irreplaceable

Irreplaceable › Guides › Pushing back on stakeholders

Influence · Product design

How to push back on stakeholders, without becoming the blocker

AI raises the pressure to ship. Here are 26 of the requests designers get, from “replace search with an AI assistant” to “skip design reviews and ship what AI makes”. For each: what you know, the replies you could give, and why the strong one works. Decide yours before you open them.

By Sarit Elisha, founder of Irreplaceable · Updated October 10, 2026

Practice it in a drill

Reply in a live conversation and see how the stakeholder reacts. Free.

Reframe the request

The ask is a solution. Find the problem behind it, a metric for it and the cheapest first move.

01

Lior, CMO

“Users don't trust our AI. Add a “Powered by GPT” badge and a friendly mascot.”

What you know: Users who abandon AI answers say: “I couldn't check where it came from.”

What's the real problem?
Strong: Users can't verify answers, so they can't rely on them.Trust follows checkability. That's designable.
Partly: The AI feels cold.Tone matters a little. Sources matter a lot.
Weak: Our AI brand isn't recognized.The users told you why they leave, and it isn't the brand.
How will you know it worked?
Strong: Answers used without re-asking, and how often sources get opened.Trust shows up as reliance on the right answers.
Partly: NPS of the AI feature.Fine as a pulse, too blunt as a target.
Weak: Badge awareness in a survey.Measures the badge, not trust.
What do you do first?
Strong: Add sources to the 20 most common question types and test.Small, fast, aimed at the stated reason.
Partly: Interview users who abandoned answers.Useful, but you already know the main reason.
Weak: Brief an illustrator on the mascot.Cute doesn't make claims checkable.

What seniors do: Trust isn't a badge. Show where every answer came from, and say clearly when the AI doesn't know.

02

Shira, Support lead

“The chatbot resolves 70% of tickets. Make it harder to reach a human so we hit 85%.”

What you know: Repeat contacts within 7 days are up 30% since the bot launched.

What's the real problem?
Strong: “Resolved” counts chats that ended, not problems solved.The repeat contacts say the 70% is inflated already.
Partly: Support costs are too high.True, and hiding humans raises them later through churn.
Weak: Users just prefer humans.An assumption. Many prefer a fast correct answer from anyone.
How will you know it worked?
Strong: Problems solved with no repeat contact within 7 days.Measures resolution, not avoidance.
Partly: Post-chat satisfaction.Asked too early, before the user knows if it worked.
Weak: Deflection rate.The metric that created the problem.
What do you do first?
Strong: Read 50 “resolved” chats that came back within a week.Shows exactly where the bot fails, and what to fix or hand off.
Partly: Add a satisfaction survey.More data, same blind spot.
Weak: Hide the human button for 10% of users.Tests how much pain users tolerate, not how to help them.

What seniors do: Make the hand-off better, not harder: the bot solves what it can, passes the rest with context, and is measured on problems that stay solved.

03

Tom, Founder

“Let's replace onboarding calls with an AI avatar of our customer success manager.”

What you know: Customers who get an onboarding call retain twice as well. Most of the call is spent on data import.

What's the real problem?
Strong: Data import is hard, and the calls succeed because a person gets customers through it.The call works because of what happens in it, not because of who's on screen.
Partly: Calls cost too much to scale.True, but it's the constraint, not the user problem.
Weak: Customers want a friendly face.Nothing in the data says the face is the value.
How will you know it worked?
Strong: Import completed in week one, and 90-day retention.Measures the thing the calls actually achieve.
Partly: Ratings of avatar sessions.Nice, but people rate novelty.
Weak: Number of calls replaced.Measures cost saved, even if retention drops.
What do you do first?
Strong: Map the top five import blockers from call recordings.Tells you what to fix in the product and what AI should guide.
Partly: Survey customers about avatars.People can't predict how they'd use something new.
Weak: Record the manager for the avatar.Building the solution before knowing the job.

What seniors do: Fix the import, then let AI guide it step by step inside the product. Keep humans for accounts that get stuck.

04

Yael, Design director

“AI can generate 50 landing page variants. Let's A/B test them all.”

