I keep hearing the same worried whisper in hallway chats and late-night DMs. “Is my boss quietly measuring me against a model?” The short answer is yes, in one way or another. Most leaders will not pit you against ChatGPT in a public duel. They run quieter checks. They compare your output to an AI baseline, nudge the constraints, and watch how you respond when the assignment stops being tidy.
I am not saying every manager is scheming. Some are just trying to keep a team sharp while the tools get faster. But the patterns repeat. If you can spot the tests, you can answer them on purpose.
Before I get practical, a grounding truth. Research on automation risk has been consistent for a decade: routine, predictable tasks are the first to go, while judgment, cross-context reasoning, and relationship work persist longer. One well-known estimate suggested that close to half of tasks across jobs were technically automatable with earlier waves of AI. More recent work on coding assistants showed large productivity gains on well-scoped tasks and smaller gains on messy ones. I read those studies as a simple rule of thumb. If your work looks like a clean prompt and a short feedback loop, a boss can swap a chunk of it with a model. If your work looks like messy people, evolving goals, and open-ended tradeoffs, the math changes.
With that in mind, here are the seven quiet checks I see most often, what each test reveals, and how I answer them in my own work.
Test 1: The procedural trap
What it looks like: You get a highly structured task that could be handled by a careful junior. Fill this template. Draft release notes from a spec. Reformat a dataset. Write a “standard” blog post in a house style. The instructions are crystal clear. The scope is neat.
What your boss is watching: Do you follow directions, or do you add the kind of context a model tends to miss? They might already have run the same input through a model and want to see if your version carries more signal than the best prompt they can write.
Why AI looks strong here: Models shine on procedure. They track lists, mimic patterns, and produce tidy answers fast.
How I answer the test: I complete the request exactly as asked. Then I add a small bonus that reflects judgment. If I am writing release notes, I add a one-paragraph “what this means for users” section tied to a real support ticket. If I am reformatting data, I flag a surprising outlier and suggest a next step. I do not sprawl. I add one punch of value the template did not ask for. That single move separates “pattern follower” from “sense maker.”
Tell-tale sign the test is running: Feedback that fixates on consistency and speed over impact. Phrases like “repeatable,” “template,” or “can we standardize this” often mean they are calibrating a procedural baseline.
Test 2: The context-switch stress test
What it looks like: Mid-task, the constraints change. A stakeholder swaps the audience from engineers to sales. The deadline moves up. The core question shifts slightly. Your boss watches how you respond to ambiguity and speed.
What your boss is watching: Do you hold the thread when the shape of the work changes? Can you keep the outcome stable even when inputs move? Context retention and prioritization are hard to fake.
Why AI struggles here: A model can revise text instantly, but it does not own continuity in the way a person does. It does not carry social memory across meetings, and it does not protect a fragile goal from scope creep unless a human tells it to.
How I answer the test: I restate the goal in one sentence, confirm tradeoffs, and propose a minimum viable path. I name what will slip and what will not. Then I volunteer to absorb coordination pain that keeps others unblocked. I am explicit about what I am protecting and why.
Tell-tale sign the test is running: You hear “ignore what we said earlier, go with this” without a written summary from anyone else. If you provide the summary, you just won.
Test 3: The stakeholder translation drill
What it looks like: You are asked to explain something technical to a non-technical person, or to brief leadership with no time on the agenda. The topic is niche, the room is mixed, and the ask is short.
What your boss is watching: Can you adjust language, tone, and detail level to the person in front of you while preserving meaning? This is not just clarity. It is social calibration.
Why AI struggles here: Models can simplify, but they do not read a room. They do not notice the VP’s frown when the acronym lands wrong, or the sales lead’s silence when pricing risk pops up.
How I answer the test: I use two versions in my pocket. A “tweet-length” core statement with a concrete payoff, and a 60-second story that anchors the concept in a customer moment. Then I ask one question that invites the listener to speak. When they talk, I adapt on the spot.
Tell-tale sign the test is running: The boss says “can you join this call for a quick sanity check” where your presence seems optional. You were invited because translation is the work.
Test 4: The zero-data day
What it looks like: You are asked for a plan or a forecast with missing numbers. “We have no survey yet. Draft the strategy anyway.” Or “Estimate impact with what we have.”
What your boss is watching: Do you propose assumptions transparently, design a way to learn, and size risk without theater?
Why AI struggles here: A model will fill gaps with fluent guesses. That can be helpful for options, but it often reads as overconfidence. Leaders read overconfidence as risk.
How I answer the test: I frame assumptions as levers, not lore. I offer a base case, a low case, and a high case, each tied to a simple driver. Then I propose a quick test that would shrink uncertainty next week. I write the first draft of the spreadsheet myself, even if I later use a model to tidy it.
Tell-tale sign the test is running: You get praise for how you think, not for being “right.” When the boss says “the shape is right,” you passed.
