18 Roadblocks To AI Adoption — Exclusive Surveys & Exec Interviews
At a recent GAI Insights learning lab conducted by the amazing Rosemary Brisco , the topic of discussion was the obstacles and roadblocks that companies are facing while trying to deploy Generative AI in their enterprises.
The key takeaway: Most companies are NOT prepared to adopt AI, according to two recent surveys.
And while AI can enhance operations, several obstacles can hinder progress. Some of these obstacles make sense — and some are outright shocking. Let’s unpack the data.
I’ll dive straight into my interpretation (hopefully in Plain English!) of the two surveys and other roadblocks I am observing in talks with customers and executives.
If you’d like, you can skip straight to the technical aspects of the surveys at the bottom (my feelings won’t be hurt — LOL!)
Roadblock 1: Getting Buy-In From The Boss
A lot of companies are trying to kick off AI projects, while the CEO and Board are running around like headless chickens. Every day, the news is flooded with AI stories, and the CEO and Board are sometimes clueless.
In surveys with CEOs, the №1 complaint is: “There are so many options and everything is so confusing, we don’t know where to start”
Until there is universal buy-in from the boss, the AI project is bound to get stuck at the slightest impediment (and trust me, there will be impediments!)
Solution: Put together a simple zero-risk, low-cost AI pilot to get the boss’ feet wet.
Roadblock 2: Boiling The Ocean
With all the glorious stories about how AI will solve world hunger, some companies are trying to literally boil the ocean.
Instead of starting with simple goals using the “Crawl, Walk, Run, Fly” rule, they are trying to kick off insane workflows that are far beyond current AI capabilities.
It almost feels like people are mad about the work that they have to do and then suddenly wake up and say “I wish AI could do this so that I can go home at 4PM”.
Solution: Start with a simple AI project with a clearly defined problem and desired outcome. It does not have to be rocket science. Remember: This is the start of a long journey.
Roadblock 3: Getting Buy-In From All Teams
This one is tricky. Every company has its AI proponents and opponents.
Never has there been a technology like this that can evoke extremely strong emotions on both sides.
Some people think that AI will solve all the world’s problems.
And some people think that it will be the end of the world.
Your company is no different.
Bob in security OR Joanne in accounting is waiting to kill your AI project. Both of them are armed with that New York Times article they read some time ago.
Solution: Keep the project virtually zero-risk. Start with low-risk public data and low-risk use cases. Don’t ruffle feathers or promote the AI project as “We are going to cut 30% of our staff”
Roadblock 4: Getting Your Data Ready
A common pushback from AI opponents is: “Will our data go to OpenAI and become part of their next AI model?”
While it is pretty clear by now that using the OpenAI API means that it does NOT train on your data, clueless opponents will still bring that up and spread the FUD (Fear, Uncertainty, Doubt!)
And no amount of convincing will calm them down. So what do you do?
You start with publicly available data like your website, helpdesk, low-risk PDFs and Youtube videos. It’s so much easier to tell the opponents “Dude, Google already has this data” — than to try and convince them about RAG pipelines and Machine Learning algorithms.
This data is available to the general public and has no restrictions on accessibility or usage. Examples include public government records, published research papers, and openly available datasets.
If you are the adventurous type, you can upgrade to using Internal data. This is data that is intended for use within an organization but is not sensitive or confidential. This might include internal newsletters, policy manuals, or training materials. While not openly available to the public, this data typically doesn’t require stringent security measures.
Solution: Start simple with publicly-available data or non-sensitive internal data.
Roadblock 5: Our Data Is Not Clean
One common misconception I usually hear is: “Our data is not clean, and I am not sure what format you need the data in. Do I have to create structured JSON and XML files so that the AI will understand it?”
A few years ago, this used to be a very valid concern. AI would be trained with clearly marked structured data and the first step was “Wait — let’s get the data ready”.
However, one of the biggest unheralded superpowers of LLMs like GPT-4 is that they are remarkably good at understanding unstructured data.
Please note: I said “remarkably good” — not perfect. It is quite stunning to see how well GPT-4 is able to understand unstructured data that is passed in prompts (e.g., RAG)
Solution: Don’t worry about data cleanliness UNTIL it is shown to be an issue in the testing.