What you know: The page gets 3,000 visits a week and converts at 2%.

What's the real problem?
Strong: We don't know why visitors don't convert, and traffic can't support 50 variants.With ~60 conversions a week, 50 variants would take months to read.
Partly: The page looks dated.Maybe, but looks rarely explain a conversion gap on their own.
Weak: We lack creative ideas.AI made ideas cheap. Insight is what's scarce.
How will you know it worked?
Strong: Conversion, in a test big enough to detect a real change.An honest result beats a lucky-looking winner.
Partly: Time on page.Ambiguous: longer can mean confused.
Weak: Number of variants tested.Activity, not learning.
What do you do first?
Strong: Watch 10 session recordings, talk to 5 lost leads, then test two strong hypotheses.Fewer, better bets that your traffic can actually judge.
Partly: Test five headlines.Better, still guessing.
Weak: Generate the 50 variants.You'd get noise, and a false winner.

What seniors do: AI makes options cheap. Traffic and insight are still scarce, so spend them on a few well-reasoned bets.

05

Dana, Head of Operations

“Let AI auto-approve every expense report under $5,000.”

What you know: Most fraud cases last year were between $500 and $2,000.

What's the real problem?
Weak: undefinedundefined
Weak: undefinedundefined
Weak: undefinedundefined
How will you know it worked?
Weak: undefinedundefined
Weak: undefinedundefined
Weak: undefinedundefined
What do you do first?
Weak: undefinedundefined
Weak: undefinedundefined
Weak: undefinedundefined

What seniors do: Automate by risk, not by amount: routine low-risk reports go through, unusual patterns go to a person, whatever the total.

06

Shira, VP Customer Success

“Build an AI model that predicts which customers will leave.”

What you know: Most accounts that left had stopped using the export feature about a month before.

What's the real problem?
Strong: There's a warning sign, and nobody acts on it.The gap is the response, not the prediction.
Partly: We can't tell which customers are likely to leave.Partly, but a strong signal already exists and nobody acts on it.
Weak: Customers don't get enough value from the product.Too broad to design anything for.
How will you know it worked?
Strong: More at-risk accounts contacted in a week, and kept.Measures the save, not the forecast.
Partly: The number of risk alerts sent to account managers.Activity, not outcome.
Weak: The model predicts churn with high accuracy.A perfect prediction nobody acts on saves no one.
What do you do first?
Strong: Alert when an account stops exporting, and try three messages.Tests the whole loop in days, with the signal you already have.
Partly: Interview ten customers who left in the last quarter.Useful, but slower than acting on the signal now.
Weak: Hire a data scientist to explore all of the usage data properly.An expensive answer to a question you can test this week.

What seniors do: When a simple signal exists, the problem is usually the response, not the prediction. Build the loop first, and add a model when a rule stops being enough.

07

Avi, CTO

“Users can't find anything. Let's replace search with an AI assistant.”

What you know: Search logs show 40% of searches use internal terms that don't appear in page titles.

What's the real problem?
Strong: People search with words our pages don't use.Specific, evidenced, and fixable in several ways.
Partly: Our search technology is old and needs replacing.Maybe, but the logs point at words, not technology.
Weak: People would rather chat than type in a search box.Nothing in the data says so.
How will you know it worked?
Strong: More first-try clicks, fewer “can't find it” tickets.Measures finding, not searching.
Partly: Satisfaction with search in the quarterly survey.Useful, but slow and noisy.
Weak: How often people use the new assistant each week.Use also goes up when people struggle.
What do you do first?
Strong: Add synonyms for the top 50 terms, compare in a week.A cheap test of the actual cause. If it isn't enough, AI is next.
Partly: Rewrite every page title in the words users use.Could help, but slower and riskier than synonyms.
Weak: Build an assistant prototype and test it with users.Months of work before trying the cheap fix.

What seniors do: Fix the mismatch the data shows before replacing the system. AI can still help, for example by suggesting synonyms, as part of search rather than instead of it.

08

Eitan, COO

“Have AI write the weekly status report for every team.”

What you know: Executives say they skip the reports because they're long and all look the same.