Test 5: The constraint negotiation
What it looks like: An assignment arrives with a deadline that squeezes quality, a scope that is too broad, or a dependency you do not control. It smells like a trap, but the trap is the point.
What your boss is watching: Will you accept bad constraints and rush, or will you renegotiate to protect outcomes? Good leaders want people who can push back with judgment and get to a healthier deal.
Why AI struggles here: A model does not negotiate scope. It will hand you a polished answer to an ill-posed question. That is useful for drafts, not for deals.
How I answer the test: I propose a smaller slice that meets the core need and prove value early. I say what I will deliver by the short deadline and what I will deliver after, with a reason tied to risk and impact. I put it in writing. Leaders notice when you trade scope for certainty instead of trading quality for speed.
Tell-tale sign the test is running: When you offer a principled counterproposal, the boss quickly agrees and thanks you. They were hoping you would ask.
Test 6: The taste test
What it looks like: You get an assignment where “good” is subjective. A landing page for a thorny product. A tagline that must feel human. A dashboard that actually gets used. You are told to “make it sing.” There is no rubric.
What your boss is watching: Taste is shorthand for a mix of pattern knowledge, restraint, and attention to small truths. Do you carry a sense of what feels right for this brand, this audience, this moment?
Why AI struggles here: Models can imitate tone, but they flatten voice. They produce averages. You can prompt around that, and you should, but brand taste usually lives in the seams that are hard to describe. Consistent taste earns trust, which turns into more ambiguous work, which is the safest work to have.
How I answer the test: I bring three options with a one-line creative rationale for each. I include one safe option that proves I understand the baseline, one spicier option that risks a point of view, and one hybrid. I pin each to a real-world reference so the team knows what “good” looks like without vague adjectives.
Tell-tale sign the test is running: Feedback uses words like “feels off” or “closer.” They are not grading grammar. They are gauging your ear.
Test 7: The ownership loop
What it looks like: You are handed a small project end to end. Define the problem, ship the thing, learn from it, and write the follow-up. No one else is chasing it.
What your boss is watching: Do you close the loop without being asked? Do you connect outcomes to changes in how the team works?
Why AI struggles here: A model can help at every step, but it does not hold responsibility. It does not feel the weight of a promise.
How I answer the test: I treat the assignment like a mini business. I define the user, the success metric, the risks, and the next iteration. After delivery, I send a short memo with what we learned, an update to our checklists, and a tiny process fix. I make it repeatable, then I move on.
Tell-tale sign the test is running: No one schedules a retro. If you do, you become the person who closes loops. That is very hard to replace.
The seven tests at a glance
| The test | Boss move | What they are watching | AI’s edge | Your human edge |
| Procedural trap | Template-heavy tasks | Accuracy, speed, plus one insight past the template | Fast pattern mimicry | Add context, spot risk, carry customer memory |
| Context-switch | Mid-course changes | Goal retention, prioritization | Quick rewrite | Protect outcomes, manage tradeoffs, keep continuity |
| Stakeholder translation | Mixed audiences | Social calibration, persuasion | Simplification | Read the room, adjust live, build trust |
| Zero-data day | Plan with gaps | Assumptions, learning plan | Fluent guesses | Structure uncertainty, design tests, surface risk |
| Constraint negotiation | Unrealistic brief | Principled pushback | Polished draft to bad ask | Redesign scope, tie to impact, set terms |
| Taste test | Subjective quality | Brand ear, restraint | On-trend imitation | Distinct voice, reference points, restraint |
| Ownership loop | End-to-end task | Accountability, iteration | Tool at each step | Responsibility, process improvement, follow-through |
How bosses benchmark you against AI without saying it
I have watched managers quietly run A/B tests. They give a model the same source materials and see how far it gets with careful prompting. They are not trying to humiliate anyone. They are trying to price a baseline. A few patterns signal this is happening.
| Signal you might be in an AI benchmark | What it often means | Healthy response |
| The ask arrives with unusually clean inputs and a very strict structure | They want to compare outputs apples to apples | Match the structure, then add one human insight past the ask |
| Feedback centers on “consistency,” “style rules,” and “speed” | They are calibrating a template for repeatable work | Offer a faster path that protects quality, or automate the repeatable part yourself |
| You are asked for 3 variants with small differences | They are measuring whether variety you provide beats a model’s sampling | Make each variant grounded in a real user scenario, not just wording |
| You hear “let’s see what the tool says” in the same meeting | They are testing live | Treat the model as a collaborator. Use it to widen options, then lead the decision |
| Sudden interest in your process, not just your result | They want to know what cannot be templated | Narrate judgment points, dependencies, and social work you performed |
What the research says, in plain English
When I explain AI risk to teams, I keep the research simple and human.
- One famous estimate suggested that a large share of tasks within jobs are automatable, especially those with routine rules and predictable data. I read that as a reminder. If your day is mostly structured handoffs, you are at risk.
- Studies of coding assistants found big gains on well-scoped problems for intermediate developers. The effect shrinks when tasks are novel or when requirements change mid-flight. That echoes what I see in practice.
