Dear Friends,
I recently submitted my final assignment for the Oxford Foundations of Generative and Agentic AI programme. The result is still pending, but while the learning is fresh, I wanted to capture what stayed with me.
One thing I have gradually accepted:
I will probably never fully catch up with AI.
The technology is simply evolving too quickly. New models, tools and capabilities seem to appear every week.
But perhaps catching up with everything is not the goal.
For those of us who are not technologists, I think there is still enormous value in understanding the foundations, common language and basic principles, so we can enough to ask better questions, communicate with technical teams, recognise opportunities and make more informed decisions about how AI should be used.
That is also one of the reasons I keep this learning journal.
40 and Figuring IT Out is partly for people like me: professionals without an IT background who are still curious about how technology is changing our work and lives.
We do not all need to build the technology.
But increasingly, we need to understand it well enough to work with it, challenge it and make decisions about it.
Module 2 moved from generative AI into agentic AI, and for me the biggest conceptual shift was simple:
AI is moving from answering questions to taking action.
Here are my key takeaways.
1. An AI agent is more than a smarter chatbot
Generative AI largely responds to us.
We provide a prompt, ask a question and receive an answer.
Agentic AI goes further. An agent works towards a goal. It may retrieve information, choose between tools, perform a sequence of activities, observe what happened and adjust its next action.
A simple way to think about it is:
Plan → Act → Observe → Adjust
This iterative reasoning-and-action approach is sometimes described through the ReAct framework: Reasoning + Acting.
The important difference is that AI is no longer only producing content.
It can increasingly participate in executing a workflow.
Imagine the difference between asking AI:
“Draft an email reminding the project team about overdue actions.”
and asking an agent to:
identify the overdue actions, retrieve the responsible parties, draft the messages, send them through an approved process, monitor responses and update the action register.
That is a very different level of involvement.
And once AI begins to act, questions about control, risk and accountability become much more important.
2. The model is only one part of an agent
Another useful lesson was understanding what sits around the underlying AI model.
A practical way for a non-technical professional to understand an agent is to think of the model surrounded by three common capabilities:
Access to information — retrieving current and relevant information rather than relying only on what the model learned during training.
Memory — retaining enough context to operate across a longer interaction or multi-stage task.
Tools and APIs — allowing the agent to interact with real systems such as databases, email, finance platforms or enterprise applications.
This helped me understand why building an effective business agent is very different from simply writing a better prompt.
The model may provide much of the intelligence.
But the surrounding architecture determines what the agent can see, remember and actually do.
3. Before building an agent, ask whether you need one
This may be one of my most practical takeaways.
More sophisticated does not automatically mean better.
A predictable, rules-based activity may still be better handled through traditional automation.
A task requiring some interpretation may only need AI assistance within an existing workflow.
A more autonomous agent becomes useful when the system genuinely needs to interpret changing information, choose between actions or deal with exceptions.
And even then, one agent may be enough.
This module explored multi-agent orchestration—using specialised agents for different activities, such as routing, analysis, execution and review.
It is an interesting idea because it resembles how organisations already divide work between specialist teams.
But more agents also mean more hand-offs, coordination, cost and possible points of failure.
So my sequence would be:
Can ordinary automation solve it?
If not, can AI assist the workflow without controlling it?
If an agent is required, can one agent perform the task reliably?
Only then—does the complexity genuinely justify multiple specialised agents?
The goal should not be to build the most impressive AI architecture.
It should be to build the simplest architecture that reliably delivers the required business outcome.
4. Beware of “agent-washing”
Another phrase from the module that stayed with me was agent-washing.
As agentic AI becomes fashionable, ordinary automation, chatbots and predefined workflows can quickly be re-labelled as “AI agents”.
But the label is less important than the value being created.
For me, an AI solution should pass three tests.
Can it work? — Technical feasibility
Can the system access the right information and tools?
Can it execute the task reliably?
How does it deal with incomplete information, exceptions and failures?
Does it pay? — Economic viability
Does the improvement in speed, capacity, quality or labour actually outweigh:
implementation,
integration,
ongoing model and infrastructure costs,
training,
monitoring,
governance,
and the cost of mistakes?
Should we allow it? — Governance and organisational desirability
What authority are we giving the system?
What happens when it is wrong?
Who can intervene?
Who remains accountable?
And does its use fit the organisation's broader responsibilities to employees, customers and stakeholders?
I like this framework because it moves the discussion away from:
“Can AI do this?”
towards:
“Can it work, does it pay, and should we allow it?”
Those are very different questions.
5. How much autonomy should we give an agent?
This became one of the most relevant questions when I reflected on my own work.
In a large organisation, I can imagine process agents increasingly handling repetitive and standardised work—data preparation, reconciliations, compliance activities, workflow coordination or parts of record-to-report.
Problem-solving agents may also support more senior work, including financial analysis, forecasting, scenario modelling and identifying unusual performance.
But that does not mean every decision should be delegated.