What's the real problem?
Strong: Reports don't show what leaders need to act on.Long and same-looking means the signal is buried.
Partly: Writing the reports takes teams too much time.Maybe, but the complaint is from readers, not writers.
Weak: Teams aren't doing enough to report on each week.Nothing in the evidence says so.
How will you know it worked?
Strong: Leaders act on more issues within the week.Measures what the reports are for.
Partly: Reports take less time to write each week.A real gain, but not the problem leaders raised.
Weak: Every team submits a report on time each week.Faster, longer reports nobody reads.
What do you do first?
Strong: Ask five leaders what they looked for last week.Shows what the report needs to contain, before anyone writes it.
Partly: Make a template with fewer, shorter sections.Helps, but you don't know yet which sections matter.
Weak: Have AI write next week's reports for two teams.More of the same text, faster.

What seniors do: If nobody reads it, writing it faster doesn't help. Find what readers need, shrink to that, and then let AI draft it.

09

Tom, Head of Trust & Safety

“Use AI to block all suspicious sign-ups automatically.”

What you know: Last quarter, 1 in 5 blocked sign-ups was a real customer who later emailed support to get in.

What's the real problem?
Strong: We block real people, and fraud still matters.Two errors to balance, not one to maximize.
Partly: Fake accounts are getting through our current filters.Maybe, but the evidence shows the opposite error too.
Weak: Our sign-up form doesn't ask for enough information.More fields hurt real customers most.
How will you know it worked?
Strong: Fewer fake accounts, and fewer real people blocked.Measures both kinds of mistake.
Partly: Fewer fraud reports from existing customers.Part of it, blind to lost customers.
Weak: The number of sign-ups the AI blocks each week.Goes up when it blocks real people too.
What do you do first?
Strong: Send unsure cases to a quick check, not a block.A middle path for the cases the model can't call.
Partly: Ask support to whitelist people who email in.Fixes it after the damage, one email at a time.
Weak: Turn on auto-blocking for the riskiest country first.Concentrates the false blocks on one group.

What seniors do: A filter makes two kinds of mistakes. Measure both, and give uncertain cases a gentler path than a hard block.

10

Maya, VP Mobile

“Add voice commands to our banking app. Everyone is talking to AI now.”

What you know: The top app-store complaint is “too many steps to send money.”

What's the real problem?
Strong: Sending money takes too many steps.What users actually said, and solvable many ways.
Partly: People want to use the app without their hands.Some do, but it's not what the complaints say.
Weak: Our app doesn't feel modern next to new AI apps.A feeling, not a problem users named.
How will you know it worked?
Strong: Time and taps to send money, and drop-off.Measures the complaint directly.
Partly: Fewer “too many steps” reviews over time.Relevant, but slow and indirect.
Weak: How many people try a voice command.Curiosity, not a better transfer.
What do you do first?
Strong: Map the transfer flow and count every step.Shows which steps can go, with or without voice.
Partly: Survey users about voice features they want.Asks about the solution, not the problem.
Weak: Build a voice prototype for sending money.Adds a new way in before fixing the long path.

What seniors do: A trendy interface doesn't shorten a long flow. Cut the steps first, then see if voice still adds something.

11

Ido, Head of Growth

“Use AI to personalize the whole homepage for every user.”

What you know: 70% of sessions start with a search for one of the same five items.

What's the real problem?
Strong: The homepage doesn't show what most people came for.They search for the same five things, so the homepage misses them.
Partly: The homepage design is old and needs a refresh.Maybe, but the behavior points at content, not looks.
Weak: Each user has very different needs from the homepage.The data shows the opposite: most want the same few things.
How will you know it worked?
Strong: People reach what they came for faster.Measures the goal of the page.
Partly: Fewer searches for the top five items.A good sign, but not the whole outcome.
Weak: Time spent on the homepage goes up.Could mean people are lost.
What do you do first?
Strong: Put the top five items on the homepage and compare.A one-day test of the simplest fix.
Partly: Interview users about what they want to see first.Useful, but the data already answers it.
Weak: Choose a personalization vendor and integrate it.Months of work before the cheap test.

