that widen every cycle.
The technology is ready. BCG's research is clear on this: "The models are ready. The challenge is everything around them — the systems, data, and workflows they depend on and the organizational capabilities needed to use them at scale."
The operating model is the challenge. And the operating model is a choice.
Sources
- MIT Sloan Management Review (2024, 2025): How to Reap Compound Benefits from Generative AI
- BCG (2025): Enterprise as Code: An Operating Model for the AI Era
- BCG (2025): Agents Accelerate the Next Wave of AI Value Creation
- BCG (2026): Scaling AI Requires New Processes, Not Just New Tools
- EY (2025): AI-driven productivity is fueling reinvestment over workforce reductions
- IBM (2025): AskHR — 94% automation rate, six years of continuous iteration
Start with one workflow.
Map it. Separate predictable from creative. See exactly where AI adds value — and where it doesn't.