AI adoption is no longer optional—it’s a competitive necessity. But deploying AI isn’t as simple as downloading pre-built models or subscribing to generic cloud services. Half the AI your business needs already exists in vendor solutions, but the other half is uniquely yours. Half of the AI you need must reflect your specific processes, data, context, and goals. And, the faster you deliver your unique AI at the lowest costs means you will gain a competitive advantage over those who will wait.
Which is Which Half?
Generative AI models, such as those used for summarization, content creation, or simple queries, can be effective starting points for many organizations. These pre-trained models offer broad utility but come with critical limitations. They may help with small tasks within broader business processes, but they still place the burden on employees to master prompting techniques, extract relevant information, and integrate it into workflows.
For example, Google’s 67-page Gemini prompting guide illustrates the complexity of effective prompting, teaching users to craft specific personas, contexts, and desired outputs for each task. However, prompting guides quickly become obsolete—Gemini has undergone 10 major updates since the October 2024 guide was published. The pace of updates outstrips the ability to train employees effectively, leaving organizations struggling to keep up.
Moreover, without robust security and governance frameworks, relying on generative AI models exposes businesses to significant risks. Sensitive data, proprietary processes, and unique terminology require protection and context-specific understanding to avoid compliance violations or operational missteps. As discussed in our podcast GenAI is Only 5% of the Solution, generative AI is helpful for general tasks but cannot address the unique challenges of your organization. The other half of your AI solution must be customized to your workflows, business rules, and goals.
To truly harness generative AI, businesses need an integrated approach—combining off-the-shelf capabilities with bespoke models tailored to their data. This ensures decisions align with your organization’s needs while maintaining strict security and compliance standards. Krista enables this by providing a framework for safely and effectively customizing AI, empowering businesses to implement solutions uniquely suited to their operations.
The Problem with Traditional Machine Learning
Traditional machine learning development is resource-intensive, requiring specialized expertise to select models, clean data, and fine-tune algorithms. This creates bottlenecks, as businesses often lack the bandwidth or resources to manage these tasks efficiently. Even organizations with dedicated data science teams face lengthy backlogs and escalating costs, limiting their ability to focus on high-value projects.
Some larger, high-consequence decisions, such as foundational business models or critical predictions, are best suited for data science teams. These complex problems may require months of effort to build and refine the appropriate machine learning models. However, the majority of higher-frequency, lower-consequence decisions are ideal for automated machine learning.
This is where Krista excels. Krista automates the entire machine learning process, from capturing and formatting data to selecting the best-fit algorithm for your specific use case. By managing these routine tasks, Krista enables businesses to implement AI quickly and cost-effectively without relying on expensive, specialized data science teams. This frees up your data science experts to focus on strategic, high-impact projects while Krista handles the rest, ensuring efficiency and agility across your organization.
How Krista Delivers Automated Machine Learning
Krista’s machine learning capabilities make it possible for businesses to deploy AI models with minimal effort. Here’s how:
Data Ingestion: Krista connects to your existing systems and extracts relevant data, preparing it for machine learning.
Automated Model Training: Krista selects the optimal algorithms for your data and trains models without requiring manual intervention.
Continuous Optimization: Models are evaluated and fine-tuned to maintain accuracy as business needs evolve.
System Integration: Krista deploys models into production and integrates them into your workflows, automating decision-making processes end-to-end.
Whether you’re categorizing emails, predicting customer churn, or optimizing inventory, Krista adapts to your use cases and scales with your business.
Why Krista’s Approach Matters
Krista accelerates AI deployment and delivers measurable benefits:
- Faster Time to Value: Begin seeing ROI from AI initiatives in weeks, not months.
- Reduced Costs: Eliminate the need for extensive data science resources.
- Improved Accuracy: Benefit from consistently optimized models tuned to your data.
- Increased Agility: Quickly retrain models as new data emerges or business needs shift.
