AI for UK Financial Institutions
Outline of Management Impacts

updated 1 May 2025

  • The coming of massive processing power brings fundamental change
  • The transformer architecture and LLMs also brings fundamental change
  • Change has accelerated with the emergence of capable AI agents
  • Enterprise valuations can be significantly raised if effective AI action can be shown
  • Overheads and costs can be radically reduced
  • The technology is moving so fast that a methodical, systemic response is needed
  • A major spur to action is that all existing IT projects may themselves be obsolescent
  • The smart way forward is to have a priority list, get results, learn, re-prioritise, apply.

BD 1 May 2025

  • AI agents have already transformed the potential of AI for enterprises, now the pace of change is accelerating thanks to emerging standard frameworks for interconnectivity
  • MCP ("Model Context Protocol"), put forward by Anthropic, is for making data and executable functions, called "tools", available to AI agents
  • A2A (Agent-to-Agent), put forward by Google, is for making it easy for AI agents to use other AI agents
  • MCP and A2A are not software, they are are open and public frameworks
  • They allow AI agents to be applied in a standard way and have been widely adopted
  • Using MCP and A2A, a chosen LLM decides how to use the tools, data, and other AI agents
  • This is normally initiated by a prompt from a human
  • Third-party data, tools, and AI agents are easily available, comparable with APIs
  • However there are many security, testing and complexity issues
  • The optimum approach is to be MCP- and A2A- ready but not yet to use third-party offerings

BD 24 April 2025

  1. Recognise that Enterprise AI is not like Consumer AI
  2. Recognise that AI Agents matter more than any particular LLM
  3. Compartmentalise enterprise operations, then use specialist AI agents for each
  4. Use historic, often messy, data to create context for AI
  5. Use AI query history to create new context and IP
  6. Capture the value of experienced human-in-the-loop
  7. Prioritise, based on cost and process pain-points
  8. Scale early for production, no need for micro-prototypes
  9. Incorporate into existing workflows, so the AI is actually used
  10. Go for speed plus feedback loops - better than perfection

BD 17 April 2025

  1. Strategic vision misaligned with AI potential
  2. Disconnect between understanding of business and understanding of AI
  3. Response inadequate to impending AI-originated obsolescence
  4. Current AI work does not pragmatically meet actual business needs
  5. Conceiving AI as merely a LLM chatbot, ignoring Agentic AI

BD 10 April 2025

Yes, we understand...
... how AI is set to change the world
... how infrastructure build-out is happening fast
... how LLMs are becoming ever more powerful
... about fine-tuning, RAG, opensource
... about AI Agents
... about using multiple LLMs incl. DeepSeek
... about MCP and A2A as connectivity standard frameworks
... how LLM hallucinations and guardrails are managed
... how long-term agentic memory builds knowledge
... how AI can convert our proprietary data into IP
... how major efficiencies can be achieved
... how innovation is accelerating
Q: BUT HOW DO WE GAIN ADVANTAGE???
A: Through the careful, phased assembly of AI agents

BD 2 April 2025

  • The agentic AI ecosystem is newly mature - only since Jan 2025
  • It is powerful for FCA and other regulatory compliance - and more widely
  • Prior AI preoccupations are now more easily manageable, incl...
  • ...choice of LLMs eg DeepSeek, fine-tuning, prompt engineering, guard rails
  • An emerging heuristic is: 80% cost reduction (= 5x performance) for many corporate functions
  • Agentic AI is relevant to all current projects, incl...
  • ...IT, data, business processes, cost optimisation, skills
  • Agentic AI requires astute management...
  • ...in projects, in strategic direction, in service provision