Roadblock 6: We Didn’t Budget For This
Now, this is a valid concern because this AI thing came out of nowhere. In fact, a common refrain I heard in 2023 was:
“We made our budgets in 2022 and didn’t expect this AI thing to show up”.
OK — makes sense — even though AI is an existential crisis for a lot of companies. Last week, I heard a customer say “We have to adapt to this despite the budget because otherwise, we will cease to exist”
Now not all companies have disposable budgets lying around — so what do you do?
You start with low-cost initiatives first.
Start with cheap on-ramps like ChatGPT For Teams ($30/month per user) or low-cost no-code systems that let you create AI pilots with ease, without using developers.
By avoiding expensive development projects (Yes — ML engineers do cost a lot!) — you get to pilot AI with some simple use cases and workflows without breaking the bank.
I’m not talking about tens of thousands of dollars here. You can create an AI pilot for less than $1000 (using DIY no-code systems)
Solution: Start with no-code systems that cost less than $1000
Roadblock 7: We Don’t Have The Time
Remember the time when AI projects would take 6 months to develop and run?
Companies would assemble teams of AI developers and then create product roadmaps for the next 6 months — treating a simple AI pilot like it’s a mega project.
Unfortunately, some companies are still in that mindset. They think an AI project is supposed to take 6 months to build and test — and then they enter a “freeze” state.
News flash: With new no-code and low-code systems, you can get your AI pilot up and running in a few days (if not hours!)
Don’t believe me? This AI pilot for a rare disease foundation was created in a few hours .
In fact, Mary in HR itself might be able to get your AI pilot running in a few days (I kid you not — I have proof!)
Worst case: Hire a Python developer on Fiverr or Upwork and get him to build a simple wrapper workflow for you with the OpenAI API.
Ninja Tip: You don’t need to hire a $200K ML engineer to create a prompt. English is the new coding language, and the best prompt engineers are now available on Fiverr and Upwork.
Disclaimer: Your definition of “quick” will vary by company. For some companies, 6 months is “quick”. For smaller companies, 6 days is “quick”. For a startup, 6 hours is “quick”. Find your own definition of quick.
Solution: Use low-code or no-code systems, or hire freelance developers to build quick, low-risk workflows.
Roadblock 8: AI Is Too Scary
Every day, the New York Times or some other mainstream media will publish some scary news story about AI.
At first, it was about hallucinations. And then about ethics. And now about misinformation. And the list goes on and on.
Given this humongous uncertainty, some companies are stuck in a “deer in the headlights” mode and taking a “Let’s wait and see” approach.
While this cautious approach might work for some (especially those with tremendous moats!), most businesses that have real competitors don’t have this luxury.
Just ask yourself: How could my competitor use AI and completely destroy my company?
Make no mistake: I am not saying that you should NOT do a risk assessment and see which of the AI risks are applicable to you. That is something I am totally NOT saying.
Rather, take a rational approach and see which risks really apply to you. This is no different than any other risk that you would take insurance for.
Solution: Do a common-sense rational risk assessment and see which of the AI risks is actually applicable to you. Treat it just like any normal risk you would consider in your business.
Roadblock 9: Which Team Should Own This?
Since AI is applicable to pretty much every team and every department, the big question arises around which team should lead the AI pilot.
For some companies, everyone wants to be a hero. And then there are some companies where no team wants to put their head on the line.
What should be the first use cases? And which team should own it?
Solution: Start with simple use cases. And let the biggest proponents and champions lead the AI pilot — irrespective of which department they might be from.
Why is this? You need early adopters and champions who will run through walls to make things happen.
It almost reminds of a “Lewis and Clark ” analogy: If you need to cross a mountain, find a horse. If you need to cross a river, build a boat.
You need people who think like this.
Roadblock 10 : No Employee Use Policy
You know that your employees are already secretly using ChatGPT — whether you have an employee-use policy or not.
There have been multiple studies published showing that employees at Fortune 500 companies have been secretly using the public ChatGPT on their phones (yes — the $20 subscription that learns on all your data!)
Solution: So rather than making employees go behind your back, have an open conversation and publish the employee-use policies.
