This doesn't address the permissions problem mentioned in the OP. That one is non-trivial.

That said, even if the offshore model is dead because of ai, there is still an issue of ai becoming more costly than a dev team that knows what it is doing. Ai is data reliant and if any aspect of that data is borked, the ai will consume a large quantity of tokens without telling you "hang on, something is not right" before the customer will.

Case in point - I worked as a consultant during covid for a large drinks conglomerate that was a globally traded company. They were getting most of their data from Nielsen, but said data was so badly fragmented and fractured it was like trying to mine gold from a turd. It took nearly a year for a whole team to get it into a semblance of quality. Even then, it took me with a team of 5 others to find out that the data cleansing we had suggested (which improved the analytics by magnitudes on what they had) still had faults.

We used LLM in python to find the low hanging fruit (double spaces, misspellings, words in different order etc), but the more complex ones required hand reviews by yours truly to get the last 20% of data quality errors resolved. By the end of it we found 50k duplicated product names, and that was for one country, on one data provider's info. Now, imagine that being amplified by 40 different data providers across 70 countries, or the same data provider providing that file having 50 different names for the same products in one country, and 20 in another. Give that to an ai without telling it "this data is messy AF" and it will trundle along presuming everything is fine when in fact it is so far out from reality that it's effectively lying to you.

Another case, another company, with kind of the opposite issue called a slowly changing dimension or SCD - a big uk based company, who had a marketing team struggling to analyse anything going on. This was down to the fact that the unique identifier for all their products, the barcodes, would be shuffled around by the supply chain team to represent different products regularly. One day you're looking at raspberries, the next day it's a blouse.

They had to build a table to say what ibarcode reflected which products across which dates internally, simply to run analysis internally. This is something which slowed down query processing to a crawl over time, and dramatically raised computing requirements. Marketing couldn't stop it, and the supply teams would change their policy as it would ramp up costs massively to stop their practices.

I'm not aware of an ai being able to go to a company and say "this is costing you lots of money" upon seeing it, as the ai would not be able to think outside the box to say "you don't have a working UID" without some context being fed to it. Meanwhile, upon flagging the issue, it took me and one other guy about 10 minutes to find a solution, to whit "create a new UID specifically for the marketing analysis". The ai would have gone "oh well, I guess I have to use this table you built" and burn through not only tokens but other resources.

Ai must presume data quality exists unless it is trained to see it doesnt. Ai must presume data is not specious at best, or trash at worst. And ai cannot make data quality better without human inputs saying "these 50 things are the same thing" or "this one thing is 50 different things, and it's all depending on the dates you are looking across". Or rather it can, but not well without oversight. What it can do is identify probability of something being a duplicate, and the way we label shit can make that dangerous for an ai to do alone without a human governing it. And people

Another thing. People understand time better than computers and much better than ai, as there are libraries upon libraries in db systems just to handle simple concepts to us like daylight savings and date formats. The worst is semantics - in business, dates and times are used all over the place. Order date, transaction date, delivery date, reporting periods vs accounting periods, etc. Humans can tell the difference inherently that transaction dates differ from delivery dates, but an ai must be told this every time. Time is multidimensional in business, and although ai can find anomalies fast than a person, it cannot reason about time on its own.

The there is R&D. This is probably one of the worst places to use ai but I'll guarantee that companies are probably using it here more than anywhere else. R&D is a highly iterative and test to fail process. In that department you expect slow progress on rate of return because when it finally succeeds in building a viable product the value is orders of magnitude greater than the dev costs, provided you have a suitably sized team for it. With an AI, the cost is exponentially higher at doing R&D because token usage and compute will go through the roof. One experiment can turn into a costing nightmare if ai is the engine doing the R&D instead of a team of highly skilled devs on a fixed salary.

Finally there is the trust problem. You get a computer to do all that assessment of quality for you without human validation saying "looks good" and no amount of compute is going to fix the errors. You get a computer to auto ingest it, and you will see the use of tokens ramp up as people smash the ai trying to get it to find out wtf is going on with their now borked reporting suite. Once trust in reporting and analytics is gone, adoption of the ai falls through the floor, and rebuilding trust is one of the hardest things for people to do once it is lost. Then you have millions spent on a tool no one wants or feels they can rely on.

Tom Baker said it best in Dr. Who - "the trouble with computers, of course, is they are just sophisticated idiots." ai is greater at finding anomalies and workarounds, or writing code at speed. What it is absolutely shit at is the human condition. It can't handle any of the following well on its own:

• ambiguity
• governance conflicts
• temporal reasoning & inconsistencies
• undocumented workflows
• legacy data
• edge cases
• compliance requirements
• organisational politics

AI costs are also not just tokens x pricing. They are much more nuanced, and need to include:

• inference compute
• vector databases
• orchestration
• retrieval pipelines
• retries
• hallucination mitigation
• monitoring & data error notification
• prompt engineering
• guardrails
• fine tuning
• evaluation systems
• human review
• governance
• latency management
• security
• data cleaning
• context window expansion

All those costs balloon the cost of an ai project massively beyond just token price, and mass adoption of an ai across a company just make those issues worse.