PAID CONTENT brought to you by
BACK TO HOME
Stop blaming the AI if your company’s data is stale
Executives blame algorithms for failed AI projects, but Vaidotas Juknys, CEO at Decodo, says the real culprit is stale, incomplete web data feeding flawed models daily.
Most AI failures don’t start with the model. Take a coding agent pulling documentation to write an application programming interface (API) integration. If it’s reading a cached snapshot of those docs from six weeks ago, it will generate calls against endpoints that have since changed or been deprecated, and the code will look completely reasonable until it actually runs. When it breaks, the model gets blamed. Almost nobody asks how old the documentation was.
That difference between what a model believes and what’s true right now is where most AI failures live. I run Decodo, a company that provides a web data access layer for fast-growing tech teams, so I end up seeing the progress up close. More of these teams are building a live data connection into the product from day one instead of adding it after launch.
The scale of it shows up in the data too. Every model has a training cutoff, so a coding assistant is, by definition, working from a documentation snapshot that’s already out of date the day it ships. Gartner’s own research on AI code assistants found that 42% of engineering teams see productivity gains of only 1% to 10% from these tools, with another 12% seeing no gain at all. The Stack Overflow Developer Survey found that trust in AI-generated code accuracy dropped from 40% to 29% in a single year, even as adoption kept climbing. Plenty of things explain this issue. But of all of them, only one never gets audited—how old the data feeding the tool actually is.
What the reviews miss
When something goes wrong, engineers argue about the model. Was the prompt off? Was the wrong model used for the job? Almost nobody checks where the agent got its information from.
That’s where it breaks. A dependency list or a page of documentation, pulled once when the integration was built, never looked at again. The agent keeps reading from it anyway, months later, as if it were still current.
An agent pulling package documentation to recommend an install command will suggest a version that has since been deprecated or pulled entirely, and the suggestion will look completely normal until the build fails.
Data doesn’t rot at one speed
Most companies treat all their data the same way, one refresh schedule applied across everything they collect. That flattens something important because information decays at wildly different rates depending on what it is. A competitor’s price can go stale within hours, while a product review tells you something useful a week later, and a regulatory filing might hold steady for months before anything changes.
I think of it as a half-life—price data measured in hours, and catalog structure sometimes stretching out for months before anyone touches it.
Very few companies tier their refresh cadence to match. They apply one policy to everything, which means the fast-moving stuff goes stale while the slow-moving stuff gets refreshed for no reason, burning budget in the wrong place. A business selling into a fast-moving retail market needs its price data refreshed by the hour. A business tracking industry regulation might be fine checking once a month. Treating both the same wastes effort on one and risks the business on the other.
The budget line nobody protects
Every AI cost review looks the same. Teams check which model they’re using, whether a cheaper one would work, and how many tokens a task actually burns through. That’s where the savings usually are, so that’s where the effort goes.
The web data feeding all of that barely gets looked at, and when it does, it’s compared with the model bill and looks small by comparison. Nobody’s tracking what that spend actually protects, so it just reads as a line that isn’t doing much. When the budget review turns into a cost-cutting exercise, it’s usually the first thing to go.
A smaller, cheaper model running on fresh data will beat a bigger model running on stale data most of the time, but stale data doesn’t show up on an invoice the way a high token count does. Teams optimize the part they can see and cut the part they can’t, and six months later the outputs are worse without anyone connecting it back to a decision made in a budget meeting.
What I keep telling teams
Most founders can tell me their models’ accuracy score from six months ago, but almost none can tell me the average age of the data feeding it right now.
Take a pricing agent that reprices against competitors every morning. If it’s reading a saved copy of those competitor pages instead of the live one, it will keep making confident decisions off information that is already wrong, and nothing in its output will look off until someone goes and checks the source.
That happens because nobody owns the refresh. Someone pulls the competitor pages once, when the project ships, and the job of keeping that current never gets assigned to a person, so it just doesn’t happen.
The fix here isn’t a smarter agent—it’s replacing that saved copy with a live connection so the pricing decision each morning is based on that morning’s data instead of whatever was true when the project launched.
That’s what our Web Data API exists to do—give a developer one live connection to whatever part of the web an agent depends on, so nobody has to remember to rerun a script by hand. It handles the operational complexity of collecting publicly available web data reliably, at scale, so the team building the agent doesn’t have to.
Note: This content was created by Decodo, a client of Business Reporter.
BACK TO HOME
The bottom line
I don’t think AI is overhyped. I think most companies are underinvesting in something unglamorous, keeping the data behind their systems as fresh as the decisions those systems are making.
The next AI project that stalls at your company will probably not fail on the science. Somebody will eventually trace it back to a feed that’s gone stale, still running on whatever schedule it was set up with on day one.
If your AI workflows need access to real-time data, it’s about time you discovered Decodo’s Web Data API.
Get started for free with up to 10,000 monthly requests here.