The Secret Life of AI: The Expectation Gap
.jpeg)
The Secret Life of AI: The Expectation Gap
How to prompt, think, and get results from any AI tool
#WorkingWithAI #Prompting #AIConfidence
Margaret is a senior software engineer. Timothy is her junior colleague. They work in a grand Victorian library in London — and in every episode, they'll show you exactly how to get what you want from AI.
Episode 2
Timothy did not bother unwinding his scarf this time. He sat down, dropped a folder on the table between them, and looked at Margaret with the expression of a man who had been proven right about something he hadn't wanted to be right about.
"It failed," he said.
Margaret looked at the folder. Then at Timothy. "Tell me."
"The quarterly report. I gave it everything — the raw data, the previous report for reference, the client's name, the deadline. I told it exactly what I needed." He pushed the folder toward her. "What came back was plausible. Polished, even. It looked like a report. It sounded like a report." He paused. "Half the figures were wrong. The conclusions didn't follow from the data I'd provided. And I nearly sent it."
"But you didn't."
"I caught it at midnight. An hour before I was going to bed." He leaned back. "I trusted it, Margaret. Completely. I handed it the wheel and walked away."
Margaret was quiet for a moment. She did not reach for the folder.
"Yes," she said finally. "You did. Though perhaps a better image than a wheel is this — you handed a very capable junior colleague a complex assignment, gave no further instruction, and expected a finished product on your desk by morning. No check-ins. No review. No questions asked."
Timothy considered that. "When you put it that way —"
"It sounds like a management problem as much as a technology one," Margaret said. "Because it is."
What Over-Reliance Actually Looks Like
"I'm not saying this to be harsh," Margaret continued, "but I want you to hear it clearly. What you've just described is not a failure of the AI. It is a failure of the handoff." She folded her hands on the table. "When you give a junior colleague a complex task and walk away entirely — no check-ins, no review, no questions — and they return something flawed, who bears responsibility for the outcome?"
Timothy said nothing.
"You gave the AI everything it needed to produce something that looked right. And it did exactly that. It produced something that looked right." She tapped the folder once. "Looking right and being right are two entirely different things. And only one of you at this table is equipped to tell the difference."
"So I'm supposed to check every single thing it produces?"
"You are supposed to remain in the loop," Margaret said. "There is a difference. You do not need to verify every comma. But a quarterly report going to a client? The figures, the conclusions, the narrative — yes. That requires your eyes. Your judgment. Your domain knowledge." She paused. "The AI does not know your client. It does not know what the numbers should feel like based on last quarter's performance and the conversation you had with the account manager on Tuesday. You do. That context lives only in your head."
The Expectation Gap — And Where It Comes From
She rose and moved to the window, looking out at the grey London morning. "The expectation gap is the distance between what you believed the AI would do and what it actually did. In your case, you believed it would function as a reliable, self-sufficient analyst. What it actually is — what it has always been — is a very capable first drafter that requires human oversight to cross the finish line."
"Nobody told me that," Timothy said.
"Nobody tells anyone that." She turned back from the window. "The companies that build these tools have every incentive to emphasize capability and very little incentive to emphasize limitation. So the marketing says extraordinary and the documentation says powerful and the demos show the best possible outcome under ideal conditions." She returned to her chair. "And then you sit down with a real problem, in real conditions, with real stakes — and the gap opens up."
"So what do I do with that gap?"
"You close it," Margaret said simply. "Not by lowering your ambition for what AI can help you achieve — but by recalibrating your expectations to match what the tool actually is. Once you understand what it genuinely does well, and where it reliably falls short, you stop being surprised. You stop being burned. You start using it correctly."
What AI Is Actually Good At — And Where It Falls Short
She pulled a sheet of paper from beneath her book and set it between them. "Let me be direct about this, because it will save you a great deal of frustration."
"AI is remarkable at generating structure. First drafts. Summaries. Rewording. Brainstorming. Explaining complex ideas in plain language. Spotting patterns in text. Writing code that follows common patterns. Tasks where a reasonable attempt is most of the value." She looked at him steadily. "It is unreliable when precision is the entire point. When the figures must be exactly right. When the reasoning must be airtight. When the output will be used without review. When the stakes of a wrong answer are high."
