(This column originally appeared in Forbes)
The age old question: build or buy?
Ella Haman is the chief technology officer for Kapitus, a financial services firm that offers financing primarily to small businesses. Like many companies in her industry, Kapitus has been investing heavily in internal AI-based systems to help their advisors make quicker and more accurate financing decisions. Kapitus has decided to build these solutions rather than rely on external software providers and Haman, as CTO, has – no surprise – been significantly involved.
Why Build Internally?
Many businesses are trying to figure out whether to build or buy AI solutions. Last year I wrote on all the steps that a small or mid-sized business needs to take in order to accomplish this. But there are other factors to consider. For those thinking of doing it themselves, here are a few things that Ella has learned.
The Starting Point Is The Data
AI is a tool. But it’s a tool that won’t work very well if it’s using unreliable data. Which is why it was critical for Kapitus to make sure their data was as complete and accurate as possible.
“AI only works if the underlying data is secure, governed, visible, and usable,” Haman said. “Most AI failures happen before models are ever deployed.”
If your company is like mine (and so many of my clients) you’ve probably got a lot of issues with your underlying business data. Empty fields. Wrong or out-of-date information. Inconsistent entries. Thinking that you can run automation – AI or otherwise – on inaccurate or unreliable data is going to get you into trouble. Haman made this a priority and took measures to clean up her underlying databases before implementing their own AI platforms.
Don’t Do Everything At Once
How do you tackle a big challenge? By breaking it down into smaller challenges and conquering them individually. An AI implementation project is like any other project. And every project manager will tell you that it’s important not to bite off more than you can chew. AI can be transformational, but only through small, measurable automation.
In the private equity world, companies like Kapitus are deploying their AI solutions so as not to try and replace their existing systems wholesale. Instead, Haman has concentrated on using AI to remove friction and human bottlenecks via small bites.
“The real gains come from automating early and repetitive steps, not replicating human judgment,” she said. “We’ve been doing this through small, measurable chunks rather than trying to build one massive solution.”
Haman points to some data – like this recent report from MIT – that shows a high failure rate for most corporate AI projects. For her this isn’t surprising, and it’s avoidable.
“The mistake people make is going for really big things at once, instead of identifying small use cases that actually fit and measuring the impact,” she said.
That said, Haman has not been averse to making bigger changes when the situation warrants.
“We’ve been trying to focus on redesigning our processes and workflows from scratch,” she said. “The longer term objective is not to build an agent that replicates what people do today—it’s to reimagine the business process itself.”
Governance Is Critical
According to Microsoft, data governance is the “overall framework of processes, policies, roles, and standards that ensures an organization’s data is managed effectively, securely, and consistently throughout its lifecycle, making it reliable, usable, and compliant for business goals, decision-making, and risk management.” It’s a lot of words to say make sure there’s oversight.
Ella says that governance over AI systems and their processes is as important as innovation and, arguably, it’s the most important part of an internal AI-based system. Firms like Ella’s have a lot of data to manage and much of it is confidential financial histories of both current and prospective clients. Misuse or loss of this data could be catastrophic.
“It’s our own private data and it comes from external sources like credit rating agencies as well as transactions and other operational sources,” Ells said. “Having it governed and usable is critical.”
This is the reason why much time has been spent creating standards, roadmaps and other forms of governance around the information that is being used by AI. To do this, the company formed its own “AI Council” to define the “rules of the road” that sets out parameters for evaluating tools, making decisions and deploying applications.
“Professionally deploying our systems — how we secure it, authenticate it, retrieve and log data – that’s where governance really matters,” she said.
When I work with clients and speak to technology leaders like Ella who have made the decision to build their own internal AI systems rather than buy the biggest common factor comes down to one word: control. For Haman, the decision to build her company’s AI systems internally was not so much because their vendors aren’t good, but because their management felt it was critical to keep full command over their data, economics and outcomes.
“We not only want control of the data, but we want to feel confident and have the ability to move quickly,” she said. “That doesn’t mean we don’t use outside vendors, but if we do we need to know the ROI.”
Doing it on your own is costly, time consuming and has its share of risks. But if done the right way, a company can have full control over their data and their processes and the AI agents they build so that their systems are working the way they want, rather than more expensive (and riskier) compromises.
