Five weeks of research. Forty-seven AI conversations. A hundred and eighty pages of exported output, all of it saved, tagged, and organized.
Her boss asked one question and none of it helped.
What happened in that Thursday review didn’t cost Priya Raghavan her job. It cost her something slower and harder to win back: the assumption, held quietly by everyone above her, that her work could be trusted without checking.
But it didn’t have to happen this way.
A big opportunity
Priya Raghavan is a senior strategy analyst at a 600-person industrial software company. Six years in. The person you hand the hard question to.
In February, the exec team gave her the biggest question on the roadmap: should the company enter the European market next year, starting with Germany, or double down on the domestic mid-market where it already wins? A recommendation was due to the leadership team in April.
She did what any capable analyst does now. She opened a chat window and started working.
And the work went well. Genuinely well. She pulled competitive landscape analysis, regulatory requirements, channel partner structures, pricing benchmarks, localization costs, three different market sizing approaches. When a thread got long she started a fresh one so the context stayed clean. She saved everything. She built a naming convention by week two.
Forty-seven threads. A hundred and eighty pages. She synthesized it into a forty-page recommendation with an executive summary, and it was, by any standard she’d been taught to apply, the most thorough piece of research she had ever produced.
She sent it to her SVP of Corporate Development on a Tuesday for a Thursday review, one week ahead of the exec presentation.
The day it all fell apart
He’d read the whole thing. He started by saying so, and he meant it as a compliment.
Then he asked what she thought they should do.
Priya started describing what the analysis pointed toward. He stopped her and said no — not what the report says. What do you think. She heard the distinction and reached for an answer and found, with a lurch she still remembers, that she didn’t have one. She had a recommendation. She did not have a position. She had never noticed the difference before, because for five weeks nothing had required her to.
She recovered enough to argue the case. It went fine for a few minutes.
Then he pointed at a number on page eleven: the eighteen-month certification and channel-readiness timeline that the entire go-to-market sequence rested on. Where did that come from?
She said she’d pull the source.
He asked if she was confident in it.
And Priya said the sentence that ended the meeting, though it took her until the drive home to understand that it had. She said it had come up consistently across her research.
He didn’t say anything cutting. He didn’t need to. They both sat there for a second while the words rearranged themselves in the air: consistently across her research meant the same model had told her the same thing five times, in five slightly different sentences, in five separate threads, and she had read that repetition as corroboration. Five sources, one source. She’d never once put those five statements side by side, because they lived in five different conversations and she had no surface on which to lay them next to each other.
Nobody escalated anything. That’s not how this works.
The recommendation went to an outside consultancy for validation. It cost $120,000 and six weeks to check work she had already done. Priya was assigned to support them, which meant spending September assembling context for people being paid to second-guess her. The Europe decision slipped two quarters. When the next big open-ended question came up in the fall, it went to a colleague, and her SVP framed that as protecting her bandwidth.
The reputation that formed wasn’t her analysis was wrong. As far as anyone knows, it wasn’t. It was worse than that: she produces a lot of material. That’s a label with no deliverable attached to it, which means there’s nothing she can fix and resubmit.
The private cost was steeper. She stopped trusting her own output. Every deck she built after that carried a low hum of how would I defend this if someone pushed, and the honest answer was usually that she wouldn’t know how.
The maddening part is that most of her research was probably fine. Some of it was excellent. She had no way to tell which was which.
How mind mapping would have changed the outcome
The failure happened in the first hour, when she opened a chat window instead of a blank map.
That single choice handed the model authorship of her structure. Her first prompt was broad, the response came back organized into neat sections, and every one of the forty-six threads that followed went deeper inside a frame she had never chosen and never examined. By week three she wasn’t researching a question anymore. She was elaborating an outline that a machine had produced in nine seconds.
That’s why she couldn’t audit it. She was checking the model’s answers using the model’s own categories, which is a closed loop. Nothing can look out of place in a structure derived from the very claims you’re testing.
Map it first and the loop breaks open.
Before the first prompt, you spend twenty minutes building your first-level topics from your own question and your own domain knowledge. Six to nine branches: regulatory path, channel economics, competitive response, localization cost, pricing, internal capacity, exit conditions. It takes twenty minutes and it is the single highest-leverage thing you will do in the entire project, because that skeleton is now an independent standard. Everything the AI produces has to earn a place on a structure it did not build.
Then, as material comes in, every claim gets placed — one claim, one branch, one note behind it recording where it came from and whether a human being ever verified it. This costs seconds in the moment and is impossible to reconstruct five weeks later. Ask Priya on page eleven.
What a mind map reveals
Once it’s all on one surface, three things become visible that a hundred and eighty pages of prose will hide from you until you’re sitting across from your boss.
Redundancy stops looking like agreement. The eighteen-month timeline, appearing five times in five voices, collapses on a map into one node with five pointers to the same origin. The illusion of consensus doesn’t survive proximity. That’s not a discipline problem Priya had; it’s a format problem. Fluent, confident, sequential text is exactly the medium in which repetition impersonates evidence.
Contradiction surfaces. Thread twelve assumed a consolidating market. Thread thirty-one assumed a fragmenting one. Both were plausible, neither was flagged, and they sat a hundred and forty pages apart. Put them on adjacent branches and you can’t not see it.
Gaps become visible. And the empty branch becomes the finding. This is the one that matters most, and it’s the reason none of this can be automated. AI will never tell you what it failed to address. It has no mechanism for reporting its own gaps, and it will answer the question you asked with the same confident tone whether or not that was the right question. But a first-level topic sitting there with nothing underneath it is a visible, structural, undeniable hole. Priya’s exit conditions branch would have been bare. In forty-seven threads she never once asked what would tell them the European bet had failed. The exec team would have asked. They always do.
There’s a version of that Thursday where he points at page eleven and she can say exactly where the number came from, which parts of it she confirmed independently, which parts she’s holding loosely, and what she’d need to see to change her mind.
That’s not a better report. That’s a person who owns her own thinking, in a room full of people who can tell the difference.
The tools got very good, very fast, at generating answers. They did not get any better at telling you what your answer is missing.
That job didn’t go away. It just got much bigger, and much lonelier, and almost nobody has been given a method for doing it.
Leave a Reply