What seniors do: When most people want the same thing, show it to everyone. Personalize later, for the minority it doesn't serve.

12

Karen, CRO

“Give every salesperson an AI copilot that listens to their calls.”

What you know: Most deals stall after the demo, waiting weeks for the buyer's IT security review.

What's the real problem?
Strong: Deals wait weeks on the buyer's security review.Where the time is actually lost.
Partly: Salespeople need better coaching on their calls.Maybe, but the deals stall after the calls.
Weak: Our demos don't convince buyers to move ahead.Buyers want to move. IT is the wait.
How will you know it worked?
Strong: Shorter time from demo to signed deal.Measures the stall directly.
Partly: Salespeople rate the copilot highly.Nice, but not revenue.
Weak: How many calls the copilot analyzes.Usage, not deals.
What do you do first?
Strong: Ask three buyers' IT teams what slowed them.Shows what to prepare in advance.
Partly: Interview the top sellers about how they close.Helpful, but the stall isn't in the selling.
Weak: Run a copilot pilot with five salespeople.Tests a fix for a different problem.

What seniors do: Look for where the time goes. Here a security pack ready before the demo could do more than any copilot.

13

Lea, Product lead

“Use AI to fill new workspaces with generated sample content.”

What you know: Most new users say they didn't know what to put in first. Workspaces still empty after day 1 rarely come back.

What's the real problem?
Strong: New users don't know what to do first.What users said, and why the workspace stays empty.
Partly: An empty workspace looks bare and unfinished.True, but the cause is uncertainty, not looks.
Weak: New users aren't motivated enough to start.They signed up. They're stuck, not unmotivated.
How will you know it worked?
Strong: More users add their own first item.Measures real use, not filler.
Partly: Fewer workspaces still empty after day 1.Sample content alone makes this look solved.
Weak: How much sample content AI generates.Output, not use.
What do you do first?
Strong: Watch five new users try to start a workspace.Shows what first step would help.
Partly: Add a welcome video to the empty workspace.Could help, but untested and easy to skip.
Weak: Generate sample content for every new workspace.Might hide the problem instead of solving it.

What seniors do: Filler content can make a metric look fixed. Help people take their own first step, then see whether AI can make that step easier.

Hold the line under pressure

A conversation in rounds. Each reply changes what they say next.

14

Maya, VP Product

“Research takes too long. Let's use synthetic AI users instead of interviews.”

Your first reply
Strong: Let's use them where they're good: stress-testing the interview guide and early ideas. Five real sessions for the decision itself.“Faster and still real. Fine.” You kept what matters and gave her speed.
Partly: Synthetic users are fake data.“So is half of what users say.” Fair point, no alternative.
Weak: OK, I'll set it up.Next quarter you ship something no real person asked for.
When they push back
Strong: Last quarter, a real user showed us the export workaround no model predicted. It changed the roadmap. I'll send you the clip.“I remember that one.” A real moment beats a methodology debate.
Partly: Research shows synthetic users are unreliable.“Research also takes too long.” Abstract evidence, easy to wave off.
Weak: Research is my area.Territory talk makes you look slow and defensive.
Closing the conversation
Strong: I'll cut the cycle to three days: AI drafts the guide and the synthesis, I run the sessions.“Three days, I can plan around that.” You fixed her actual problem: speed.
Partly: Let's revisit next quarter.Postponing keeps the tension, and the old pace.
Weak: We'll keep the current process.The problem she raised is still there.

What seniors do: Use AI to make research faster, not fake. Real people stay where the decision is.

15

Ron, Head of Growth

“Let's have AI personalize prices based on each user's willingness to pay.”

