We had the privilege of delivering a masterclass at Futurescot’s annual Public Sector AI conference in Glasgow in May, Scotland’s leading forum for exploring how artificial intelligence can reshape public services.

The auditorium was full of leaders genuinely committed to making AI work, not just talking about it. And for those in the public sector, making it work means one thing above all: predictable costs, as much as predictable outcomes.

Our session focused on what the industry doesn’t discuss enough – the gulf between a promising AI pilot and a fully scaled, cost-effective solution in live service.

There are several reasons for that gap, and in the public sector there are particular ‘traps’ that AI projects can fall into, as well as common mistakes. Here’s our take on some common traps to watch out for, some thoughts on how they might be avoided, and cause for optimism on how the AI infrastructure landscape may be shifting.

The Pilot Trap – 95% of Pilots Fail to Go Live

Across the enterprise landscape, we’re seeing a familiar pattern. A team runs a successful proof-of-concept, generates impressive numbers and then the project stalls. It doesn’t fail dramatically, because pilots are designed to prove a point. Delivery is designed to serve people efficiently with the lowest possible cost of AI infrastructure. The gap between those two goals is wider than most organisations plan for.

The Token Trap – Spiralling Costs

The ‘Token Trap’ is when the cost of AI exponentially grows despite the unit costs of the technology coming down. For scaled enterprise deployments (e.g. in software engineering or automated customer support), a fraction of a penny per token compounds into 6-or-7 figure monthly bills.

According to Gartner, agentic models, for example, require between 5-30 times more tokens per task than a standard GenAI chatbot, and can perform many more tasks than a human using GenAI. While lower token unit costs will enable more advanced GenAI capabilities, these advancements will drive disproportionately higher token demand. As token consumption rises faster than token costs fall, overall inference costs are expected to increase.

Workflow First, Technology Second

The most common mistake we see is starting with the model and working backwards to find a use case. The questions that matter are simpler: Where does a caseworker spend two hours doing something that should take twenty minutes? Where does a citizen wait three weeks for a decision that could be made in three days?

Pega’s AI Manifesto captures this directly. One of its nine principles states that intelligence is useless if not put into action in business processes and customer interactions. The manifesto urges organisations to start with outcomes, identify the workflows to impact, and only then determine which AI capabilities drive value. For public sector organisations navigating stretched budgets, this isn’t just good advice: it’s the difference between a project that delivers and one that doesn’t.

Augmentation, Not Automation: Citizens First

The theme that generated the most discussion in our session was that AI in public services should augment human judgement, not replace it – with the citizen, not the system, always being the starting point.

Public sector workers bring contextual knowledge, empathy, and discretion that no model replicates. The goal should be to free them from tasks that don’t require those qualities, so that they can bring more of them to the moments that do. The Pega AI Manifesto puts it plainly: AI is augmented intelligence and it is best with humans in control. If a solution makes life easier for a council and harder for the resident, it isn’t a success.

PegaWorld 2026: Ending the Token Tax

Fresh from Futurescot, one announcement from PegaWorld 2026 in Las Vegas grabbed attention for exactly the right reasons. Pega announced that with Pega Infinity 26, organisations can design, build, and run agentic workflows without paying per token. Their Predictable AI architecture shifts heavy reasoning to design time, making runtime agents faster, more reliable, and dramatically cheaper to scale.

This matters enormously for public sector delivery. Token-based pricing has been a quiet killer of scaling ambitions since unpredictable costs make it nearly impossible to build a business case beyond the pilot stage. A flat per-case model changes that entirely. As Pega CEO Alan Trefler said: “Token-maxxing can only lead to unsustainable costs and unpredictable results.”

Finally, Scotland has real strengths to build an AI environment for a collaborative public sector culture, national digital ambition, and genuine willingness to have honest conversations about what AI can and can’t do. The path from pilot to delivery requires discipline about workflow, honesty in planning, and a commitment to building systems that serve citizens first.

Contributing author: Kajal Kaspate, Pega AI Innovation Lead