BD 28 March 2025

  • Autonomy / FSD is not yet feasible, ie humans-in-the-loop are still necessary - regulators will want to see this in their domains anyway
  • The costs of gathering client in-house data can be high - but this becomes client IP once stored in a semantic database
  • New agentic AI systems can be incorporated into workflows using appropriate tools
  • Employees' careers will be radically affected, appropriate policies are needed
  • It is important for directors to have a wide understanding of what is happening with agentic AI, they still have ultimate responsibility
  • The AI ecosystem is evolving fast, staying flexible is important

BD 21 March 2025

  • AI agents are modules that do specialised, well-defined, compartmentalised jobs
  • They have access to data and tools that are designed for the job
  • Where LLMs tend to be amorphous and sprawling, AI agents are focused and specific
  • Within the AI agent, the LLM's job is to make best use of specified information to generate answers to user prompts
  • An AI agent is analogous to a human "agent", but taking seconds to do what a human would do in days - but still fallible
  • Broadly AI agents can be viewed as doing things right, but not necessarily doing the right things: humans-in-the-loop are still needed, but their effectiveness is greatly increased
  • AI agents provide a way around the problem that LLMs themselves are not reliable sources of truth
  • AI agents can use tools which safely make messy private data useful, thus unlocking IP
  • AI agents mean that private data, believed to comprise 90% of the world's information, can safely be made AI-available to the data owner
  • Training of LLMs is expensive and its usefulness may be superseded by AI agents
  • It is important that AI agent design retains the ability to switch between LLMs - different ones are good at different jobs
  • Specified information eg FCA Handbook is supplied to the AI agent in a well defined RAG pipeline
  • The specified information is usually stored in semantic vector or graph databases
  • AI agents radically reduce hallucinations
  • AI agents can include methods to store and learn from history of user AI interactions, and use the history to provide better responses
  • AI agents tend to proliferate, they need to be managed through automated workflow systems

BD 14 March 2025

  • Apply AI to FCA and other regulatory compliance
  • Apply same methods to QA, standards, mandates, guardrails
  • Manageable scale
  • Pragmatic early AI experience
  • Create an early AI trophy
  • Increase compliance efficiency by 5x
  • Reduce friction, increase agility
  • Safely crystallise huge value of proprietary data
  • Leverage lessons into wider AI applications.
  • Retain adaptability for burgeoning AI ecosystem

BD 7 March 2025

  • Powerful for FCA, PRA and other codes
  • A high value, pragmatic AI use-case
  • A method, not a software package
  • No need to use frontier LLMs
  • Can use opensource incl. DeepSeek
  • Option to switch between LLMs
  • Radically enhances human value
  • Minimises LLM fine-tuning
  • Embeds corporate AI guard rails

BD 28 February 2025

Cost. 95% lower cost than comparable LLMs.

Business Landscape. DeepSeek appears set to shift the balance of power away from LLM providers towards LLM applications, software, and users (such as financial institutions).

Quality. Generally, amongst the best performing LLMs.

Architecture. Quality is achieved through elegant use of Reinforcement Learning applied to a large number of advanced and difficult test cases, combined with sophisticated filtering techniques.

Security. If accessed via DeepSeek API, security is sieve-like. If run locally, security is amenable to normal disciplines.

Opensource. DeepSeek is opensource, giving its inner workings exposure to armies of developers and millions of pairs of eyes - arguably the most powerful quality test.

Local Version. Using Ollama or similar, the DeepSeek engine can be downloaded to run locally, and achieve high levels of performance and security.

Agentic RAG. For many purposes the most effective way to apply DeepSeek is as part of regular Agentic RAG pipelines.

LLM Menu. It is usually smart to enable any AI agent to access any model suitable for purpose. DeepSeek can simply be added to the menu of possible options.

Risk Management. The menu approach finesses the need to make once-forever commitments. If doubts emerge, eg over DeepSeek's provenance, the menu option can simply be turned off.

Human-in-the-Loop. Applications which retain humans-in-the-loop, such as Compliance eg with the FCA Handbook, are particularly well placed to consider DeepSeek - the human provides the commonsense test.

BD 21 February 2025