The State of AI in Enterprise: A 2025 Practitioner's Outlook
Beyond the hype — a candid assessment of where AI is genuinely delivering ROI and where organisations are still burning budget.
Two years into the generative AI era, the gap between organisations getting real value from AI and those burning budget on pilots is widening. Having shipped AI systems for a range of enterprise clients, we have a clear view of what's working and what isn't.
Where AI Is Actually Delivering ROI
The highest-ROI AI use cases we've seen are all variations of the same pattern: taking something a human currently does slowly, inconsistently, or at high cost — and either automating it entirely or making the human dramatically faster. Document processing, customer query triage, code review assistance, and data extraction from unstructured sources top the list.
What these use cases share: they have clear success metrics, the data quality is manageable, and the human-in-the-loop oversight requirement is well understood. They are not trying to replace human judgement — they're augmenting human throughput.
Where Budget Is Still Being Burned
The losing pattern is equally consistent: an AI strategy built around capabilities rather than problems. 'We need an AI chatbot' is not a problem statement. 'Our support team handles 800 repetitive queries per day that could be resolved without human intervention' is. When you don't start with a measurable problem, you can't measure whether your AI solution solved it.
The Engineering Reality of Production AI
The single biggest surprise for organisations moving from AI pilot to production: the model is 20% of the work. The other 80% is data pipelines, evaluation frameworks, monitoring infrastructure, prompt management, safety guardrails, and the human review workflows that keep the system trustworthy.
Organisations that treat AI as a model selection exercise — rather than a software engineering challenge — consistently underestimate implementation timelines by 3–5×. Budget accordingly, or don't start.
Found this article useful?
Share it with your network on LinkedIn.
Key Takeaways
- Highest ROI comes from augmenting human throughput, not replacing human judgement
- Start with a measurable problem — never with a technology capability
- The model is 20% of the work; infrastructure and monitoring are 80%
- Budget 3–5× longer for production AI than a proof-of-concept suggests
- Human-in-the-loop oversight is not a weakness — it's a production requirement
Ready to act on this?
Talk to our team about applying these ideas to your organisation.
Start a Conversation