Chris Towers, Roar Media
We were joined by some brilliant organisations at the AI in Business Conference last month in London with talks coming from Unilever, GSK, CBRE, Visa and more. Amongst them our fantastic event partner Transition Technologies PSC UK&I were on stage sharing some perspectives, here we break down the key takeaways from an informative session.
For the past two years, enterprise AI strategy has been driven by one headline: LLMs will change everything. The assumption was simple, if a model can pass the bar exam and write poetry, surely it can help a bank underwrite loans, or help a manufacturer plan procurement. But anyone working in enterprise AI already knows the truth: a standalone LLM is not an enterprise solution. An LLM can predict words, but it can’t tell a risk officer how affordability is calculated in their internal credit model. It can’t compute exposure trends for a specific corporate borrower. It can’t pull ERP data or validate it against policy.
That gap between general intelligence and enterprise grade decisioning is where the real transformation is happening. And the most forward-leaning organisations are moving past “chat with documents” toward what many are calling agentic AI.
Let’s unpack why.
LLMs Are Evolving from Knowledge Models into Decision Engines
Most large enterprises now understand the limitations of text-only models trained on public data. Yes, they can reason and generate content. But business users don’t ask abstract questions, they ask operational ones:
- “How is our affordability score calculated according to internal policy?”
- “Can we use Open Banking data in our risk models—and under what conditions?”
- “Give me the current exposure, PD movement and red-flags for HorizonTech LLC.”
Those aren’t hypothetical examples, they’re real prompts from an AI demo grounded in internal documents and structured company data. The answers are policy-accurate, reference specific governance requirements (e.g. Open Banking Data Handling Standard), and even incorporate explainability constraints such as SHAP methods.
That is not text generation anymore.
That is enterprise reasoning.
And importantly—it illustrates the future architecture:
LLM → retrieves facts → queries data sources → interprets results → guides action
The shift is unmistakable: from predicting words to achieving goals.
Talking to Documents Was Step One — Talking to Data Is the Real Prize
Early enterprise experimentation focused on RAG, vectorising PDFs and letting an AI assistant answer questions based on internal content. It was a huge improvement over hallucinated “best guesses,” but it was also limited.
Static documents don’t capture:
- real-time exposure
- covenant breaches
- PD shifts
- liquidity position
- utilization changes
- network ownership relationships
When the knowledge lives in databases, not PDFs, a retrieval system must evolve. And that’s where tool calling and text-to-SQL enter the picture.
A modern agent can:
- Interpret intent (“calculate exposure change”)
- Generate structured SQL
- Query the right system
- Return numbers in context
- Flag compliance-relevant anomalies
That is no longer a chatbot.
It’s a performer.
Why Multi-Agent AI Matters
Enterprise workflows aren’t monolithic. A credit decision isn’t one question, it’s a chain of analytics steps:
- retrieve policy
- pull risk data
- evaluate thresholds
- test eligibility
- surface rationale
- escalate exceptions
Expecting a single LLM to manage all of that is unrealistic.
The emerging pattern is multi-agent orchestration:
- A retrieval agent finds policies.
- A compliance agent checks governance rules.
- A data agent retrieves exposure and PD.
- A reasoning agent synthesizes.
- A presentation agent structures results.
This is how future enterprise systems will operate, not as a single “gen AI app,” but as an ecosystem of specialised AI workers.
The Data Bottleneck: Enterprises Need Unified Access Layers
Anyone who has tried to point an LLM directly at production data knows the pain:
- fragmented systems
- CRM / core banking / credit models
- different schemas
- inconsistent timestamps
- no governance layer
- no context for interpretation
Enterprises need a unified data representation; AI Data Marts, vector layers, metadata governance, and model control planes. Without that foundation, agentic systems collapse under their own complexity.
A modern strategy includes:
- structured and unstructured data harmonization
- vector DBs for semantic retrieval
- Data Marts for analytic retrieval
- MCP servers or tool orchestrators
- Explainability as a first-class requirement
Without this, enterprises have “cool demos” but no deployable systems.
Business Value Is Shifting Toward Intelligent Decisioning
The investment wave is moving decisively beyond chatbots:
- Decision support
- Automated adjudication
- Risk assessment
- Planning and forecasting
- Compliance guardrails
- Process automation
Agentic systems are enabling system-to-system intelligence, not just “ask me anything” interfaces. Think about adjusting a PD threshold and instantly modelling portfolio impact, no analyst required.
This is execution, not conversation.
Enterprises Are Asking the Right Questions: Where’s the ROI?
Smart organizations start with three questions:
- Is there a business problem AI can realistically solve?
- Do we have data maturity to support it?
- What is the roadmap to adoption?
This is where advisory models are maturing: correlation studies, investment prioritisation, readiness assessments, and cross-industry benchmarks (finance, telecom, manufacturing).
It’s no longer “should we do AI?”
It’s “what part of our P&L moves first?”
Agentic AI Is the Next Platform Shift
Look at current market behaviour:
- Spend is moving toward automation and intelligent decisioning
- Multi-agent collaboration is becoming a baseline capability
- AI ecosystems are shifting toward autonomous analysis and rules-based action
The organisations that master orchestration, data, governance, tool layers, and compliant reasoning will unlock the competitive advantage. Everyone else will still be testing chatbots against PDFs.
And here’s the strategic takeaway:
Enterprise AI is no longer about answering questions. It’s about taking action grounded in organizational truth.
That is what will differentiate operational AI from entertainment.
Final Thought: The Winners Will Make AI Boring
The hype cycle is fading.
Now the work begins and the ROI follows.
The companies that succeed in 2025 won’t chase magical models. They will:
- operationalise data
- orchestrate agents
- enforce governance
- automate decisions
- integrate with systems
- prove outcomes
When AI disappears into workflows and compliance dashboards, that’s when it’s won.
Enterprise AI is becoming boring and that’s when transformation gets real.
Find out more about Transition Technologies PSC UK&I here.