Most companies already use ChatGPT and Claude.
Take a board forecast. Someone pastes in a spreadsheet, a Slack thread, and a paragraph of context.
The model helps, until reality comes back in. ARR is using last quarter's definition. The spreadsheet is not the version in the board folder. The Slack thread has an exception for one enterprise customer. The formula needs to change, the deck needs different framing, and the update still has to go through approval.
The model got smarter. The workflow stayed manual.
That is the part people underrate.
To use AI well, you have to give it the same information a good employee would need, often more. It has to know that "ARR" means the board-approved definition, not the dashboard default. It has to know which spreadsheet wins, which Slack exception still applies, and which approval path matters.
Then you have to ask whether the pipeline is any good. Is it fast? Permission-aware? Does it pull the right context without flooding the model? Does it write back cleanly?
Search, permissions, parsing, citations, evals, review, artifacts, writeback, and memory are all hard on their own. The product is making them work together efficiently.
When those pieces live in the user's head, every run starts over. The smart employee becomes the retrieval layer, policy layer, QA layer, writeback layer, and memory layer.
Good employees learn source precedence, manager taste, approval paths, customer exceptions, broken reports, and the difference between technically correct and accepted.
AI has to learn the same way: from what gets used, rejected, and approved.
A fixed clause, rejected source, changed formula, rewritten account brief, or renamed heading should not disappear into chat history. It should become captured correction: operating memory, precedent, a better starting point next time.
The useful object is not a paragraph. It is the changed deck, the corrected formula, the evidence trail, the approval, the writeback, and the memory for next month.
Buying AI seats gives every employee a faster engine. But a Ferrari engine in a Jeep is not going to win an F1 race. The rest of the machine matters: documents, context, existing systems, review, writeback, and memory.
Doe upgrades that machine around the work. The more work moves through it, the more it knows how your company works. Next month's forecast should not start from a blank chat. It should start from the deck, the formula, the approval, and the correction that already happened.