Transcript

Paul:
Welcome to another Parallel Project Training podcast. I’m Paul Naybour, and with me today is James Gardner. Hello, James.

James:
Hi Paul, how are you?

Paul:
Really good, thank you. Today we are going to talk about AI and how it can be useful for project managers, both now and in the future.

Before we get into that, you are someone who knows a lot about AI. Could you tell us who you are, what you do, and also say something about your podcast, newsletter and other content?

James:
Yes. I mean, you say I’m an expert. I certainly wasn’t an expert, and I’m not sure I would call myself one now. I would say I am someone who is insanely curious about this area.

My background is as a quantity surveyor. I am a chartered QS and a Fellow of the Royal Institution of Chartered Surveyors. I spent most of my career delivering projects, mainly in the education sector. I worked on projects for many universities, including Oxford colleges and Imperial College.

I have always been interested in the future of the industry. Even when I was at university in the late 1990s, my dissertation looked at what the industry might look like in ten years’ time. It was interesting to revisit that recently.

After around twenty years in the industry, I had the opportunity at Gleeds to move into what was then called the Insight and Analytics Department. That gave me a strong grounding in data and allowed me to put some of my interests into practice. Eventually, that became part of a wider digital team.

AI has now become an important part of that work. That includes AI deployment, AI development, AI training, and related activity.

We were experimenting with AI as far back as 2016 or 2017, well before the current generative AI hype. We did some work with Imperial College, where one of the people I was working with was an AI professor. He allowed me to borrow a couple of his undergraduate students, and we built our first machine learning model then.

That opened the floodgates for me in terms of understanding what AI could mean for our industry. It has taken a while for the industry to catch up, but when GPT models started to emerge, even before ChatGPT, it became obvious to me that this was going to be a big deal.

I now have the pleasure of deploying that technology across the organisation. I also do work with RICS, which is a real passion of mine. I want to help make sure the industry moves in the right direction.

For RICS, I have been part of the Construction Professional Group Panel, looking at the future of the industry from a construction point of view. I was also part of the working group for the recently released professional guidance on the responsible use of AI for surveyors.

What I think is important is that I also understand projects. I have worked in projects for many years and understand the pain points involved in delivering them.

Paul:
And what is Project Flux?

James:
Project Flux is my labour of love. It started as a newsletter because I realised I was spending my weekends reading about AI just to keep up with its rapid development. I thought I might as well turn that reading into a newsletter to help other people understand what was going on.

It started on LinkedIn and gained a lot of interest. It has now become quite a well-read newsletter in the industry, with around 6,000 subscribers.

We then started doing podcasts as well. We interview authors, academics, practitioners and people from the technology side. You can find it at projectflux.ai. It is free to subscribe, and the aim is simply to keep people informed.

Paul:
So you are at the cutting edge.

What we thought we would talk about today is what project managers can do with AI day to day, but also what might happen in the future.

Before the podcast, we had a discussion and grouped this into three layers.

The first layer is the basic use of AI, which we called knowledge assistance. This is where AI is used as a kind of intelligent encyclopaedia. You load it with project management knowledge, guidance, textbooks and in-house procedures, and it helps people understand what the guidance says.

The second layer is decision support. This is where AI has access to corporate systems, such as planning, finance, drawings and project status information. It can then act as an intelligent query tool or virtual dashboard, helping people understand what is going on in a project without always having to phone people or ask around.

The third layer is organisational learning. This works more at portfolio level. It is about looking for trends and patterns in the data. We know projects often fail for familiar reasons. Perhaps AI can help us understand those patterns and identify warning signs earlier. That includes things like benchmarking, real-time benchmarking, lessons learned and double-loop learning.

We thought this structure would be useful because AI is a huge subject, and people interpret it in different ways.

James:
They do. Most people think of AI as ChatGPT because that is their main exposure to it. But AI has been around for a long time. The term was coined back in the 1950s.

There is a lot of confusion, and many people are still trying to make sense of it. So I think it is helpful to clarify what we mean.

Paul:
Let’s start with level one: knowledge assistance. Have you seen AI used in that way, and what does it mean to you?

James:
That is where most people are at the moment. When people first get access to ChatGPT or Copilot, they often use it as a glorified Google.

Interestingly, there was a study recently suggesting that click-throughs from Google search results have fallen significantly. That matters because Google’s advertising model relies on people clicking through to links. But now people often get an AI overview directly from the search engine, powered by Google’s Gemini model.

So people are already getting used to writing a query and receiving an answer directly, rather than scrolling through pages of links.

Paul:
I have seen that with clients. Some organisations have large libraries of procedures, processes, standards and rules. They want people to access that information quickly, so they use AI as a better search tool to find answers.

James:
Yes, but there are a few issues with that.