For mid-market organizations, Krista levels the playing field. Even without a dedicated data science team, businesses can leverage AI to drive meaningful outcomes.
Real-World Impact
Krista’s clients consistently experience transformative results by automating routine processes and enabling their teams to focus on strategic initiatives. For example, a leading customer service operation reduced call handle times by 30% and increased first-call resolution rates by integrating Krista’s AI-driven automation. Another client, a global retailer, used Krista to streamline call center operations and onboarding, cutting onboarding time from weeks to just a few days while improving agent performance and customer satisfaction.
By automating repetitive tasks like email categorization or support ticket triage, Krista helps businesses operate at machine speed, ensuring quicker resolutions and higher productivity. This not only reduces operational costs but also empowers employees to engage in high-value work, driving both efficiency and job satisfaction across the organization.
Take the Next Step
AI adoption doesn’t need to be intimidating. Start by identifying processes with high volume and clear ROI potential. Krista can help prioritize these initiatives, guide your journey, and automate the unique AI models your business needs to thrive. Let Krista show you how automated machine learning can transform your operations today.
Links and Resources
Speakers
Transcription
Scott King
Hey everyone. Thanks for joining this episode of the Union podcast. I’m Scott King, your usual host, and I’m joined by our co-hosts, Chris and John. Hey guys.
We’re going to talk about AI and where you’re going to gather it from. Half of the AI you need, Chris, is available—you can subscribe to it, download it, or use a service to help you generate content or use tools like LLMs. The other half is unique to you. It doesn’t work for anyone else; it only works for your business. You can’t borrow it from a competitor or download it and expect it to work automatically. What exactly do we mean? If half is available and the other half is unique, can you give us some ground rules?
Sure. Everyone’s reading about AI, learning about models and machine learning. We’ve talked about this before, but every business operates differently. That’s why you’re not your competitor. Different parts of your business may need to solve the same problem, but in different ways. For example, you want to read customer service emails and categorize them based on the actions needed. That’s a categorizer.
We need to see your customer service emails to understand the words and terms you’re using, then build a categorizer tailored to that. You can’t take the same categorizer and use it for accounts payable or receivable. While both involve categorization, the data, processes, and keywords are completely different. Tailoring it to your business needs is key. You might need 30 categorizers across your company.
This type of machine learning or AI works differently depending on its purpose. A categorizer for accounts payable won’t be the same as one for receivable because money flows in opposite directions. Similarly, customer service categorizers for sales or support queries are different. Each is built to understand your specific terms and processes.
John had a great analogy. We have 3D printers because sometimes you just need one specific part. You don’t need 400 of them; you need one of a specific size to complete a project. That’s why people love 3D printing. If something fails or doesn’t fit, you can create one part tailored to your needs. The term “bespoke” applies here.
People need to realize that while many tools exist, they don’t understand your business. They don’t know how you want to answer questions or use your terms. We had a customer with 20 acronyms unique to their company. You can’t buy AI online and expect it to know those. It’s trained on internet data, not your company’s terminology.
For example, “yield” at one company meant renewals. Everyone knew it referred to renewals. You won’t find a machine learning model trained to interpret that correctly when handling sales questions.
Scott King
That’s like how “turnover” in the UK means revenue, but here, it’s something completely different. Interesting. John, can you expand on what Chris mentioned about unique AI and categorizers? Does this go down to the specific technology you’re using?
Chris Kraus
Yes. Absolutely.
Scott King
You have to learn from people, right? Let’s dive into an “if-then” scenario. We’ve talked about the half of the AI you can buy, and we’ve covered that extensively in podcasts. AI must do something—it has to orchestrate. LLMs are only 5% of the solution, and we’ve shared plenty of resources on that.
The other half of AI requires expertise to assemble it inside your business. That’s the unique-to-you part. Chris, if I decide I need categorizers, walk me through what happens if I have one or more data scientists on payroll—or none. It sounds like a daunting task.