Your first reply
Strong: When users compare prices and find out, trust breaks overnight. What's the goal: revenue per user? There are cleaner ways to get it.“Yes, revenue per user. Go on.” You moved to his goal.
Partly: That's unethical.“Airlines do it.” A label, not an argument.
Weak: Cool, I'll design the flow.The screenshot of two prices goes viral within a month.
When they push back
Strong: Personalize the offer, not the price: the right plan at the right moment, one price list for everyone.“Same upside, no headline risk. OK.” Better alternative, same metric.
Partly: Let's keep the differences small.Smaller unfairness is still unfairness when it's found.
Weak: Let's ask legal first.Escalating before proposing makes you the blocker.
Closing the conversation
Strong: A/B test it on revenue per user, with refunds and complaints as the guardrail.“Ship the test.” A guardrail turns ethics into a number everyone watches.
Partly: Survey users afterwards.Too late to catch the damage.
Weak: Track conversion only.Conversion rises right up until the backlash.

What seniors do: Personalize value, not prices. If you couldn't explain it on the pricing page, don't ship it.

16

Dana, VP Revenue

“Add a full-screen upgrade prompt after every AI answer. Conversion is the only number the board asks about.”

Your first reply
Strong: Let's test it against a prompt shown when someone hits the free limit.“Fine, as long as I get a number by month end.” You turned an opinion into an experiment she controls.
Weak: That's a dark pattern. People will hate it, and I don't want my name on that design.“Then I'll find someone who will.” A moral stand with no alternative loses the room.
Weak: Sure. I'll mock up the full-screen version today so engineering can start this sprint.Conversion bumps for a week, then cancellations and one-star reviews arrive.
When they push back
Strong: Constant prompts teach people to dismiss without reading. Let's track dismiss rate.“Add it to the dashboard.” You gave her a metric that protects users.
Partly: Can we run this past the research team first, before we commit to a direction?Reasonable, but it sounds like a delay, not a plan.
Weak: Every study I've read says people hate popups, so this will hurt conversion, not help it.“People also hate paying.” A generalization, not evidence.
Closing the conversation
Strong: Let's agree now: if the limit version keeps more people at 30 days, it wins.“Agreed.” The decision is made before anyone gets attached to a variant.
Partly: I'll check in with you after launch, and we can go over the conversion numbers.A follow-up without a decision rule means the louder voice wins.
Weak: Let's ship both versions to everyone and see which one people respond to better.Two prompts at once, and no way to tell which one hurt.

What seniors do: Don't argue ethics in the abstract. Offer a better-timed alternative, test it on the stakeholder's metric, and add the metric that protects users.

17

Omar, Head of Sales

“Our biggest client wants the AI to auto-approve their refunds. Build it this sprint or we lose the renewal.”

Your first reply
Strong: Let's make it a setting: auto-approve under a limit they pick, ask above it.“They'll take that.” You kept the deal and a person in the loop for large amounts.
Weak: We don't build custom features for a single client. It sets a bad precedent.“Then you explain it to the CEO when they leave.” A policy with no alternative ends the talk.
Weak: OK. Let's auto-approve all of their refunds and see how it goes.Until a fraud spike drains their refund budget over a weekend.
When they push back
Strong: Every auto-approved refund gets logged, and their admin can reverse it for 48 hours.“Their finance team will love the log.” Accountability became a selling point.
Partly: We can ship it now and add logging in the next version, once it's proven.Speed now, blind spot later: the first dispute has no record.
Weak: Engineering thinks this is risky, so I'd rather wait for their review before we commit.Pointing at another team makes you a messenger, not an owner.
Closing the conversation
Strong: If it works, we offer the setting to every enterprise account next quarter.“So it's a product, not a favor.” You turned a one-off into roadmap.
Partly: Let's call it a one-time exception for this client and revisit it next year.Fine for now, but the next big client will ask for the same thing.
Weak: Let's wait and see whether any other customers ask for the same thing.No owner and no plan, so the next request is another fire drill.

What seniors do: Turn a custom demand into a configurable, reversible feature with a clear owner. Then decide whether it belongs in the product for everyone.

18

Lior, Engineering lead

“We're behind. Let's skip screen reader support for the AI chat and add it after launch.”