First, we have trained ourselves over the last two or three decades to use Google through keywords. With Google, you tried to shorten your search to a few key words to find the right result.

That is not the best strategy with AI. AI needs more context. It needs to be told what to do. The more useful context you provide, the better the answer is likely to be.

If you give it very little information, it is more likely to hallucinate because it still wants to give you an answer. If you do not provide enough context, it may fill in the gaps.

That is where people sometimes have a bad experience. They use AI like Google, expect a better result, and sometimes get a worse result.

It requires a different mindset and a different kind of training.

The other issue is that we are quite lazy. In the past, if we went to a library, we had to find the right book and check sources. With Google, at least we still had to look at the links and make a judgement about whether they were reliable.

Now we can get an answer immediately, and we do not always check the citations or sources. That creates a risk, particularly where the AI makes things up.

So level one is where most people start, but many people get stuck there because they are not using the tools in the right way.

Paul:
I think this has strong applications where the body of knowledge is constrained. For example, we have used it with exams that are based on a set text.

You can define the source material for the model and say, “This is the text you should use to answer questions.” You can also give it sample questions and define the sort of response you want.

The user may enter a very short prompt, but you can design system instructions in the background that expand that into a more complete request.

James:
Exactly. People are effectively placing their own body of knowledge on top of the model.

You have the capability of the model, whether that is GPT, Claude or another system, and then you layer your own information on top. That directs the model towards your information first.

It is a common way of getting started. It is also fairly easy to do now. You can do it personally in Copilot or similar tools by creating a project and putting information into that project folder. You can also create a small agent within Copilot.

That is a good first application.

Paul:
And nowadays you can attach a folder on Google Drive or SharePoint and tell it that this is the content you want it to search through.

James:
Yes. It is getting easier and easier to personalise.

Paul:
The level of hallucination also seems to be going down. AI is getting better at using sources when you provide them.

James:
It is. That is partly because context windows are getting larger. It is also because the models are becoming more powerful and better at checking their own work.

We are also getting better at using them. We are learning how to prompt them more effectively.

That said, there are still issues. People need to understand that. I do not like the phrase “human in the loop”, because it implies that we are somehow below the AI. But the principle is important. We still need human judgement. We should not simply pass off AI-generated information as our own without checking it.

Paul:
Before we move on, you mentioned the context window. Could you explain what that is?

James:
The context window is the amount of information the model can take into account at one time.

When these models first emerged, you could only put a limited amount of information into them. You could not easily upload documents, for example.

These models work using tokens. Very roughly, a token is around three-quarters of a word. So when you hear that a model has a context window of one million tokens, that means it can work with a very large amount of text.

The bigger the context window, the more information the model can synthesise and understand. That usually improves the quality of the response and reduces the likelihood of hallucination.

The models have also become better at understanding attachments, images, tables and PDFs.

As a rule of thumb, a bigger context window usually leads to better performance.

Paul:
Let’s move to level two, which we called decision support.

This is where AI becomes a top layer, but you give it access to tools or functions. It can call business systems and bring in costs, schedules, drawings and dashboards.

There are a couple of concepts here. One is an API, which is an established way of pulling information from one system into another. Another is MCP, or Model Context Protocol, which allows the AI to access tools. For example, it could call a cost system, a risk register, or a project status tool.

Have you seen people doing this?

James:
Yes. MCP is a big deal. The best way to think about MCP is like a USB-C connection for your AI.

When USB-C came out, it gave us a standard way to connect hardware. MCP is a standard way for AI to connect to software. It allows the AI to talk to different systems.

If software such as Microsoft tools, Procore or another platform has an MCP available, it can plug into your AI environment. The AI then shifts from something that talks about your project to something that can work inside your project.

You can still connect systems using APIs, but MCP makes it much easier. It can appear as a switch in tools such as ChatGPT, Copilot or Claude. You turn it on, and the AI has access to the relevant system.

Personally, I use Notion to keep my notes. There is an MCP for that, so Claude can connect to my Notion workspace. I have similar connections to email and calendar. That is a massive change because the AI becomes more agentic. It is no longer just giving responses; it is working inside the tools.

I think it will not be long before it is unusual for a software tool not to have an MCP.

Paul:
We have seen this used with risk registers, change logs, issue logs, Gantt charts and the general data project managers produce.

James:
Yes. Take Procore, for example. If your project data is in the Procore environment, and Procore has an MCP, you can connect that data to Claude, Copilot or another AI system. The data becomes interoperable.

Another simple example is email. This is where you start to see a separation between people using AI as a glorified search engine and people using it in a more advanced way.

The people using AI in a more advanced way are giving themselves agentic capability. AI is doing tasks for them. That is not a 10% productivity improvement. In some cases, it may be 50% or even 100%, because they effectively have agents working for them.