She paused, then added: "And remember what we established last time — plausible and accurate are not the same thing. The AI does not retrieve facts from a verified source. It generates what a correct answer should look like. For anything numerical, factual, or verifiable — check it. Every time."
"That describes most of my actual work," Timothy said quietly.
"Which means AI is your collaborator, not your replacement," Margaret said. "It handles the parts of the work where a strong first attempt is valuable. You handle the verification, the judgment, the final call. That is not a limitation to resent. That is simply how the collaboration works."
How to Set Expectations Before You Start
"So how do I get this right going forward?" Timothy asked. "Before I start a task — what should I be asking myself?"
Margaret considered this. "Three questions. First: what is the cost of this being wrong? If the cost is high — client-facing, financial, legal, reputational — your review must be proportionally thorough. The AI's confidence in its output has no relationship to its accuracy. A wrong answer arrives with the same smooth tone as a right one."
"Second: am I giving it enough context to succeed? The gap between a vague prompt and a precise one is enormous. Not just the task itself, but the constraints, the audience, the format, the tone, the specific outcome you need. The more context you provide, the narrower the space in which it can go wrong." She held up a finger. "One practical note here — AI works only with what is in the current conversation. It has no memory of previous sessions, no access to your files unless you paste them in, no knowledge of your organisation, your client history, or your industry norms unless you provide them. Every session starts from zero. If you assume it remembers something from last week, you will be disappointed." She paused. "Some platforms are beginning to introduce memory features that carry context between sessions. But even then — never assume it holds the right context. Verify it. The principle does not change."
"And third: am I staying in the loop?" She looked at him directly. "Not hovering. Not second-guessing every sentence. But checking in. Reading what comes back with your full attention. Asking yourself — does this match what I know to be true? Does this feel right? Would I be comfortable putting my name on this?"
Timothy picked up the folder and looked at it differently now. "I didn't ask myself any of those."
"No," Margaret said. "You handed it the wheel and trusted it to know the road." She paused. "It doesn't know the road, Timothy. It knows roads in general. The specific road — your client, your data, your context — only you know that."
"So the quarterly report — how would I approach it now?"
Margaret nodded, as though she had been waiting for exactly that question. "You give it the structure. Ask it to draft the narrative sections — the executive summary, the context, the commentary. That is where a strong first attempt is genuinely valuable. The figures you pull yourself from the source data and insert directly. You read the conclusions it draws and ask — do these follow from what I actually know about this client? Then you put your name on it." She looked at him steadily. "Three steps. Draft, verify, judge. In that order. Every time."
The Recalibration
He was quiet for a long moment. Outside, a pigeon landed on the windowsill, regarded the library briefly, and departed.
"I think I expected it to be smarter than it is," Timothy finally said.
"Most people do, at first. And then something goes wrong and they decide it is more foolish than it is." Margaret closed the folder and slid it back toward him. "The truth is somewhere between the two. It is genuinely capable, within specific boundaries. Your job is to learn those boundaries — not from the brochure, but from your own experience, over time. Every task you give it teaches you something about what it can and cannot do. That knowledge compounds."
"Patience and flexibility," Timothy said.
"And now — calibrated expectations." She picked up her book. "Come back next week. We will talk about what to do when you hit a wall — when the AI keeps giving you the wrong thing no matter how many times you try. Because that will happen too. And there is a way through it."
Timothy stood, finally unwound his scarf — then wound it back again, the morning still being what it was — and headed for the door.
"Margaret."
She looked up.
"The report. I fixed it myself. Took three hours."
"I know," she said. "But you knew it was wrong. That is not nothing."
He nodded once and left.
She listened to his footsteps fade down the corridor, then returned to her page.
Next episode: The Art of the Prompt — not a technical tutorial, but a human skills episode. What you'd tell a very smart new colleague on their first day, and why the quality of your question determines everything about the quality of your answer.
Aaron Rose is a software engineer and technology writer at tech-reader.blog.
Catch up on the latest explainer videos, podcasts, and industry discussions below.
Comments
Post a Comment