Your first reply
Strong: Keep the two things that lock people out: keyboard focus and announced replies.“Two things I can do in a day.” You shrank the scope instead of fighting the deadline.
Weak: Accessibility isn't optional. We can't launch something part of our users can't use.True, and it ends the conversation without a plan he can say yes to.
Weak: Fine, we'll do it after launch. Can you add it to the backlog so we don't forget?“Later” rarely comes, and blind users can't read the replies at all.
When they push back
Strong: Unannounced replies mean a screen reader user asks and never hears back.“I didn't realize it was that basic.” A concrete failure beats a principle.
Partly: There could be legal exposure here, and I'd rather not find out the hard way.Can be true, but fear without specifics sounds like stalling.
Weak: Most of our competitors support screen readers in their AI chat already.“They have more engineers.” A comparison isn't a reason.
Closing the conversation
Strong: I'll test it with a screen reader on Thursday and write the two tickets myself.“Then it's done.” You took on the work, not just the opinion.
Partly: Can QA check it with a screen reader before we launch and report what breaks?Delegating the check is fine, but nobody owns the fix.
Weak: Let's bring it to the retro and decide as a team how to handle it next time.After launch, with real users already locked out.

What seniors do: Separate what blocks people from what polishes the experience. Keep the blockers, defer the rest, and own the check yourself.

19

Noam, CEO

“AI can generate our screens now. Let's skip design reviews and ship what it produces.”

Your first reply
Strong: Let's ship AI screens on low-risk pages now, and keep review for checkout.“So we get the speed where it's safe.” You said yes to his goal, with a boundary.
Partly: AI screens look finished but miss edge cases. Users will notice, and so will support.“So fix the edge cases.” True, but it sounds like defending your job.
Weak: Design review is how we keep quality. Skipping it is how products get worse.“That's what you'd say.” A principle from the person it affects sounds like self-interest.
When they push back
Strong: Let's compare support tickets and drop-off on AI pages against reviewed ones.“Fair test.” Now the answer comes from data, not from who's in the room.
Partly: I can show you three AI screens with problems we caught in review last month.Good evidence, but three examples won't outweigh a cost argument.
Weak: Other companies tried this and had to roll it back after a few months.“Which ones?” A vague claim falls apart under one question.
Closing the conversation
Strong: I'll write a checklist so reviews of AI screens take 15 minutes, not a day.“That I'd approve.” You fixed his real complaint, which was speed.
Partly: We could hire a contractor to review the AI screens instead of the team.Moves the work, but adds cost he didn't ask for.
Weak: Let's keep the current process until we have more data on quality.The status quo is exactly what he wants to change.

What seniors do: When a leader wants speed, don't defend the process. Agree where the risk is low, measure the rest, and make the safeguard fast enough to keep.

20

Ronit, Head of Data

“Let's train our model on customer chat logs. The terms of service already allow it.”

Your first reply
Strong: Allowed isn't the same as expected. Most customers never read that part of the terms.“Fair, but legally we're covered.” You moved the question from law to trust.
Weak: That's a privacy violation, and I think we'd lose customers over it if it got out.“Legal says it isn't.” Calling it a violation is easy to dismiss.
Weak: Fine by me if legal signed off. It's their call, not a design question.It is a design question: nobody designed how customers learn about it.
When they push back
Strong: Let's first check how many logs contain health, money or family details.“That's a lot more than I thought.” A number made the risk concrete.
Partly: We could ask the privacy team to write a clear policy before we go any further.Useful, but a policy doesn't tell you what's in the data.
Partly: Can we use public data instead? It's safer and nobody can complain.Safer, but it may not teach the model what she needs.
Closing the conversation
Strong: Let's tell customers, let them opt out, and strip personal details.“That I can do in a sprint.” You turned trust into three concrete steps.
Partly: Let's just make sure the model never repeats anything word for word.Helps, but customers still never agreed to it.
Weak: Let's go ahead and handle any complaints if they come up later.By the time complaints come, it's a news story.

What seniors do: Legal permission is the floor, not the bar. Find out what's in the data, then make the use visible and give people a choice.

21

Gal, Founder

“Give the support bot a human name and photo, and don't say it's AI. Customers are nicer to people.”

