Enterprise Challenges to adopting GenAI

While various forms of Artificial Intelligence(AI) have become embedded in business processes over the past decade, the world has turned its attention to one particular branch of the technology, Generative AI(GenAI). ChatGPT and other large language models (LLMs) are the foundation of many exciting, innovative solutions popping up in all corners of the enterprise, but is your organization actually ready to drive value from adopting these new capabilities? In many cases, probably not. Today, lets talk about why I say that and cover the three main obstacles for bringing AI to the enterprise.

Trust

I'd argue the most pervasive hurdle GenAI solutions need to overcome is that of trust. Leaving aside the hype that AI is coming to replace all of our jobs, there are still plenty of anxieties surrounding GenAI capabilities. This lack of trust stems from:

Probabilistic Nature of GenAI

• When you type 1+1 into a calculator, you expect to get 2 as the answer every single time. That is a 'Deterministic' behavior. The machine is programmed to run the same operation and give you the same result every time. Gen AI doesn't do that. Its 'Probabilistic.' There are a range of right answers and the one you get this time may be different than then one you got last time. Not to mention, hosted models are constantly being tweaked. Usually, if you ask for 1+1 you'll get 2, but sometimes you'll get 11, sometimes 3. Large Language Models (LLMs) in particular are crawling through all the information they can find, and if a joke or wrong answer pops up enough, it might sway the results you get out of the model.

Mistakes at scale

• GenAI in the enterprise is still a new technology, and like every new technology, there are heaps of horror stories. Chatbots being tricked into bad mouthing the company they represent is bad, but what happens when the AI figures out giving away money pumps up customer sentiment and inappropriately promises discounts and refunds? Air Canada found out, its liable for statements made by its chatbot, even if it goes against their policy. When a person makes a mistake, it can be fairly contained. If an AI makes that same mistake, it scales substantially.

Change Resistance

• Some of your people prefer the stability of a known process even if something new could make their work substantially easier. Others are always hopping to the next big thing in hopes they can gain an edge. If your organization leans towards stability, any change, Gen AI adoption included, will be met with resistance and its going to take time and exposure to this new way of working to break it down.

Data

Many organizations suffer from fragmented, un-managed, or non-existent data. AI needs data to perform, and in particular, it needs your organization's data to specialize its output for your company. Models trained on generic, broadly available, data are putting on incredible demos, but if you want them to be effective in your specific use cases, they will need your data. The lack of effective data strategy, organization, and readiness pose a massive roadblock to bringing GenAI into the enterprise.

Additionally, the concern around data is paired with the concern around trust. Who owns the model, where is your data being moved to and what is being stored? If you limit what data the AI gets access to, and you should to a certain extent, that will in turn place limits on what capabilities you can expect the AI to deliver. As one example, if it can't touch PCI(payment card industry) data, it can't handle some payments or end-to-end sales transactions. Healthcare has another class of data to worry about PHI(Protected Health Information). While running AI and processing data locally is one way of letting organizations leverage existing policies to manage this new tech, you may be limited in what models you can run and take on substantial infrastructure costs to maintain compliance.

Integrations

LLMs on their own have disruptive, but limited, uses; think content generation. Their real power will come from action and that has to come from integrating with other systems and platforms to get work done. If you see the phrase 'Agentic AI', you are seeing an AI intended to work outside of its own interface, an AI with a job. Just as before, this requires a level of trust and autonomy for a new technology that works in a fundamentally different way than any program you've worked with before. Aside from the trust concern, legacy systems just may not be capable of integrating with this new technology. Your original developers may no longer be with the company, the integrations may be overly complex and burdensome, or there aren't programmatic interfaces to link the new technology up to. If your existing systems aren't ready, the value you derive from new GenAI capabilities will be inherently limited.

GenAI is poised to re-invent work as we know it, but there are some large obstacles that enterprises need to contend with to bring it to life. Creating trust in something so new and different is paramount. Tackling messy data and integrating with legacy technologies can also pose a challenge. Still, these are solvable given the time, resources, and dedication.

Have I missed any of the big hurdles you've seen? Have you broken through any of these? If you have, please share your experience!