OpenAI’s September 10 release of ChatGPT for Financial Services points to a practical shift in enterprise AI: the winning product is not the most flexible blank canvas. It is the one that removes setup from the work your team already has to do.
That matters because most AI projects do not die at the moment of model selection. They die in the space between a promising answer and a usable workflow. OpenAI is packaging financial data, connected sources, and analysis workflows into the product because that is where the work breaks. The lesson for business leaders is simple: if your AI plan depends on every team building its own connective tissue, adoption will stay fragile.
TL;DR
AI adoption improves when setup disappears from the user’s day. Before buying another general-purpose tool, pick one recurring decision, map the data required to make it, name who trusts the output, and measure whether the workflow gets faster or cleaner under normal business conditions.
The product is moving closer to the work
The old software pattern was access first, workflow later. Give people a tool, train them, and hope the useful process emerges.
That pattern is weak for AI.
With AI, the gap between access and value is wider. A finance team does not only need a chatbot that can reason over numbers. It needs the right data available, the right sources connected, the right permissions set, the right review loop in place, and the right confidence level before anyone uses the output in a forecast, memo, diligence process, board packet, or client recommendation.
OpenAI’s product language is telling. The company describes ChatGPT for Financial Services as bringing together built-in data, connected sources, and financial analysis workflows. That is not a model announcement. It is a packaging decision.
The product is admitting something leaders should notice: the hardest part of AI at work is often not getting an answer. It is making the answer land inside a workflow people already trust.
Setup is adoption friction
Every extra step between intent and useful output taxes adoption.
Someone has to find the right file. Someone has to check whether the number came from the current version. Someone has to format the output into the template the business actually uses. Someone has to explain why the answer is safe enough to use.
That work is usually invisible in an AI demo. It is painfully visible in production.
MIT’s NANDA research found that 95 percent of enterprise generative AI pilots were failing to produce measurable return. That number should not be read as “AI does not work.” It should be read as “most organizations are still bad at turning capability into operating change.”
Setup is one of the places that failure hides.
If the analyst has to rebuild the same context every time, the AI becomes another tab. If the manager has to verify every source manually, the AI becomes another risk surface. If the output does not match the format the organization already uses to make decisions, the AI becomes a novelty layer on top of the real work.
The useful implementation question is not “can the AI do this task?” It is “how much work does a normal employee have to do before the AI can help in a way the business trusts?”
Do not confuse flexibility with usefulness
Leaders like flexible tools because they appear to preserve optionality. Buy one platform, let every department find its own use cases, avoid overcommitting too early.
That sounds reasonable. It often creates a quiet mess.
A blank canvas is only useful when the team has time, skill, and ownership to turn it into a working system. Most teams do not. They have meetings, deadlines, reporting cycles, customer escalations, month-end work, and vendor cleanup.
A narrower tool with the right data and workflow shape can beat a broader tool that requires every user to become an AI product manager. The business impact does not come from theoretical range. It comes from repeated use inside a high-value process.
For financial services, that might mean faster company research, cleaner market summaries, or less time assembling background material before judgment begins. For another business, the same principle could apply to sales handoffs, contract review, support escalations, or client reporting.
What leaders should do this week
Pick one recurring decision that already matters.
Not a vague department goal. Not “improve productivity.” Pick a decision with a visible output: approve a vendor, prioritize accounts, brief a client, update a forecast, prepare a renewal, review a contract, or decide which exception needs human attention.
Then map five things.
What data does the person need before making the decision? Where does that data live today? What format does the final output need to take? Who reviews or relies on it? What would prove the workflow improved after two weeks?
That map will tell you whether AI is ready to help.
If the data lives across five disconnected places, solve the access problem first. If nobody agrees what a good output looks like, write the standard before adding AI. If the review owner is unclear, the workflow is not ready. If the success measure is “people used the tool,” the project is already too soft.
Measure the operating result instead. Time saved in prep. Fewer handoff errors. Faster first draft. Shorter review cycle. Less rework after manager review.
OpenAI’s financial services release is worth noticing because it points away from AI as a general assistant sitting beside the work and toward AI as a structured participant inside the work.
That is where the next adoption gap will open.
Some organizations will keep buying flexible AI access and wondering why usage plateaus. Others will keep removing setup from the workflows that already drive business value.
The second group will look less experimental. They will also get more done.
Sources: OpenAI, “Introducing ChatGPT for Financial Services,” September 10, 2026; MIT NANDA, “The GenAI Divide: State of AI in Business 2025.”