This is your chance to even create a curated list of tools that employees can use. For example: ChatGPT For Teams was released last week and is starting to show excellent employee productivity gains (in a relatively safer environment — which means: this one does NOT train on your data!).
Roadblock 11 : Aligning The Goal
Doing an AI pilot project is great — but equally important is to decide upfront what the goal of the AI pilot is.
For some reason, some companies think that AI is the land of milk and honey. And that it will solve all their problems. I often hear “I wish AI could just do what Bob does”.
This would be like trying to go to Mars — before you can go to the grocery store.
Remember: You are taking a “crawl, walk, run, fly” approach.
领英推荐
You can establish some metrics such as reduced costs, higher customer engagement, and better lead conversion to gauge AI effectiveness.
Solution: Keep it simple. If the goal is to learn how to crawl, set the expectations accordingly. You don’t need to have the project fly on the first attempt. The very act of getting started puts you ahead of the competition.
Unfortunately, “Do Nothing” is not a good strategy on this one .. you will likely cease to exist. Maybe not immediately but eventually. See this termites vs tornadoes analogy from Harvard professor, Shikhar Ghosh
Roadblock 12 : No AI Champion
Literally every AI pilot I have seen starts with an AI champion. And trust me: I’ve seen thousands of successful AI pilots till now.
The common factor in all successful AI projects: An AI champion.
This is someone who will run through walls and do what it takes to make the AI pilot work.
For example: My favorite all-time AI champion is a person (name withheld!) working in the IT department of a top school district. Historically, this industry is the slowest to move onto any technology. However, this AI champion barraged through all the hurdles (legal, procurement, fear, et al) to successfully deploy an AI pilot.
Solution: Put together a “Tiger Team ” of AI champions if you need to.
Roadblock 13 : We Don’t Know Where To Start
With all the confusion around AI, everyone and their mother is now claiming to be “AI”. Software that has virtually no AI component are suddenly re-branding themselves as AI products and changing their domain name to .ai
What this does is: It creates confusion in your teams as to the best place to start for an AI pilot.
And when there is confusion, people freeze! (aka: “Deer in headlights problem”)
Solution: Use the WINS framework and see which use case is “In the Crucible”
Roadblock 14 : Do We Need Nvidia H100s?
A few months ago, Bloomberg decided to spend $100M and build his own LLM. Last week, news emerged that GPT-4 beats it out on most finance tasks.
So basically, what Bloomberg spent $100M to do, you and I get for $1.
The main lesson to be learned here is: Do you really really need to start putting money into infrastructure for AI? Or can you piggyback off someone else’s hardware and software.
In particular: Can you just piggyback off the OpenAI API or use a RAG SaaS — rather than investing in infrastructure and re-creating software (that you have no core competency in!)
See my previous post about “Build It Or Buy ” it — this is an important decision that needs to be made, before investing in infrastructure and MLOps.
Another example: Last week, I heard of another company spending $500,000 in the Azure cloud to create a RAG pipeline. While it is technically possible that this was a necessity, it should be the absolute last resort — after making sure that this is a path you want to go down.
Solution: Piggyback off SaaS APIs — before investing in infrastructure.
Roadblock 15 : We Don’t Believe in AI
AI being a new technology, there are some skeptics in the midst. It’s difficult to ask rank-and-file employees to believe in something that they have not experienced before.
It’s a natural human tendency. In fact, I have to confess: It took me a while to get onto Twitter and Whatsapp.
Side story: I got onto Whatsapp only after a visit to India and my uncle asking me “Are you on what’s up?”
Remember: People woke up to AI only after ChatGPT launched and shocked the world. The world needs champions and hero case studies.
Your company is no different. Until your company has tasted some success with initial AI projects, nobody is going to become a believer.
Solution: Start simple and make heroes of your AI champions and AI pilots.
So first: Get the employees to believe — even if it is in the slightest possible way.
One of my favorite case studies is where a company deployed an internal chatbot for customer support staff. Within days, staff productivity had gone through the roof with a huge cut in “time to resolution”. This “victory” then led to widespread deployment across the company.
Roadblock 16 : Our Employees Are Clueless About AI
On a recent call with an employee at a large Fortune 1000 company, the person asked me “What is an LLM?”
