TL;DL - To be able to charge for POCs in markets that typically do not pay, structure your POCs so that the customer gets value even things don't work out. And don't call it a POC.
You open a car dealership to sell cars. And given that your product is a significant purchase for your customers, you find that they naturally want a test drive before making a purchase decision. You provide test drives since it costs you a few dollars in gas - a fraction of the sale - so you consider it a reasonable cost of the sales process and eat the cost.
One day you find that the cost of providing a test drive shot straight up and now the test drive costs you 10% of the (expected) sales price. You start charging customers for test drives to offset the costs and "to put skin in the game". You find that some customers with deep pockets don't mind and considers it 'normal' in some neighborhoods but in others you get huge push back.
This is the situation startups that sell AI/ML heavy products find themselves in. Due to the not-so-automated process of data science and high cost of data scientist, providing a 'test drive' with the customers' data costs much more then would an average software product. Large enterprises understand the need to do a paid proof of concept and consider it natural cost of evaluating new solutions. But smaller companies the economics only makes sense if the POC is free (or close to free) and some international markets (e.g. South Korea) POCs are expected to be free regardless of the size of the enterprise.
Plus, "Why would I pay for something I don't know if it would work for me?". This push back is so much more valid in the AI/ML space. Due to the trial and error nature of the 'science', it's largely impossible to know if certain predictions can be made at certain accuracies before hand.
So, how do you expand into these segments and geographic markets without breaking the bank by charging for POCs?
At the end of the day, its about the expected return (below) and Cost. . R being the return (impact of the solution/service), P being the probability (likelihood of succeeding). Some organizations would never (or couldn't) invest in anything that is below a certain probability and above a certain Cost, no matter how great the Expected Return is.
And AI/ML projects typically have a lower P and unclear R. Thus a lower or unclear Expected Return. Seems pretty bleak.
But the answer is to increase the P and make the R crystal clear and adjust Cost to match the new Expected Return.
Redefine the POC from a 'we'll do our best but if its doesn't work, there's nothing we can do' to a 'even if what you want cannot be achieved, you'll still come out ahead' pitch . Redefine the POC so that the customer gets something in return regardless of how it turns out- and the something should be very well defined for a clear R (return). For example, insights on how to better manage their data along with best practices, limitation of their data and system basically anything that gives the customer something that is valuable and even better if its a (presentable) by-product of your existing product or process. And if the something is by-product of your existing process, it pretty much guaranteed to materialize. Thus essentially bringing up the P to 100%. Be creative on but you'll need to good understanding of your customers to do so.
"Provide value every step of the way. "