For example, I used AI to help prepare for this podcast. We had a previous call to discuss the topic, and I had a transcript and notes in Notion. Claude could connect to those notes and help me prepare.

For project managers, who often run from meeting to meeting, this is very practical. AI can prepare briefing notes, summarise previous discussions and help you walk into meetings better informed.

Paul:
For portfolio managers and programme managers, it is also powerful. You could ask, “What are the top ten risks across my portfolio?” or “Which changes are running late?”

When I used to run a PMO, directors would ask questions like that, and I would spend two days building pivot tables. AI can pull that data together much more quickly.

James:
It can. But the key point is that you still make the judgement.

The AI takes away the grunt work. It helps you pull the data together, but you still need to decide what it means and what to do.

You can also use AI as a thought partner. That is something people often miss. They see it as a content creator, for example, to write emails to sponsors. But it can also help you think through scenarios.

Project managers spend a lot of time dealing with people, difficult conversations, negotiations and disputes. AI can help you prepare for those situations. That use case is still underused.

Paul:
Yes. If you link the contract documentation to the project status information, then in an NEC contract, for example, you could pull in compensation events and ask what the relevant clauses are, what the real issues are, and how you might approach the discussion.

James:
Exactly. And you can take that further. If you have a dispute with a contractor and a meeting next week, you can ask the AI to role-play what the contractor might say. You can use it to rehearse the conversation.

I have also created my own board of advisers. I got the idea from Jeff Woods, who wrote The AI-Driven Leader. The idea is that, with AI, you can create a simulated board of advisers.

You can choose people, living or dead, who you would like advice from. For example, you might ask Warren Buffett for financial advice or Steve Jobs for advice on innovation. You can create that as a reusable skill and call on it when you have a problem.

Paul:
And Steve Jobs says, “Just go for it.”

James:
Actually, the thing I get most often from my Steve Jobs caricature is, “Say no to more things.” He was famous for returning to Apple and cutting most of the product line so the company could focus.

I would encourage people to try this. It is a brilliant way to use AI for more than just search.

Paul:
That is interesting. One of the exams we support includes an interview-based assessment. People really struggle with the role play. It is hard to practise because most people do not often experience a technical, case-study-based one-to-one interview.

We trained a GPT to simulate those interviews. It was great practice. I would not say it was perfect, and I would not say the grading was always right, but it was very useful for asking questions and giving feedback.

James:
Absolutely. It is a brilliant use case.

Paul:
Let’s move to level three, because we are running over time.

Level three is about looking for trends. If you have collected data across a portfolio of projects, and you have data from the start to the end of projects, AI can help analyse that.

As humans, we are not very good at lessons learned. We often write them at the end, put them in a drawer and never look at them again. Or we write them with an optimistic gloss. We say, “It was always going to be hard, but we did well.” That makes it difficult to capture the real lessons.

So our level three is what we might call decision guidance or organisational learning. This is where AI performs a neutral analysis of patterns, shapes and trends in the data and communications. It can help identify the key lessons, themes and decision points.

For example, what did people talk about most? Was the main issue stakeholders? Scope? Ambiguity? Contractual issues?

You are essentially building a model of the conversations within a project to understand where the main conflict points were. I find that exciting because it is difficult for people to see these patterns manually.

James:
We are not good at synthesising large volumes of data. Projects generate millions of data points, and it is easy for patterns to be buried.

This is where it is important to distinguish between types of AI. People think of AI as ChatGPT, but ChatGPT and similar tools are generative AI. They are probabilistic. If you put the same prompt into ChatGPT twice, you may get two different answers. That is how the system is designed.

By contrast, more traditional analytical AI and machine learning models can be more deterministic. If you put in two plus two, you expect four every time.

Increasingly, tools will know when to use which approach. If you are dealing with numbers and large amounts of numerical data, you may want a more deterministic system. But much of the important project data is not numerical. It is in emails, reports, meeting notes and sentiment.

Generative AI is very good at detecting patterns in that kind of language. It can identify sentiment and changes in the way people are talking. When you combine that with numerical data, you can get much richer insight into what is happening on a project.

Paul:
I tried to explain this to someone recently. AI takes written information and converts it into numbers, or vectors. It is like looking at the night sky. You can see clusters and formations, and you can see things that are related. The challenge is then how to interpret that data.

James:
That is exactly right. These models convert words into numbers and then work out the relationships between them. That is how they understand language.

It sounds strange when you ask how something that converts words into numbers can sound so human, but in some ways that is similar to how the brain works. Neural networks were designed to replicate aspects of how neurons work.

It is incredibly powerful, but we do not fully understand how it works. We understand the input and the output. The difficult part is the billions of parameters in the middle. We do not fully know what is happening there, and that is a little scary.

Paul:
I have experimented with embeddings and presenting them as a two-dimensional map, often called a UMAP. It lets you look for clusters in the data.