Your first reply
Strong: Customers who find out feel tricked, and they'll post the screenshots.“Hm. That would be bad.” You named a cost he cares about.
Weak: Pretending to be human is lying to customers, and I won't design that.“It's just a name.” He hears a lecture, not a risk.
Weak: Sure, a friendly name and photo should make the chat feel warmer.Until a customer asks “are you a real person?” and the bot says yes.
When they push back
Strong: A friendly name is fine. Let's keep it and say it's an assistant in the first line.“So I keep the warmth.” You kept what he wanted and dropped the deception.
Partly: Several places now require bots to say they're bots, so this could be a problem.Possibly true, but he'll ask legal, not you.
Partly: Let's ask customers in a survey whether they'd mind talking to a bot.Slow, and people say one thing and do another.
Closing the conversation
Strong: Let's try the honest version for a month and measure rudeness reports.“If it holds up, fine.” You turned his worry into a measure.
Partly: We can add a small “AI” tag in the footer of the chat window, out of the way.Better than nothing, but almost nobody reads the footer.
Weak: Let's just try it and switch back if people complain about it.The complaints come after the trust is gone.

What seniors do: Keep the warmth, drop the deception. Disclose plainly where people will see it, then measure whether the fear was real.

22

Yoav, Head of Product

“The board wants AI in every part of the product by Q1. Let's add it to all five features at once.”

Your first reply
Strong: Five at once means five half-done features. Let's pick one and do it well.“Which one?” You turned a mandate into a choice he can make.
Partly: We don't have the people for five AI features in one quarter.True, but he hears “no” and still owes the board an answer.
Weak: OK, let's add a smart suggestion box to each of the five features this quarter.Five boxes nobody uses, and the board asks why usage is flat.
When they push back
Strong: Most time is lost in reporting. That's where AI should start.“That's a story the board will like.” Evidence picked the place.
Partly: Let's ask users which feature they'd most like to see AI in.Useful, but people ask for what they can imagine, not what helps most.
Weak: Whichever feature is easiest to build, so we can show something fast.Fast, and likely the least useful.
Closing the conversation
Strong: Let's show the board one feature that works, with usage, and a plan for two more.“That's a better slide than five demos.” A result plus a roadmap beats a checklist.
Partly: I'll write a memo explaining why five at once is too risky.Reasonable, but a memo against the board's ask rarely wins.
Weak: Let's demo all five to the board, even if some of them are only clickable prototypes.Prototypes on a board slide become deadlines.

What seniors do: When AI is a mandate, make it a choice: pick where it helps most, prove it there, and show the board a result instead of a checklist.

23

Hila, VP Engineering

“The agent emailed 3,000 customers the wrong renewal price last night. Whose call was it to skip the review step?”

Your first reply
Strong: It was mine. I'll explain why, what went wrong, and what we're doing now.“Thank you. Go on.” Owning it first lets the room move to fixing it.
Partly: It was a team decision during the sprint. We all agreed it was low risk.Technically true, but it reads as spreading the blame.
Weak: The prices came from the pricing team's sheet, so the data was wrong at the source.Maybe true, and now the room is about blame, not customers.
When they push back
Strong: First, a correction email to all 3,000 today, and we honor the lower price.“Do it.” Customers first, before the post-mortem.
Partly: First, let's figure out exactly how it happened before we tell customers anything.Understanding matters, but customers are acting on wrong prices now.
Partly: First, we pause the agent completely until we've reviewed everything it does.Safe, but it stops the useful parts too, and fixes nothing for customers.
Closing the conversation
Strong: Anything with prices or money now needs approval, and I'll check the other flows.“Good. Send me the list.” A specific fix with an owner.
Partly: Let's add more tests to the agent so this kind of error gets caught earlier.Helps, but tests miss the next new kind of error.
Weak: We'll be more careful with the agent from now on, and review things more often.“More careful” isn't a change anyone can check.

What seniors do: When something you decided breaks, say so first. Fix it for the people affected, then change the rule that let it happen.

24

Eyal, Engineering manager

“Showing sources under each AI answer is slowing us down. Let's drop it to make the launch date.”