- There is also a line of work on “algorithm aversion” that showed people lose trust in models after seeing them make errors, yet other work found that when you give people a little control or transparency, adoption rises. Translation for us. The more you show your reasoning and bring stakeholders into your process, the more your judgment will be preferred over a black box.
I use those three findings to prioritize. I reduce the share of my day that could be a clean prompt. I increase the share that depends on relationships, live judgment, and original data.
How I make myself harder to replace by AI
I move toward interfaces. People who translate across functions survive change. I look for seams where marketing meets product, or research meets sales. That work always includes nuance that tools cannot see on their own.
I create artifacts that outlast the task. Checklists, examples, and one-page explanations save teams time. I write them as I work. If a model can do the base work, I own the standards.
I make the messy parts visible. I narrate the parts of my process that do not show up in the final asset. The stakeholder I convinced. The hidden constraint I removed. The decision I made because of history the deck does not capture.
I involve users early. When I bring a real quote, a support ticket, or a sales transcript, the conversation moves from taste to truth. Models can simulate users. They cannot look someone in the eye.
I leverage models in public. I do not pretend I am not using AI. I show how I use it, where it helps, and where I stop. That positions me as the person who gets leverage without losing judgment.
A one-week plan to raise your “human premium”
| Day | Practice | What you do | Outcome |
| Monday | Map replaceable tasks | List your weekly tasks. Tag any that could be a clean prompt. | You see your risk surface |
| Tuesday | Build a “plus-one” habit | For one templated task, add a single insight past the ask. | Small, visible value past procedure |
| Wednesday | Stakeholder coffee | Ask one partner what “good” looks like to them and why. | Better taste, real references |
| Thursday | Write a mini standard | Turn a repeated fix into a 10-line checklist. Share it. | You own the standard, not just the task |
| Friday | Run a micro-experiment | Pick one assumption and test it in 48 hours. Report back. | You reduce uncertainty and show initiative |
Repeat weekly. The compounding effect is real.
How to respond to each test in the moment
| Test | First sentence I say | First artifact I create |
| Procedural trap | “I will match the template exactly and add one note on impact.” | Final deliverable plus a short “so what” |
| Context-switch | “Let me restate the goal so we keep the target steady.” | A one-paragraph goal and tradeoffs note |
| Stakeholder translation | “In one line, here is the value. Then I have a 60-second example.” | Two-tier brief, a question for the room |
| Zero-data day | “Here are the three assumptions driving the range. I will test one now.” | Simple scenario table and a scrappy test plan |
| Constraint negotiation | “I can deliver X by Friday that solves the core use. Y comes next.” | A scoped milestone and a small demo |
| Taste test | “Here are three directions with references for each.” | Options grid with rationale and links |
| Ownership loop | “I will own the loop and send a one-pager on what we learned.” | Short post-mortem plus a tiny process change |
Real-world frictions that help your case
I keep a running list of small, real-world details that a model does not capture well and that a boss actually values.
- Institutional memory. You remember why the pricing page avoided a certain phrase after that legal issue last spring.
- Live constraints. You know the only designer free next week is also on the iOS release. Your plan respects that.
- Soft landings. You schedule the rollout on Tuesday because the support team’s busiest day is Monday.
- Relationship credit. Sales will actually push your deck because you involved two reps in the early draft.
- Taste for the brand. You can feel the difference between “clever” and “on message” because you have sat with users and heard how they talk.
These details are small on paper. They matter in shops where trust is currency.
What to say if your boss asks “why not just use AI for this”
I do not get defensive. I stay specific. I say, “Let’s use the model for the first draft and for research spread. I will own the decision points, the translation to stakeholders, and the final call on tone. If the draft clears 80 percent of the work, we both win. If it does not, I will show you where it failed and fold that into our checklist.”
That frame treats AI as leverage and you as the owner. Leaders want owners.
A brief checklist for leaders who want to keep humans in the loop
If you are a boss reading this, here is what I ask my own teams to do.
- Write the goal, not just the task.
- Benchmark with models honestly, then publish the standards so the whole team benefits.
- Reward people for closing loops, not just shipping assets.
- Push repeatable work into tools and lift people into judgment work.
- Capture taste with examples, not adjectives.
- Celebrate the person who fixes a constraint before it breaks the team.
When leaders manage the system with that spirit, these tests feel less like traps and more like a ladder.
The quiet conclusion I live by
AI is already inside our work. That is not a threat by itself. The threat is becoming the person who only ships what a clean prompt can produce. I avoid that fate by leaning into judgment, translation, and ownership while still squeezing every drop of speed from the tools. Bosses are running subtle tests to see who can do both. I try to be the person who answers each test with clarity, then writes down the play so the team gets faster together.
If you remember nothing else, remember this. Treat every neat, templated ask as an invitation to add one honest inch of human value. Then, on the messy work, take the wheel. That mix is very hard to replace.