In 2015, when I was given the chance to lead a services portfolio that included AI, I was quite excited for two reasons. One – it perfectly suited my geeky interests (especially my passion for math) and two – it seemed to me that I was back in 2000 with ERP, with every company wanting to do something significant with AI . Four years of working closely with assorted clients on AI projects, building teams, and getting asked similar questions all the time – I think it is probably time to step back and share a few learnings from the initiatives that went exceedingly well and those that did not get anywhere.
To keep the blog to a reasonable length, I am going to make some groupings and generalizations and I do have a worry that some nuance will get lost as a result. So if something needs to be challenged – pls do so in the comments, and I am happy to explain what is on my mind in some more detail.
While there is a lot of debate about AGI, potential job losses and so on – I don’t think that is what is hurting adoption at the moment in commercial companies. So I am ignoring a discussion on those kinds of topics here.
As always – all this is strictly my own personal opinions, and not that of my employer.
1. Adoption has not exactly matched what ERP did in the 2000s
AI adoption no doubt has increased significantly – but certainly nowhere near what we saw when ERP exploded twenty years ago. There are many reasons – fragmented market, lot more hype and scare tactics thanks to social media being a thing now, AI initiatives being considered a science fiction type project etc. Also, I think the academic influence on this field is a bit of a double edged sword. On one hand – its a young field in many ways and it needs a lot of R&D. On the other hand – we also need a strong focus on applications and not just the theory. Till that balance happens – I don’t think adoption will catch fire. I also think going after low hanging fruit alone without a road map of how AI will influence the business has played a part in adoption not being where we all expected it to be.
2. “Fire and forget” does not work at all
In a large part because of ignorance, a lot of initiatives did not understand that as data changes over time – so will the quality of output of that AI makes out of it. So you will see great results and declare a Proof of Concept a success, and then a while later you start wondering why the decisions look so bad. A lot of attention need to be given to the life cycle aspects and only a minority of projects do so now.
3. Focus on business is key
There are plenty of courses on AI and they are inexpensive and generally of high quality. And yet – awareness of what problems AI can actually solve is still very low amongst business and IT executives. Pardon me for saying this – but anyone who says things like “use AI to re-imagine xyz” should be stopped and questioned on specifics before allowing to continue their pitches. We should be way past the point where we should be talking at such a high level . The conversation should be about what unsolved problems can we solve now with AI, what extra benefits can the business get with an approach that uses AI to already solved problems etc.
4. Focus on IT is ALSO key
AI is not all business oriented either – it needs a LOT of IT aspects to be taken care of if you need it to be widely adopted in a company. This is the lack of nuance that gets me worked up with a lot of commentary on the topic . To make it sound fancy, a lot of people say things like “AI is not an IT initiative, its a business initiative”. Newsflash – they are plain wrong ! . It is both and it needs different things to fall in place for business and IT. A lot of care and thought needs to go into making sure that you have a process (and tech components and people) in place to manage the whole life cycle of AI in a company – and it is non-trivial. It is absolutely possible to to do a bunch of pilot projects without such a platform approach – but it will get to a tipping point very fast where the lack of a platform approach will hurt.
5. No Good Data, No Good AI
This is true for all things in IT – but I have seen that AI will expose your lack of discipline in data much more than most other initiatives. You can create very smart solutions in a POC mode with AI techniques and they may solve very complex problems. Then when you decide to deploy at scale, you realize that your data is not in a form that can allow for an enterprise wide deployment . I have lost count of how many times I have come across this even for departmental level roll outs. It was no different in late 90s with ERP – a good part of the coding I have done in my life was to write transformation routines to fit data into something ERP could use. It’s astonishing that even today we have not largely solved the problem with data. Pro tip – include budget to get data into shape when you estimate AI projects !
6. Quality of service
There is a whole array of things like performance, security, ethics, CI/CD etc that get ignored in POCs that get declared as a success. And then when we try to take it into production – it hits a sequence of walls and AI gets labelled as “only good for POC”. While it might be too early to say “best practices” – there are sensible tried and true approaches to be used with AI that you should consider from the start and include in the scope with time and budget. If it does to get into production, what is the whole point anyway ?
7. API, custom build, commercial vs open source platforms etc
Like it or not – almost every company will end up with a mix of all these while going through their AI projects. Open source based custom builds will probably be the largest component eventually – but there are so many good reusable commercial products around now that it makes very little sense to custom build everything from scratch. Companies that have no ground rules established on how to approach this typically end up wasting a lot of time and money.
8. Developers have a big role, along with data scientists
Thanks to a lot of dashboard based demos – a lot of companies start on AI thinking that all they need is a team of data scientists and may be some visualization experts. The reality is that high value usually comes from integrating AI functionality into applications that run the business. That needs API work, data wrangling , DevOps integration etc which need engineers . I know several projects that under estimated this in a big way which led to stoppage of work till they could find more budget.
9. Ecosystem of talent
AI is the poster child of the war on talent. It is hard to source, recruit and retain talent in this field. A big part of the planning for large scale AI work needs to consider the risk of finding and keeping the right talent. This will need a combination of working with academic institutions, consulting companies, free lancers, recruiting and training in-house talent and so on. Its a LOT of work and generally under estimated by orders of magnitude. You generally don’t need an army of people like what we needed when ERP became hot. But you need high quality – and everyone else needs them too. And since AI generally needs more care and feeding throughout its life cycle – thats an added layer of thoughtfulness that needs to happen !
10. Education
I cannot over emphasize the need for training and education – and a lot of cross pollination across the organizations. There are a LOT of different ways of solving problems with AI techniques. If the team is not aware of it, and can’t debate and try out – you probably won’t maximize the effectiveness of what AI can do for you. Similarly it needs a bit of education to understand how to interpret what AI tells you. I often do a level set on basic ideas of curve fitting, probability etc for my clients to make sure they are not misinterpreting me.
11. Change Management
I have lived through the ups and downs on how change management is perceived by my clients from my ERP days. It used to be the first thing to be cut off the budget and almost in every single case there was a big price to pay. AI – in many cases – is significantly disruptive for how business and IT operates daily. And when the humans involved in it are not proactively prepared for change – the resistance/fear/anger etc usually leads to suboptimal use of AI and sometimes total failure. In any case the top-down approach from ERP days is not useful in AI context – simply because the whole process is a series of experiments and not very deterministic like an SAP implementation.
12. Communication
In almost every other context I would have clubbed it under change management. But comms needs a special mention when it comes to AI. Within the team, and across the org – there is always uncertainty about AI and the best way to get things done, and what benefits are expected and so on. It’s just the nature of the beast. If you don’t spend the time and effort ( and money ) to communicate clearly – it can single handedly make it impossible to have a successful AI initiative.
And Vijay Says... 