Chris Kraus
If you have data scientists on staff, give them the hardest, most complex projects. You might review your projects and realize Krista can handle 45 of them quickly, delivering 30% of the value. Let Krista do those automatically. Then let your team focus on the project that might take a year to research and deliver a 40% lift.
The key is to ensure every AI effort contributes to end-to-end automation. As John said, models decay over time. You don’t want to spend 12 months building a model and then another 12 months building an application to host it. By that point, your model could be outdated. Krista can host the model she builds automatically, offering one platform to manage both. It’s not a question of one or the other—do both.
Many mid-market companies don’t have data scientists, but that doesn’t mean they can’t benefit from AI. Krista becomes your data scientist. Present Krista with the problem: Are you predicting a number? Categorizing something? Identifying alert levels or email types? Ask practical questions about what your team does today. For instance, if someone reads an email and determines its category, that’s labeling—Krista can do that.
Even if your data is messy, Krista can start as a novice. If you lack labeled data, that’s fine. Krista will ask clarifying questions like, “What is this?” over time. For example, she might ask if an email is about a product return or a service return, and you’ll guide her. As Krista interacts with your data and processes, she learns and improves, eventually automating the task entirely.
If Krista is trained on a hundred examples over a week, she’ll build a model automatically. You can then have her notify you when she thinks she has an answer. You can approve or deny her suggestions, and if she’s unsure, you can provide the answer. Over two to three weeks, as she gathers more examples and learns from daily tasks, Krista becomes increasingly accurate.
At a certain point, you’ll notice your team trusting her decisions more. For example, if you set an accuracy threshold at 60%, you might adjust it to 70% or even 80% as confidence builds. Eventually, you let her handle routine tasks entirely, focusing only on exceptions.
This process works whether you have a data scientist or not. If you have labeled data, Krista can leverage it. If you don’t, you can still get started by training her through everyday interactions. The key is understanding your biggest problem and desired outcome. Don’t pursue machine learning because it’s trendy—use it to solve a real business problem.
People who think machine learning is glamorous haven’t done it. The majority of time is spent dealing with messy data—exporting, transforming, and managing pipelines on hardware that’s often insufficient. Sure, there’s exciting work in shaping models and comparing algorithms, but it’s rare.
That’s why you want software to handle as much of the data processing as possible. Krista excels at building her own models based on your data and decision-making processes. While Krista is great for automating routine, high-volume decisions, truly high-consequence models—the ones foundational to your business—still require top-tier data scientists.
The challenge is that as soon as you have a data science team, you’ll face a backlog of projects that could stretch years. Krista alleviates this by taking on most of the routine work, allowing your team to focus on critical projects.
This creates a symbiotic relationship. As Chris said, building models isn’t enough. You also need to deliver them, integrate them with systems, feed them data for predictions, and involve people for validation when needed. Incremental training from real-world use is essential to keep models healthy. Krista handles all of this, whether for her own models or those built by your data science team.
We’ve worked hard to address constraints in adopting AI and accelerating tech change. One of the most rewarding aspects is seeing companies transform. Initially, many are skeptical and hesitant. Three months later, they’re excited and ready to expand AI’s role in their operations.
At its core, AI is about providing sufficient data to make decisions. If the data is lacking, decisions can’t be made. Krista operates with this principle, ensuring transparency and confidence in every decision. She helps businesses navigate the journey from uncertainty to efficiency, showing that AI is just math—but math that transforms how work gets done.
Even if traditional AI practices don’t support this approach, we’ve created one that gives you the confidence to act on predictions. You’ll know whether the AI is certain enough to proceed or if it needs further clarification. Sometimes, the AI might tell you it doesn’t know the answer yet, but it’s working toward it.
This process mirrors how businesses make decisions. You never have complete data or unlimited time. Decisions are made with incomplete or imperfect information. We’ve designed Krista to bring AI into this real-world framework, empowering you to make decisions with confidence and transparency.