I was hoping to use diary data, but the project was shelved because nobody had written the diary entries. I still think it would be very useful at portfolio level.

The problem with lessons learned is that your perception and my perception of the project might be different. But if we look across a whole portfolio, we may start to see patterns and shapes.

James:
It is very interesting. People’s definitions of success also vary. Another issue with lessons learned is that they capture a point in time. Then everyone moves on to the next project and repeats the same mistakes.

What I find exciting is the idea of recursive self-improvement, where models improve as they work. For our industry, that could be powerful because we would not have to wait for a lessons learned report at the end of each project. The model would keep improving as it goes.

Paul:
And it could learn not just from one project, but from similar projects with similar characteristics.

James:
Exactly. Every project is unique as a whole, but if you break it down into constituent parts, you will find similar elements in other projects.

If we can capture that learning into broader models, everyone can benefit. Society can benefit too, if we get this right.

Paul:
We should briefly talk about risks and concerns. We are both enthusiastic about AI, but there are risks. What should people be worried about?

James:
I am an evangelist, but I am also very aware of the risks.

The biggest risk, in my view, is loss of control. AI is moving incredibly fast, and we are in an arms race between the major developers. New models are being pushed out all the time.

There are concerns about safety, cyber risks and the speed of release. Some models or capabilities may be held back temporarily so organisations have time to patch vulnerabilities.

The broader concern is that the companies building these systems are becoming extremely powerful. They are almost like empires. Yet legislation and governance are struggling to keep up.

In the UK, we do not have one comprehensive AI Act in the same way as the EU. We are also affected by the EU AI Act because we work with organisations in Europe. We still have GDPR, and there are OECD principles and different approaches in Canada, China and individual US states.

But there is no consistent global consensus on how to manage AI. That creates gaps.

My worry is that there may be a major incident that forces everyone to slow down and take stock. We are already seeing prominent figures in AI raising concerns. Geoffrey Hinton, for example, left Google so he could speak more freely about the risks.

Paul:
It reminds me of when I graduated. I studied electronic engineering, including microprocessor coding and microprocessor-based systems. Microprocessors were new then, and the technology accelerated very quickly.

Within a decade, we went from very little to IBM PCs, desktop computers and eventually microprocessors in almost everything, even washing machines.

James:
In some ways this is not new. You can go back to the printing press, the loom, or even the introduction of handwriting. Socrates was apparently worried that writing would damage memory.

Technological transitions have always caused concern. What is different now is the speed. This is moving at a pace that is very difficult to manage.

Paul:
I am sure people said something similar when the steam engine came out.

James:
Yes, maybe. Perhaps people will look back and wonder what we were so worried about. I hope so.

Paul:
But you are right about the risk of accidents.

James:
Aviation is a good example. When powered flight was invented, there was no comprehensive legislation. It often takes accidents or major incidents before society decides universal rules are needed.

Paul:
If people find this podcast useful, where should they go for more information?

James:
RICS has professional guidance on the responsible use of AI, which people can download for free.

There is also Project Flux, my newsletter, at projectflux.ai. It is free to subscribe.

I also do work with APM, and people can connect with me on LinkedIn to follow what I am doing in the industry.

Paul:
There is also a very good series from IBM on YouTube, with whiteboard sessions explaining the technology. I found those helpful because they explain terms such as context window and RAG. There is a lot of jargon in this area.

James:
There is. But there is no shortage of material. IBM, Google, OpenAI and others have all produced useful courses and videos. YouTube is your friend.

I always say people should take responsibility for their own AI literacy. There is no excuse really, because much of the material is free.

Paul:
And if anyone wants to give me some data, I am very interested in looking for patterns.

James:
There you go.

Paul:
Thank you very much for your time. I hope this has been useful. I think you want me to reciprocate and do a podcast for your channel.

James:
Yes, I would love you to. I will send the invite and we can return the favour.

Paul:
Excellent. I wish you all the best with this new AI revolution. It is going to be fascinating.

Do you think projects will still be overspent and late in ten or twenty years’ time?

James:
That is a good question. There will always be humans involved, and there are things AI will not fully replicate, such as negotiation, emotional intelligence and relationships.

I would be disappointed if we are not much more successful at delivering projects. The fundamental issue is that we need to deliver more projects, and deliver them better.

I am optimistic, but five years is a long time to look ahead in the current climate.

Paul:
Do you think the pharaohs were told their pyramids would be built on time? That is the question.

James:
I am sure they could put us to shame with some of the things they achieved. Or maybe they were part of an advanced society. There are always rumours about that. Maybe they had AI. Who knows?

Paul:
Maybe it has come full circle.

James:
Maybe it has.

Paul:
Brilliant. Thank you very much.

James:
Thank you, Paul.