Roughly 90% of large organisations now use AI for coding, and 86% of those have it in production, according to Anthropic's January 2026 State of AI Agents Report. The problem now defining enterprise AI is the gap between AI activity and AI return. The technology works, but most organisations still lack the operating model to convert engineering capability into measurable outcomes across the business.
That was the question running underneath every session at our Applied AI Symposium, hosted at Indicium AI’s London office on 14 May 2026 and sponsored by Anthropic and AWS. Around 150 senior leaders gathered to take it on: C-suite, Heads of Data and AI, engineering directors, and product and strategy leaders from companies including JPMorganChase, NatWest Group, Citi, and EDF.
Ralph Ramos of Anthropic opened with The Agentic Revolution keynote. The Financial Services panel followed with leaders from Experian, Schroders, Howden, and AWS. The Energy & Utilities panel brought together Shell, AWS, and independent advisor Damien Buie. The day closed with a fireside chat featuring Louis Joubert of LSEG (London Stock Exchange Group), Anna Bratton of Anthropic, and Steve Bryen, Indicium AI's CTO.
Across all four sessions, the same themes kept returning.
1. The unit of AI work just got bigger
Ralph opened with a chart that reframed how most leaders had been thinking about AI productivity. It traced how the "unit of work" that AI now performs has grown year on year: from a line of code in 2023, to a function in 2024, to a feature in 2025, to a project in 2026. With each step, the model has done more of the production and the developer more of the steering.
The reference customer points are striking. Spotify reports 90% of engineering time saved on migrations. Rakuten ran an autonomous refactor across a 12.5-million-line codebase in seven hours, running in production. Stripe converted 10,000 lines from Scala to Java in four days against an estimated ten engineering-weeks, and is now pushing 1,200 pull requests to production per week. Ralph's framing: "The question isn't whether your company adopts. It's whether the rest of your company catches up to your engineers."
Louis Joubert of LSEG flagged the counterweight to all of this in the fireside chat: every line of code AI produces, is also a line that has to be maintained. Code bloat is a real enterprise liability, and as developer leverage scales, so does the maintenance surface area.
2. The constraint has moved, from model to organisation
If the unit of work has scaled, the binding constraint has shifted with it. The Financial Services panel was direct: the model-quality arms race has slowed into incremental improvements, and the real constraint is now the harness around the model: domain context, MCP (Model Context Protocol) layers, skills libraries, and organisational wiring.
For a decade, the enterprise mantra was "data is the moat." The argument now is that context is the moat: workflow logic, institutional memory, and the domain metadata that determine whether AI generates something generically competent or genuinely valuable. As AWS's Mathias Athwal put it on the panel, context should be seen as an asset in the business, and the process around it is now the constraint. Documentation, MCP layers, and skills libraries are no longer overhead; they are strategic intellectual property.
Ralph put numbers behind the same idea. Roughly 65% of the average knowledge worker's day is "the middle" that AI compresses: chasing data, reconciling spreadsheets, summarising threads, drafting the first pass. Planning and the judgement calls a regulator might ask about remain firmly with humans. The leverage is enormous, but capturing it means redesigning how teams work, not buying more model capacity.
Richard Bruckshaw of Schroders gave a concrete example of this redesign: an agentic Software Development Lifecycle (SDLC) in which AI is embedded across build, test, and review stages, replacing manual testing with automation. As Richard put it, the speed-up means controls have to become more explicit, not less.
3. Trust, agency, and the regulated industries
The Energy & Utilities panel grounded the agentic conversation in a sector where the cost of a poor decision is measured in megawatts, safety incidents, and grid reliability rather than basis points. The tension is structural: heavy regulation, critical national infrastructure, and multi-year IT and capital cycles.
Ofgem (the UK energy regulator) and its AI sandbox, together with the EU AI Act, are reshaping the cost-benefit calculation, and the panel argued AI may end up redefining the energy system itself, alongside sustainability commitments such as Shell's 2050 net-zero target. AWS's Carol Yan framed the operating-model question for energy leaders: value creation happens decentrally, but value acceleration happens centrally.
The Financial Services panel arrived at the trust question from the opposite direction. As consumers expect AI to manage meaningful parts of their financial lives, agentic trust becomes a new problem: an AI acting on someone's behalf, making consequential decisions, while identity verification and fraud detection get materially harder when the entity making a request might itself be an agent.
In the fireside chat, Louis Joubert framed the discipline that regulated environments demand in a single line: "Don't trust until you have verified, and keep checking." Anna Bratton of Anthropic extended it: even as frontier models grow more capable, model drift is a production reality, and enterprises still need to design human-in-the-loop checkpoints into agentic workflows. The Financial Services panel's version: the old approval gates were load-bearing. Rebuilding that discipline as an automated process is the real design challenge.
4. From project to operation: closing the GenAI divide
The most quoted study of the afternoon was MIT's GenAI Divide, which found that 95% of enterprise GenAI pilots show no measurable P&L impact. Its conclusion went further: "This divide does not seem to be driven by model quality or regulation, but seems to be determined by approach."
The 5% of organisations delivering measurable returns shared four characteristics: they were sourced from the frontline, measured on business outcomes, owned by one accountable leader, and built with people who had done it before.
Steve Bryen pressed the same point in the fireside chat: "The job is not coding. It's building systems, architecture, operations. The job is about delivering outcomes to the business." The missing capability in most enterprises is operational rather than technical. Running AI in production is an operation, sometimes a 24/7 one, and demands governance, MLOps, change management, and partnership structures to match.
Louis Joubert named the three forms of readinesses that determine whether enterprise AI scales beyond pilots: infrastructure, data, and operational. Missing any one turns "AI in production" into a single fragile system rather than a sustainable capability.
The LSEG and Moody’s data-through-Claude partnership point in the same direction: enterprise-grade AI is becoming an ecosystem play. As Anna Bratton put it, there is no established blueprint for partnerships at this scale; what makes them work is mutual trust and deep domain expertise from both sides.
Ralph closed on a Novo Nordisk case study. Clinical study reports cut from ten weeks to ten minutes, with a 95% reduction in verification resource, built by a team of eleven led by a molecular biologist. Frontline, outcome-measured, single owner, borrowed expertise. Exactly the MIT GenAI Divide profile for companies delivering measurable AI returns.
The path forward
1. The constraint has moved. Context, governance, and organisational wiring are the new differentiators, not model quality.
2. The old friction was load-bearing. Rebuild the discipline that approval gates provided into automated processes before the gap becomes visible in production.
3. Source from the frontline. Outcome-measured initiatives owned by a single accountable leader outperform centrally specified ones.
4. Run AI as an operation. Production-grade AI is a 24/7 capability, not a quarterly deliverable.
5. Solve agentic trust before the product reaches customers. Identity, accountability, and human oversight need designing in, not retrofitting.
To benchmark your AI operating model or request an AI Transformation diagnostic, talk to the Indicium AI team.