Scott King
John, you mentioned decisions, finding answers over time, and the backlog. Let’s tie those together. What does the cycle time look like today when starting a project? How does that change in 12 or 24 months?
John Michelsen
It’s a constant challenge—what I call a physics problem. Integrating systems that don’t cooperate with Krista often becomes the longest part of any project. Whether it’s a simple or complex integration, working with systems that resist communication slows things down.
If Krista is orchestrating people already connected through communication channels, things move quickly. Similarly, if systems have clean, well-designed APIs, integration is straightforward, often taking four to six weeks. Our team is usually showing significant progress within days of a kickoff meeting.
However, when we encounter systems that only operate through user interfaces and resist integration, the process becomes more complex. The good news is that time tends to solve these issues. Systems either modernize, get replaced, or we build a one-time integration.
John Michelsen
Once Krista connects with a system, you won’t need to integrate it again at a wiring level. Krista acts as a bridge, allowing any other system using Krista to access the connected capabilities without additional development. This isolates the rest of the company from the complexities of backend systems, making modernization and system swaps easier.
While it’s not effortless, it simplifies a significant challenge—managing change for people. Instead of employees showing up to a new system and feeling lost, Krista handles the transition. This eliminates much of the confusion and frustration that typically comes with change.
Scott King
Easier, not easy.
John Michelsen
We can solve challenges that were once almost impossible. Consider all the SaaS software that was supposed to be easy to replace—how easy is it? The reality is, it’s hard. Why? Because your people are the integration strategy. They’re the ones using it, and their understanding is built into the processes.
When you replace a system, you disrupt this tribal knowledge, and now even that isn’t accurate. This makes change difficult. The key is AI trained on your data in near real-time. That’s how you build the models that transform your 10-step processes into six or even two steps. Let the machine handle the repetitive work.
Why two steps instead of zero? Because AI doesn’t know everything. People excel at solving challenging, interesting problems. We’re not as effective at repetitive, mundane tasks, and that’s where machines should take over. Automating this work improves accuracy, reduces turnover, and helps people focus on more engaging responsibilities.
John Michelsen
Teams with nearly 100% turnover aren’t scared of losing their jobs—they’re already planning to leave. Krista hasn’t caused a single person to lose their role. Instead, employees move into more interesting positions. Some even become Krista authors, leveraging their knowledge to improve workflows. Their expertise becomes an asset to the organization in new ways.
Scott King
That’s the goal—automate processes so machines or software handle the work, freeing people to refine it and accelerate learning. Nobody wants to do repetitive tasks endlessly. Chris, if someone wants to start, how do they pick a process or categorizer to try first?
Chris Kraus
The key is practicality. Don’t focus on tasks you do once a year—that won’t justify the investment. Look at tasks performed daily, like email categorization. Some organizations have teams of three to ten people categorizing emails all day. Those employees aren’t happy doing 100 to 500 of these tasks every day.
AI is perfect for these scenarios because it’s solvable and provides a high return on investment. If you list ten processes, we can help you prioritize. For example, categorization and answering questions are straightforward for AI. We can analyze your goals, the frequency of the tasks, and their importance to identify where to start.
Scott King
John, any final thoughts before we wrap up?
John Michelsen
Focus on highly repetitive tasks first—they’re your top targets for automation. Start by having the machine handle as much of the work as possible. Then, identify friction points where AI can handle 80% of the work. Introduce models to address those gaps.
Think of it like 3D printing. As you build, you realize you’re missing a specific piece that a machine can handle. That’s your bespoke solution. Most business processes don’t take weeks—they take minutes but sit idle in email queues waiting for someone to act. This causes delays, lost revenue, increased costs, and even compliance penalties. When machines complete these processes end-to-end, it happens in minutes, guided by people where necessary.
Scott King
Exactly. Tasks sit in email queues until someone gets impatient and escalates them. Thanks, everyone, for the great insights. If you have questions, reach out to us at Krista.ai. We’re here to help with building your own ML. Until next time.
