The build vs buy question needs to be handled differently for AI solutions. The old 80-20 rules of SaaS don't apply cleanly.
I was analysing a real AI use case for a tier-1 construction contractor, with operations spread over most of the Middle East. The problem was real and reflected in both lost revenue as well as large manual effort costs. So I worked with them to figure out what's the best approach given the current state of things.
A standard build vs buy analysis. What are the off-the-shelf solutions that can handle this, vs what will it take / cost to build it all in-house.
For ERPs, CRMs, WMSs and other SaaS setups of old, the standard rule was to have about 80% of the business processes adapt to the solution and have about 20% customisation for org-specific elements. Most solutions either had an on-prem or a single-tenant option to take care of data segregation, and there would be a defined mechanism for business process updates to flow to other tenants.
In the AI era, data is the currency, not the business process. The question becomes how well the proposed model works on this org's data. For the scenario I was helping evaluate, the available solutions had been primarily trained on US-focused data. For GCC notations and terminology, most wouldn't give more than 60% accuracy, meaning automation would still need manual oversight of every model output. With such a high exception rate, the time saved or resources reassigned would be minimal. Actual business benefit didn't justify the costs involved.
These models are trained only on publicly available data. For regional quirks in data, models trained on relevant local data are needed. For SaaS, the 80% generality came from learning from tenants, with process improvements flowing back to all of them. AI models can't replicate that loop. Tenants can't allow their data to be used for training, due to IP and NDA requirements. On-prem using the proprietary models gets preferred but still has the same model training issue.
On the build question, it was possible to fine-tune open-weight or proprietary models using own data. Just like older build options, this has the same risk-ownership problem. You don't know what the outcome of internal development is going to be. It could give an accuracy of 90% or 40%, depending on execution.
In the current market, getting great ML Engineers and Data Scientists who can help achieve that 90% accuracy is a very costly affair. The payback period to do this properly can go up to 2 years.
So you end up having to decide between not-fully-fit-for-purpose off-the-shelf solutions running on-prem, or building something whose payback period is long and uncertain. Continuing to absorb the current manual effort may cost less than buying a solution capped at 60% accuracy or building one that might only reach 40%. Maybe you're better off living with the problem itself for now.
#AI #BuildVsBuy #DigitalTransformation #AIAdoption