Why winners rebuild around models: how ‘future-built’ organisations capture AI gains

What the evidence is actually saying

The headline from the research cited in the original piece is straightforward: it’s not enough to licence a model. Organisations that capture the largest revenue and cost benefits are those that rebuild processes, incentives and infrastructure around AI. BCG’s analysis of “future-built” firms finds roughly five times the revenue impact and three times the cost reduction versus laggards, and it notes material budget shares going to agentic systems (about 15% in their examples). Those numbers point to a pattern: gains concentrate where AI is embedded rather than bolted on.

Why rebuilding matters (not just buying)

Putting a model behind an existing interface or asking teams to use a cloud API rarely changes the sociotechnical dynamics that determine long-term value. Three things typically need to change together:

• Incentives and ownership. Who benefits when the model improves throughput, quality or sales conversion? If the reward sits with a central AI team while product teams face the costs of integration, adoption stalls.
• Modern infrastructure. Models need low-latency inference, data plumbing for continuous improvement, observability and secure access. Legacy stacks create friction that erodes realised value.
• Processes for human–AI collaboration. Teams must learn how to work with model outputs as routine artefacts — editing, validating, escalating and feeding back corrections into training data.

Practical guidance: how to become “future-built”

Below are concrete steps you can take to shift from pilots to persistent productivity gains. They map to the same levers BCG highlights: capability investment, reinvestment, and a willingness to allocate budget to more agentic approaches.

• Start with an outcomes map.

  Don’t begin with a model. Map business outcomes (revenue flows, cost buckets, customer moments) to evidence of where AI could shift those metrics. Prioritise opportunities where incentives are already aligned — teams that gain from improved metrics should also control the change budget or receive shared returns.

• Create cross-functional squads with skin in the game.

  Put product managers, engineers, data scientists and the domain SMEs into one accountable team. Make success measurable (A/B tests or KPI deltas) and ensure savings or revenue lift is at least partially reinvested into the operating unit.

• Invest in the plumbing before the glamour.

  Cheap model access is seductive, but production value comes from observability, feature stores, retraining pipelines, and secure data flows. Treat these as first-class products so iteration is fast and safe.

• Make human–AI routines explicit.

  Define clear workflows for when a model’s output is accepted, edited or escalated. Train people to provide feedback in a way that can be instrumented for learning. A small set of repeatable routines turns one-off assistance into scalable productivity.

• Allocate a purposeful budget to agentic capabilities.

  BCG’s examples suggest meaningful spend on agentic AI — systems that can take multi-step actions within constraints. Reserve a portion of AI investment to pilot agents where they replace repetitive tasks or coordinate workflows, and measure their downstream impact on cycle time and quality.

• Define governance that enables fast iteration.

  Lightweight guardrails (risk tiers, rapid rollback, human-in-the-loop checks) are better than heavyweight approvals that delay learning. Governance should enable safe experimentation and capture lessons to scale the winners.

• Reinvest returns into people and architecture.

  Rather than centralising cost savings as immediate margin expansion, reinvest a portion into training, hiring and platform improvements. That compounds return and reduces the gap between early adopters and laggards.

What success looks like in day-to-day work

In future-built teams, interacting with an AI is routine: agents or model outputs appear in the same workflow as other tools, humans know when to trust or correct outputs, and data from those interactions flows back to continuous improvement pipelines. The organisation measures not just raw model performance but downstream business metrics — revenue per customer, lead-to-deal time, error rates — and ties those to reinvestment decisions.

Common pitfalls to avoid

• Treating models as a one-off productivity feature and failing to fund platform work.
• Centralising savings without giving teams ownership of reinvestment decisions.
• Neglecting human workflows: automation that removes human agency will create brittle, low-adoption systems.

How to measure progress

Track both technical and sociotechnical metrics: deployment frequency, mean time to recover, proportion of decisions where model output was accepted, downstream KPIs (conversion, cost per transaction) and reinvestment rate. These metrics let you see whether a model’s outputs translate into organisational advantage, rather than just technical maturity.

Expanded from

This post expands on a point from Sociotechnical systems in software and AI: Who actually captures the productivity gains?.

Sources

Sociotechnical systems in software and AI: Who actually captures the productivity gains? (original article summarising BCG research)