I. Introduction:
Let’s be honest, the generative AI genie is out of the bottle, and there is no going back. Work as we know it has changed dramatically in the last few years, as more and more knowledge workers are expected to use AI to increase their output while simultaneously accelerating timelines. According to Anil Dash, 500,000 tech workers have been laid off since ChatGPT was introduced. No one can attribute the exact causes of those layoffs to AI alone, but it’s safe to say that, as a result of smaller workforces, leadership is looking for efficiencies, or in other words, ways to do more with less.
That is all fine and good on paper (or pixel), but the iron triangle of “you can have fast, cheap, or good. Pick one” remains true. For example, the promise of ‘vibe-coding’, where a person can tell an AI what the final product they desire should be in a few sentences, does arguably produce a fast result, but is it good? For those of us who survived the first dot-com bubble, web 2.0, a recession, the second bubble, the metaverse, and now find ourselves trying to adopt AI – I would say that ‘vibe-building’ is impressive, but it’s rarely good.
The issue lies in LLMs trying to be good and be everything to everyone. Think of it like hiring a management consultant who knows your industry but nothing about your specific business to set your roadmap. One of two things will happen. The first scenario is that you will get a lot of advice in the form of a thick document that sounds very impressive. Then you will see the cracks, your workforce will get anxious, attrition will begin, and you will have to work overtime to correct the errors. The second is that you trust that the advice was sound and stay the course. Chances are good that the initiative will fail.
Think of it like this. AI is trained on both facts and fiction, but it doesn’t always know which is which. It aims to please, and as a result, you frequently get a mix of both.
In my experience building with AI tools over the last 4 years, if you don’t give AI the context it needs to really understand the deep context, it will just give you technically correct, well-sounding, impressive-looking output, but that almost always lacks the quality of traditional methods. AI is here to stay, though. So how do we maximize its ROI while doing more with fewer resources?
II. The adoption gap:
Before we can improve the output of Generative AI tools, we need the workforce to use them. According to Greg Shove, CEO of Section – an OpenAI Enterprise training partner,
“The results are clear – most employees are untrained and unprepared to effectively and safely prompt AI to unlock meaningful productivity gains,”
This aligns with my experience leading a UX team at Cisco. Generative AI tools adopted the empty ‘Google Search box’ or blank-canvas interface in an effort to do everything for everyone, and as a result, were not suited to support employees who didn’t know how to apply them to their specific processes and tasks. The response to low workforce adoption rates was typically to spend time and money on training. That money spent makes achieving ROI harder.
There is another imperative issue to be addressed. That is the one where AI companies use your input in their future training, creating a data security risk. The AI never forgets. AI companies offer ‘enterprise’ licenses that provide security, but they are cost-prohibitive for anyone but the largest companies.
A lesser-known but equally effective solution is to use the API endpoints from AI companies. In almost all cases, especially at foundation model providers, data entered through the API is not stored or used for training. The trade-off is that the API doesn’t include an interface. It’s up to the user to build their own.
The last adoption hurdle is that of choices, or more specifically, the overabundance of choices. Companies are struggling with what is called ‘shadow AI’, where employees use sanctioned products on their own with company data. This could be due to fear of being seen as lazy for using AI, comfort with other products, or a lack of features and quality in the products they are provided. On the latter, it differs from the blank-canvas issue. Rather than not knowing what they need for their specific task, they know what they need but can’t get it, so they go elsewhere.
III. Examples of ROI from AI adoption:
Chat Agency AI has become our solution for these challenges. It started out as a collection of AI integrations into our proven business consulting process – built from over two-decades helping the world’s largest brands launch new products. At each stage and step, we identified an AI tool that produced good output, on par with traditional methods, but this meant knowing when to use each tool, rewriting the input prompts each time, collecting and storing the output, and maintaining and managing multiple iterations as our projects progressed.
Our solution has been to build a platform that follows our process, organizes our ideas and projects, and obfuscates the prompts and model selection to the experts who maintain it. Some of the results we’ve captured from customers are:
- 1. Automotive B2C Startup: A solo founder compressed six months of research into a single sprint, uncovering untapped market segments that led to a successful provisional patent filing and $300k in saved consultant fees.
- 2. Financial Services Firm: A 10-person banking product team scaled their strategic capacity by 40x—moving from 5 ad-hoc reviews a year to 200 data-backed evaluations—eliminating management projects in favor of fact-based roadmaps.
- 3. Construction Company: A $200M construction firm avoided a six-figure consulting engagement by using Chat Agency to build the business requirements and financial models needed to successfully pivot their internal software into a market-ready B2B product.
IV. Importance of security context in the professional services industry
What we have found is that by building a front-end tool specifically designed for a process we know works, we have been able to address many of the challenges companies have faced with AI ROI.
Firstly, by curating the best AI models for each step of the process and storing the results in our own database, we achieve higher-quality results than a single AI provider or individuals trying to manage it themselves.
Secondly, we have greatly reduced shadow AI use by ensuring the tools can meet our users’ specific needs. An added benefit has been knowing that our data is not ending up in future model training.
While there was a cost to building a custom tool, that cost was quickly recouped by avoiding the need to continually provide and maintain training resources.
V. Conclusion:
The answer isn’t to just license Claude or ChatGPT. One model doesn’t cover all the bases — though AI companies will try. You also can’t lock down corporate environments, so no one can experiment. Everyone who is curious is experimenting at home, outside the business context, and bringing those expectations back to work. Corporate environments will benefit from specialized apps for high-value processes that leverage AI and mirror business context.
In some cases, it will make sense to buy outside tools (like Chat Agency AI) that are already aligned with their business content. There is also a need for multi-model sandboxes (like the core codebase of Chat Agency AI) where they can quickly build AI tools for their proprietary needs.
