How IT Departments Accidentally Locked Their Companies Into the Wrong AI—And What It's Costing Them
Somewhere around 2023, a decision got made in thousands of American companies. Not a careful, deliberate decision backed by pilot programs and capability benchmarks. More like the kind of decision that happens when a VP forwards a TechCrunch article and someone in IT says, "We already pay for Microsoft 365, so let's just turn on Copilot."
Fast forward to today, and a lot of those companies are sitting with enterprise AI contracts they can't fully justify, workflows that kind of work but not really, and zero visibility into whether they actually chose the right tool—because they never compared it to anything else.
This is the enterprise AI trap. And it's more common than anyone wants to admit.
The Microsoft Effect (And Why It's Not a Compliment)
Let's be honest about how most large organizations ended up with ChatGPT or its Microsoft-wrapped cousin, Copilot. It wasn't a procurement process. It was gravity.
Microsoft has deep roots in corporate IT. The licensing relationships, the existing infrastructure, the familiar interfaces—it all creates a path of least resistance. When AI became impossible to ignore, IT departments didn't have to go looking for a solution. One landed in their inbox as an add-on to software they were already paying for.
That's a brilliant distribution strategy for Microsoft. It's a potentially costly shortcut for the companies that went along with it.
The problem isn't that ChatGPT or Copilot are bad tools. They're not. The problem is that "good enough to demo well in a conference room" and "right for your specific workflows" are two very different things. And most enterprise AI rollouts never got past the first evaluation.
What a Real Capability Evaluation Would Have Revealed
Here's what tends to happen when companies actually run structured comparisons instead of defaulting to the household name:
Legal and compliance teams often find that Claude, built by Anthropic, handles nuanced document analysis and policy interpretation with more caution and precision than ChatGPT. Anthropic's focus on AI safety translates into outputs that are less likely to confidently state something wrong—which, in legal contexts, is a feature worth paying for.
Developers who get handed a ChatGPT enterprise license frequently end up using GitHub Copilot or Cursor on the side anyway, because those tools are built around coding workflows in ways that a general-purpose chatbot simply isn't. The enterprise license pays for something the dev team doesn't actually use.
Research and knowledge work teams often get more mileage from tools like Perplexity, which grounds responses in real-time sourced information, compared to a model that may be working from training data that's already months or years stale.
None of this means ChatGPT loses every comparison. But it does mean that "we went with the one everyone's heard of" is a procurement strategy that was always going to produce mixed results.
The Hidden Costs Nobody Put in the Budget
When companies talk about the cost of their AI tools, they usually cite the licensing fee. That's the visible number. The invisible ones are harder to track but often larger.
Prompt engineering overhead. General-purpose models require more careful, detailed prompting to produce useful outputs in specialized domains. Teams end up spending hours crafting and refining prompts that a purpose-built tool would handle out of the box. That's real labor cost, and it compounds.
Output verification time. When an AI tool has a reputation—or a track record in your organization—for confident inaccuracies, people spend more time checking its work. That's fine as a habit. It's expensive as a systemic pattern.
Shadow AI adoption. When the official enterprise tool doesn't actually meet team needs, people find workarounds. They start using personal Claude accounts, running queries through Perplexity, or pasting company data into tools that haven't been vetted by IT. This creates exactly the kind of data governance problem the enterprise license was supposed to prevent.
Opportunity cost. Every month a team spends using a mediocre tool is a month they're not getting the productivity lift that a better-matched tool would provide. That gap doesn't show up on a balance sheet, but it's real.
Companies That Went Back and Did the Work
Some organizations have started running the comparisons they skipped the first time. The pattern that emerges is consistent: the more specialized the use case, the more it matters which tool you pick.
A mid-sized law firm that had deployed Copilot across the organization ran a three-month parallel test with Claude for contract review tasks. Their finding wasn't that Claude was universally better—it was that Claude's tendency to flag ambiguity and hedge uncertain interpretations made it significantly more useful for that specific workflow. The firm now runs both, with Claude handling document analysis and Copilot covering general drafting and scheduling tasks.
A regional healthcare network that had standardized on ChatGPT for administrative work found that when they tested specialized medical documentation tools, turnaround time on clinical notes dropped substantially. The general-purpose model had been workable. The purpose-built alternative was actually good.
These aren't indictments of ChatGPT. They're illustrations of what happens when you match the tool to the job instead of the job to the tool.
The Lock-In Problem Is Real, But It's Not Permanent
One of the reasons companies stay with suboptimal AI setups is that switching feels complicated. There's training to redo, integrations to rebuild, and the organizational inertia of convincing people to change software they've already gotten used to.
But the lock-in here is softer than it looks. Unlike legacy enterprise software with years of custom development baked in, most AI tool usage is relatively portable. Prompts can be adapted. Workflows can be rebuilt. The switching cost is real but manageable—especially compared to the ongoing cost of using the wrong tool at scale.
The first step is running an honest audit. Which teams are actually getting value from the current setup? Which ones have built workarounds? Where are the verification bottlenecks? What's the actual usage rate of the enterprise license versus what was projected?
That audit tends to be uncomfortable. It also tends to be clarifying.
The Smarter Move Going Forward
The enterprise AI market is no longer a one-horse race, even if the marketing budgets make it look that way. Anthropic, Google, Mistral, Perplexity, and a growing list of domain-specific providers are building tools that compete seriously with OpenAI's offerings—and in specific contexts, outperform them.
The companies that will get the most out of AI over the next few years aren't the ones that moved fastest in 2023. They're the ones willing to do the less glamorous work of actually evaluating what they've got, comparing it to what else exists, and making decisions based on evidence rather than brand familiarity.
Your company chose ChatGPT because everyone else was. That might have been a reasonable call at the time. Whether it's still the right call is a question worth asking—with data, not assumptions.