BBVA built 20,000 GPTs and uses about 4,000. What makes a keeper?

There's a number I keep coming back to. At BBVA, employees have built more than 20,000 custom GPTs, and the bank reports that roughly 4,000 are now in frequent use, according to a Fortune commentary by Will Drover. Do the subtraction and that's around 16,000 tools somebody made that aren't in frequent use. You could read that as a story about scale, but I read it as a story about which tools survive, because something made those 4,000 worth coming back to and that's the part I want to copy.

One bot, one recurring question

The clearest example I've found is one of BBVA's own. In Mexico, before a company can open an account, sign a contract or get credit, the bank has to run a legal check called a bastanteo, which confirms the person sitting in front of them actually has the authority to sign for the company.

According to OpenAI's State of Enterprise AI report, that check used to lean on a specialist legal team answering the same repetitive questions from branches, which meant delays and bottlenecks. So BBVA built a generative AI chatbot on ChatGPT Enterprise that gives branches instant access to standardized, pre-validated legal FAQs and documentation guidance, and BBVA's Legal Services team developed and reviewed the content behind it.

The report says the bot now automates more than 9,000 queries a year and let BBVA redeploy the equivalent of 3 FTEs toward producing over 11,000 bastanteos per year. To be fair, this is OpenAI's report about a customer using OpenAI's product, so treat the numbers as vendor-reported -- the same report also says BBVA regularly uses more than 4,000 GPTs, which at least lines up with the Fortune figure.

Nothing about that bot is flashy. It answers one question that comes up constantly, for one team, using documents a legal team already stands behind, and I think that plainness is exactly why it has lasted.

Drover's piece is mostly about something else, and he makes a fair point. He calls it the builder activation gap, "the distance between the many people who could build with AI and the few who do," and his argument is that most companies have far more people who could make a useful tool than people who ever actually try.

The gap is real, but closing it on its own just gets you more attempts. Get everyone building, celebrate the count, and you'll probably end up with your own version of 20,000 GPTs and no clear idea which ones matter. How many tools you built is the wrong scoreboard, and what I want to know is how many are still in use six months later and what each one replaced.

There's a hint of the same pattern in the wider data too. Anthropic's March Economic Index report found that since August 2025, usage of its first-party API has become more concentrated, with the top 10 O*NET tasks now accounting for 33% of traffic, up from 28%. My reading -- and it's only my reading -- is that when companies wire AI into real systems, they keep pointing it at a smaller set of jobs they do over and over, so repetition is where this stuff actually earns its keep.

Four tests to run before you build

Before anyone spends a week building something, I want to know four things, because that's how I'd tell a keeper from a toy.

First, does the question recur? The bastanteo bot handles thousands of queries a year, but if your tool answers a question someone asks maybe twice a quarter, people will have forgotten it exists by the third time it comes up.

Second, is there a source document behind it? BBVA's bot sits on top of approved legal guidance. A tool with nothing behind it is just a chatbot with a nicer name, and people tend to stop trusting it after its first confident wrong answer.

Third, is someone accountable for the answer? BBVA's Legal Services team developed and reviewed the content, and any tool like this needs someone whose job includes updating it when the rules change. Tools nobody owns drift out of date until people quietly stop using them.

Fourth, does it replace a step or add one? Branches used to send these questions to the legal team and wait, and now the common ones get answered on the spot. If your tool means opening one more tab on top of the old process, it's competing with habit, and habit usually wins.

Three out of four is probably fine for most tools, but if something scores zero out of four, what you've got is a demo rather than a tool anyone will keep using.

I don't think those 16,000 are all waste, either. None of my sources say what happened to them, so I'm not going to guess, but in general a lot of abandoned tools are how people work out what's worth building... it's common to make three bad ones before the fourth turns out to be the one your team actually uses.

So I wouldn't tell anyone to stop experimenting -- I'd just stop counting the experiments as the result. What's worth celebrating is the tool that quietly handles more than 9,000 questions a year, not a dashboard showing how many tools got made.

That shape, one workflow with the documents behind it and a team that owns it, is pretty much what we build toward at Workbricks, but you don't need us to run the tests. Pick the question your team answers most often, find the document that already answers it and name the person responsible for keeping it current, and you'll have most of what a keeper needs before you've built anything.

Count the keepers, not the builds.