AI is inching closer to wide-spread adoption in the enterprise. 46% of CIOs recently interviewed by Gartner intend to invest in AI technology in the next year.
One question we hear a lot is cost: if you are planning to use AI, what can you expect to pay to make it worthwhile?
Though the depends and is likely different for each organization, we're seeing a consistent pattern in spending on AI R&D efforts as well as common ballpark budget figures.
What do you want to get out of AI? That is the key question that drives any discussion of cost. We see two common answers.
For most companies at this stage of AI development, the answer tends to be: we want evidence that further technology spend on AI will result in measurable ROI.
For others, there is already a clear use case, market opportunity or product feature that they have in mind. Their AI spend is more focused on production implementation rather than R&D discovery.
In both scenarios, a project budget can be broken down into three categories.
Technology research. First you'll need to understand which AI technology and approach is the right solution to your business need. There are numerous open-source and proprietary solutions, and understanding the technical requirements, tradeoffs and benefits of each will require a significant amount of time and resources.
Procurement and ongoing services/subscriptions. Once you identify the right approach, you'll need to either buy or subscribe to the right AI infrastructure to support your project. Even if you're using an open-source library for the heavy lifting, you'll need the infrastructure to support it, likely compute clusters that support parallel processing. Pricing here tends to be volume based, so be sure to factor in the amount of data or processing tasks you'll need to run in order to implement your AI capabilities.
Customization. AI needs to be customized to be useful. The current state-of-the-art requires that, at minimum, you need to train the models with domain specific information, and often you need to create customize algorithms to match specific use cases. The more customization, the higher the cost. Additionally, customization is going to require you hire experts - given the demand for smart AI folks, even off-shoring expertise isn't going to reduce rates by very much.
Budget Scenarios to Plan For
We've encountered three typical project budgets for AI development that seem to be pretty consistent. These are baselines, and specific requirements will typically increase the cost (see the three categories above).
AI Exploration - $300k
For those companies that are just getting putting a toe into the AI waters, we encounter a basic research exploration project. These typically are budgeted for about $300k and include a report and usually a sample of proof-of-concepts to fully test different technologies and approaches. These projects are very heavily weighted to using off-the-shelf components with little customization in order to determine what the baseline functionality is capable of, and how much customization will be necessary to tailor the technology to be truly valuable.
AI Implementation - $750k
Once you understand what you need to do, what technology you'll use to do it, its time to implement. This means getting the infrastructure in place, integrating with your existing technology or business processes and initial setup of the tooling. Though you may be able to use the learnings from an earlier R&D project, you will likely be investing significant amounts of money in new infrastructure development.
We often see the budget for this type of project balloon because of one simple mistake: the team forgot to consider the time and money it will cost to refactor existing technology to fully support modern AI tooling. Smart organizations will recognize that rebuilding older, clunkier parts of the tech first will result in overall lower cost.
AI Scale - > $1m
Initial implementation will tell you whether AI is your killer new feature; if so, you'll need to scale. This is where the hard work starts. Scaling AI gets expensive for two reasons: tuning and data volume. As you scale the feature to a large number of users, you’ll need to continue to invest significantly in additional training and refinement to handle the ever increasing edge cases. This requires lots of trial-and-error tweaking of the algorithms and the training data sets. The second expense will be handling increasing data volume. AI needs data to train and data to process. Both of these will add both hard infrastructure costs as well as development costs.
All in, to go from concept to scaling production system is likely going to be between $1 and $2 million of spend, but spread out over a couple of years.
Costs are going down for AI. As more ‘solution’ services arrive, the cost of the discovery and implementation will likely go down. However, its worth noting that AI will, for the foreseeable future, be a technology that requires a high level of customization for specific businesses. This will always increase cost over other software that is ‘off-the-shelf’, and organizations should plan accordingly.