Experienced commercial managers and directors bring business context, stakeholder understanding, professional judgement and accountability.
They ask questions that may not sit neatly inside the available data:
Does this result make commercial sense?
What happened on the project that the system cannot see?
Which assumption is unrealistic?
How might a stakeholder interpret this decision?
Who should we speak to before taking action?
So rather than thinking about agents as either “autonomous” or “not autonomous”, I increasingly prefer graduated autonomy.
My simple rule would be:
The easier an action is to verify and reverse, the more autonomy I am comfortable giving an agent.
If an action is low-risk, bounded, transparent and easily reversed, the agent may be given more freedom.
If the consequences are financially material, legally sensitive, ethically complex or difficult to reverse, human involvement should become much stronger.
That feels more useful than applying one governance model to every AI use case.
6. Transformation economics is more than headcount savings
AI transformation discussions can quickly become conversations about efficiency and labour reduction.
Those benefits are real. But they are only one side of the business case. There is also:
implementation cost,
integration cost,
governance and cybersecurity,
training,
the organisational learning curve,
supervision,
errors during transition,
loss of institutional knowledge,
and disruption caused by changing roles and responsibilities.
Sometimes the short-term cost of introducing AI may even be higher than the labour cost it initially replaces.
That does not necessarily mean the transformation is a bad investment.
It means we need to assess value over both the short and long term.
Does the technology create additional capacity?
Does it improve quality or speed?
Can it scale?
Does it allow people to focus on higher-value activities?
What new risks does it introduce?
And importantly:
What human capability might we accidentally remove while optimising today's process?
For me, this makes gradual transformation particularly important.
The objective should not simply be:
How many roles can we remove?
It should be:
How do we redesign the operating model so that technology and human capability become stronger together?
7. What becomes more valuable when AI becomes cheaper?
The Moravec’s Paradox concept:
Machines can perform some activities that humans historically regarded as intellectually difficult calculation, pattern recognition and certain forms of logical analysis with extraordinary efficiency.
Yet activities that humans perform almost instinctively can remain remarkably difficult for machines: fine physical movement, navigating unpredictable environments, interpreting subtle human behaviour and responding appropriately to another person.
This makes me question the simple prediction that the future workforce will divide into highly paid strategic thinkers at the top and low-paid manual workers at the bottom.
Perhaps something more interesting happens.
As digital intelligence becomes increasingly accessible and potentially more like a commodity some deeply human capabilities may become more valuable, not less. Such as:
Trust.
Credibility.
Physical presence.
Relationships.
Fine motor skills.
Contextual understanding.
Professional judgement.
Accountability.
I am not suggesting that every human-centred job will automatically become highly paid. But I do think scarcity and differentiation may shift.
If AI eventually becomes embedded in almost every workflow and everyone has access to similar digital intelligence, the competitive advantage may increasingly come from what remains difficult to automate.
8. Human oversight only works if humans retain capability
This connects back to something I reflected on in Module 1.
We often talk about keeping a human in the loop. But the phrase can create a false sense of security.
If AI increasingly performs the foundational work through which young professionals traditionally learn, where will future professional judgement come from?
People develop judgement by learning the basics, seeing real cases, investigating exceptions, making mistakes, receiving feedback and gradually understanding not only what happened, but why it happened.
If we automate all of that learning experience away, we may eventually have humans supervising AI who no longer understand the work deeply enough to challenge it.
So human oversight cannot only be a governance control.
It also requires a human learning pipeline.
Organisations will need to deliberately preserve opportunities for people to develop domain knowledge, causal understanding, professional judgement and accountability—even when AI can produce the immediate output more efficiently.
Otherwise, “human in the loop” may eventually become little more than a box we tick.
My Module 2 takeaway
Module 2 helped me see agentic AI less as a futuristic technology and more as an operating-model and transformation question.
I will probably never know every model, framework or technology that appears. And I am becoming increasingly comfortable with that.
What I do want is enough foundational understanding to keep asking useful questions:
Do we actually need an agent?
What level of autonomy is appropriate?
Can it work?
Does it pay?
Should we allow it?
What happens when it fails?
Who remains accountable?
And which human capabilities must we continue to develop rather than automate away?
For those of us without an IT background, perhaps this is one of the most important forms of AI literacy.
We do not need to build every technology we use.
But we increasingly need to understand it well enough to participate confidently in decisions about where it belongs, where it does not, and what we should never stop doing for ourselves.
That is one of the reasons I keep learning.
Not to catch up with everything.
Just to remain curious enough to stay part of the conversation.
Still learning. Still curious. Still figuring it out.
Lydia
See you next time.
P.S. If this letter found you at just the right moment, I’d love to hear about it. Join my mail list and let’s figure it out together — one AI-shaped step at a time. Join the weekly letter list.
☕💌 If you’d like to fuel my next cup of coffee and keep this journey going, you can:
Your support keeps the ideas flowing and the coffee brewing ✨