And this person was in charge of deploying an AI pilot at his company.
I understand: AI is not the most critical thing at every company. I get it — not everybody breathes and sleep AI like the rest of us.
Solution: Invest in AI training.
You can create training programs that include onboarding for new hires and ongoing education for existing staff. It does not have to be super-complicated.
Even having the in-house prompting expert do a demo for fellow employees will have tremendous impact.
And it does not have to be expensive. I make sure that every employee has ChatGPT Plus ($20 subscription) from Day 1. Once they have played around with it, I give them some training on prompting techniques like “chain of thought”.
Ninja tip: As a startup, I give my employees reimbursement allowances to hire prompt engineers on Fiverr for prompts and workflows they need.
Roadblock 17 : We Don’t Have KPIs For AI
While it’s early in the game to assess AI, most companies run on KPIs — so it’s important to define them straight from the start.
It doesn’t have to be complicated KPIs .. it could be simple KPIs like usage, number of queries/employee, monitoring chat logs, etc.
If you are deploying as a public-facing customer service bot, you can define KPIs like “ticket deflection” or reduction in “time to resolution”. The more usage the chatbot gets, the better for such customer-facing use cases.
Solution: Think about what metric would define “success”.
For example: I had a CMO recently tell me “I don’t know the exact numbers, but our ticket volume is down tremendously. We are super happy about this — and now deploying AI wherever we can”.
So basically: Directional data is fine — you don’t need exact numbers — but at-least have an idea about what you are optimizing towards.
Another reason for this is: When it’s time to decide whether your AI pilot is going to full-scale deployment, the CFO will ask questions about the expected ROI.
Roadblock 18 : Resistance to Change
When it comes to new things like AI, people and organizations often hesitate. It’s like when you’re used to a certain way of doing things, and suddenly, someone suggests a completely new way — it can feel uncomfortable and scary.
This is what we call “Resistance to Change.”
There are many factors to this:
In fact, one attendee in the learning lab said: “My client is worried that the AI is too good — and so there are concerns around what this will do to the human capital.”
Solution: Show AI as a productivity enhancer — NOT a replacement for humans. Start with small AI projects and show how they help.
Survey 1: GenAI Insights Learning Lab Survey
A study conducted in January 2024 surveyed AI consultants working with large corporations to identify key barriers to entry for adopting new technologies. The primary barrier was a lack of security protocols, followed closely by limitations in technical infrastructure, difficulties in problem identification, and a lack of employee training — all tying for second place according to the AI consultants.
Survey 2: ToTheWeb AI-Readiness Assessment
A second survey polled 135 professionals, mainly from the education, technology, and finance sectors, about their organizations’ preparedness to adopt AI. The respondents indicated that lowering costs and improving customer satisfaction were the top motivators for implementing AI solutions.
KEY FINDINGS ACROSS BOTH SURVEYS
As businesses navigate the complex AI landscape, they encounter a multitude of barriers that hinder adoption. The journey towards AI adoption is filled with both technical and operational hurdles, each requiring careful consideration and strategic management.
Below are the key challenges that organizations face in this endeavor:
For Visual Learners
Here is the full recording of the learning lab.
Shadow Usage of AI Tools Creates Risks
Fourteen months after ChatGPT’s launch, many companies continue blocking employee access to AI tools, yet OpenAI’s network data reveals employees at 80% of Fortune 500 companies are using ChatGPT at work regardless.
In companies where AI tools are allowed, there is often no reimbursement for the AI tools employees license themselves. The lack of official AI usage policies heightens the risk of mishandling sensitive corporate data by those untrained in proper data security practices.
By understanding these combined technical and non-technical challenges, companies can tailor their approach to AI adoption, ensuring they’re better equipped to address these issues and leverage AI successfully.
Assess Your Organization’s AI Readiness Score
This data-driven assessment tool quantitatively evaluates your business in key areas such as strategy, data management, technology infrastructure, talent pool, and governance.
By scoring your existing processes, the scorecard not only highlights your strengths but also pinpoints areas of improvement. With a clear view of your organization’s gaps, you can drive your AI roadmap and prioritize efforts to close those gaps.
If you found this article interesting, you are welcome to follow me because I write more about such AI topics.
Software