The Engineering Consultancy Productivity Challenge

When I first started in the world of engineering consultancy, in the mid-1990s, hourly selling rates sat in a range between £50 and maybe £120 for a top principal.  Designers and grads were generally at the lower end and the more technical things got, the higher the rate.

Thirty years later and the rates are… exactly the same (in the UK – very different in other countries, especially the USA).  There are various reasons for this, including offshore competition and subtle changes in working practices, and that can give the impression that productivity has improved; after all, clients are getting the same work but paying less in real terms.

This is a paradox. Productivity has not improved.  The business model is unchanged – you quote the work, do the work, check the work, and deliver the work in almost exactly the same way as you did thirty years ago.  There are small improvements but nothing like the radical changes caused by automation in manufacturing and other sectors, and there are graduates available - the UK produces more STEM graduates per head than any other European nation - but we're short of experienced seniors.  A consultancy company uses its accumulated store of human knowledge and experience, and there has been no substitute for that.

Until now, perhaps.  Might AI finally provide the equivalent of industrial robots to the consultancy industry? I need to be clear here about what I mean: Machine Learning (ML) has been part of engineering analysis for many years and I was coding neural nets to replicate non-linear control systems in the early 90s.  But can the combination of ML with Large Language Models (LLMs) could finally be the Big Leap Forward.

Naturally, the big software companies are working on it so it’s coming, and there are advantages – it shouldn’t take long for relatively simple tasks like reporting, proposal writing, and data analysis to be automated.  However, the disadvantages tend to attract most attention: don't LLMs make things up?

Public AI tools hallucinate, and that’s because they’re trained and exposed to the whole internet – all of it, and we all know how reliable that is.  Public tools struggle to discriminate between website that say the earth is flat and the NASA database.  Private AI instances, trained on your company internal data only, are the way to go, but even then they need a LOT of data to work well.  I tried to ask an AI to write me an OpenFOAM input deck and… well, it didn’t end well. 

But it will be possible, and with the big software companies pushing, it will happen.  Yes, there will be issues with quality assurance and repeatability, but they can and will be solved.  A private AI won’t struggle with IP issues, won’t be a security risk, and will continue learning faster than its human counterparts.

According to Dell’s Lee Larter, speaking at the recent IMechE Advanced Nuclear Reactor Design Conference, the problem is going to be power.  Under the UK government’s AI Opportunities Action Plan, Dell is already working with Rolls-Royce and others on how AI will accelerate the design process. The route will be via Digital Twins, taking existing models as a template and gradually allowing AI to use them as a basis for design changes.

However, the latest NVIDIA GPU consumes 1.4 W, and in the recent US-UK collaboration agreement a 120,000 GPU cluster was announced, consuming the entire power output of an entire SMR nuclear reactor.  How many will the UK need to remain competitive, in a world where suddenly engineering productivity isn’t driven by the size of the skilled workforce, but the number of chips in a nation’s AI clusters? 

There is some hope.  In this excellent article, Ambrose Evans Pritchard describes developments in graphene and conventional super-cooled chips that should cut energy demand for AI clusters by 90% by the early 2030s.  The UK is a world leader in graphene chips, although we should be wary of or historically poor performance in converting raw intellectual property into revenue.  New AI large language model algorithms use less energy, and perhaps the AIs themselves can suggest solutions.  Crucially, the UK ranks third in the world in AI intellectual property, after the USA and China, and its universities remain world-class.

Technology has a habit of making fools of us all, and I’m not going to predict when these challenges will be solved, or how.  But I am confident that they will be solved, leaving the engineer with less to do.  Whether this liberates engineers to do the one thing that they can do, and AIs can’t – imagine – is yet to be seen.  But it thousands of proficient engineers were suddenly free to do nothing but invent, what might happen?

I bet rates won’t go up.

 

About the Author

Dr Simon Rees CEng FIMechE MRINA has over thirty years of experience in engineering consultancy, with roles up to Managing Director level.  He has been involved in both acquiring businesses and being acquired, including managing the subsequent integration activities.  He is a director of Ithuriel Ltd and can be contacted on simon.rees@ithuriel.co.uk .