Scale AI from pilots to profit
Enterprises are racing to scale AI beyond pilots, but data quality, infrastructure, cost, and governance remain the critical bottlenecks. The MIT Technology Review Insights playbook urges companies to invest in data liquidity, adopt a multi-AI vendor strategy, and define ROI that captures both efficiency and new value. The near-term winners will pair discipline on risk and compliance with focused, business-specific use cases.
Points clés
- The survey polled 205 executives and data leaders across 11 industries; 88% were C‑suite (20% CEOs, 18% CIOs, 19% CTOs, 15% CDOs), at companies over $500 million in revenue across North America (31%), EMEA (25%), Asia‑Pacific (26%), and Central/South America (18%).
- Macro forecasts remain bullish: PwC projects a 14% lift to global GDP by 2030 ($15.7 trillion), Oxford researchers estimate 40% of mundane tasks could be automated, and Goldman Sachs sees $200 billion in AI investment by 2025.
- Adoption is broad but shallow: 95% use AI and 99% expect to, yet 76% have only one to three use cases; only 5.4% of US businesses produced a product or service using AI in 2024, while half expect enterprise‑wide deployment within two years.
- Most firms won’t build LLMs; they will apply platforms from providers such as Microsoft and Adobe, use proprietary models like OpenAI’s, and operate in “multi‑AI” environments; finance‑focused tools like Charli AI target high‑value document data.
- Cost benchmarks are steep: Google’s Q3 2023 capex hit $8 billion largely for AI; training GPT‑4 cost about $78 million in compute, while Gemini Ultra cost roughly $191 million.
- Budget pressure hits the “squeezed middle”: 47% of firms with $500 million–$1 billion in revenue cite budget constraints (vs. 22% average), while larger firms are more likely to have AI in production.
- Spending is set to surge: 9 in 10 plan to increase AI readiness outlays in 2024; for data investments, 41% expect a 10–24% rise and 37% expect a 25–49% rise.
- Data is the choke point: the top bottlenecks are data quality (49%) and data infrastructure/pipelines (44%); among $10 billion+ firms, infrastructure (55%) and data quality (52%) are most limiting.
- Risk governs pace: governance, security, and privacy are the biggest brake on deployment speed for 45% (65% at $10 billion+ firms), and 98% of executives would forgo first‑mover status to ensure safety.
- Regulation is intensifying: AI‑related laws grew from 1 in 2016 to 37 in 2022; the EU AI Act and the US executive order under President Biden emphasize risk‑based oversight, mandatory testing, and human supervision.
À retenir
Start where the real mess is: your data. Clean it, label it, and make it flow, because even the smartest model can’t fix a data swamp. Pick one or two business‑specific use cases that move a KPI you actually care about, resist the urge to build your own LLM in the garage, and set ROI metrics that track both time saved and new revenue. Embrace a multi‑AI stack, bake in governance on day one, and keep an eye on regulators—nothing kills momentum like a surprise audit. Do this, and you’ll scale AI without setting your GPU budget (or your reputation) on fire.
Sources
Quiz sur le document: 10 questions