Your first reply
Strong: Which part is slow? If it's previews, let's ship plain links for now.“Plain links take a day.” You found the cheap version of the feature.
Partly: Without sources, people can't check the answers, and they'll trust it less.“Will they really notice?” A fair point, but abstract.
Weak: OK. We can launch without sources and add them in the next version.Launch reviews say the AI “makes things up”, and there's no way to check.
When they push back
Strong: In our tests, people clicked a source on 1 in 4 answers before trusting them.“That's more than I expected.” Behavior beats opinion.
Partly: Legal will probably want sources anyway, to cover us when answers are wrong.Possibly, but now it's a legal question you can't answer.
Weak: Our competitors all show sources, so we'll look worse without them.“They launched a year ago.” Comparison isn't a reason.
Closing the conversation
Strong: I'll write the plain-links ticket today and test it with five people Friday.“Then we keep the date.” You owned the smaller version.
Partly: Let's ask the PM to decide whether sources are in scope for this launch or not.Escalating is fine, but you had the answer.
Weak: Let's push the launch a week so we can do the full feature properly.“Not happening.” A bigger ask than he came in with.

What seniors do: When a safeguard is said to cost too much, find out which part costs. There's often a cheap version that keeps the protection.

25

Dafna, VP Operations

“With AI, designers should ship 3× more screens. Let's track screens per designer per week.”

Your first reply
Strong: Screens are easy to game. Let's track problems solved and time to a decision.“Fair, if I can still see speed.” You kept her goal and changed the measure.
Partly: Fine, as long as we also count how many of those screens get thrown away later.A partial fix: the number still rewards volume.
Weak: Design isn't a factory, and creative work can't be counted in screens per week.“Everything can be measured.” A slogan against a number loses.
When they push back
Strong: Our biggest win last quarter was a screen we deleted. It would score zero.“Good example.” A real case shows what the metric would miss.
Partly: Designers will feel watched, and the best ones may start looking elsewhere.Possibly true, but it sounds like protecting the team from scrutiny.
Weak: As far as I know, no good design team measures its output like this anymore.“So we'd be first.” A vague claim, easily turned around.
Closing the conversation
Strong: Let's pilot both measures for a month and keep the one that predicts results.“Deal.” Now the data picks the measure.
Partly: I'll prepare a deck on how design value should be measured, for next quarter.Thoughtful, but by next quarter the metric is already in place.
Weak: Let's push back as a team and refuse to report any productivity numbers.“Then I'll pick the numbers for you.” Refusing hands over control.

What seniors do: When a leader picks a metric that rewards the wrong thing, don't refuse to be measured. Offer a better measure and let a short pilot decide.

26

Assaf, Head of Growth

“Let's tune the AI feed to maximize time spent for our teen users. Engagement is our main KPI.”

Your first reply
Strong: Time spent includes 2 AM scrolling. Let's check when teens actually use it.“I didn't know it was that late.” A fact about his own users moved him.
Partly: Can we wait until the policy team says what's allowed for under-18 users?Sensible, but it hands the decision to someone else, and slows it.
Weak: Optimizing teens for screen time is unethical. I don't want to work on that.“It's a feed, not a casino.” Moral language without a fact gets dismissed.
When they push back
Strong: Let's optimize for teens coming back each week, not for longer sessions.“That still grows the business.” A healthier goal he can still win on.
Partly: Let's add a screen-time reminder after an hour, to show we care about it.A gesture, while the feed still optimizes the same thing.
Weak: Engagement isn't a real KPI anyway. Let's measure something else instead.“It's my KPI.” Dismissing his goal ends the talk.
Closing the conversation
Strong: Let's test it on 10% of teens and track return visits and late-night use.“Small and measured, fine.” The test answers the question for both of you.
Partly: I'll write a short ethics review of the feed and share it with leadership.Useful later, but it doesn't shape this decision.
Weak: Let's roll it out to all teens now and look at complaints in a quarter.A quarter of late nights before anyone looks.

What seniors do: When a goal can harm a vulnerable group, find the fact that shows it, offer a goal that still grows the business, and test small.

Practice the conversation

Knowing the reply is different from finding it in the room.

Pushback drills put you in the conversation, with a stakeholder who reacts to what you say.

Try a pushback drill Free. About 3 minutes.
↑