Category: ai

LinkedIn’s AI could have you freezing your butt off in Alaska

Rakesh Agrawal

I love rampies. Not only could I not get off the ground without them, I couldn’t leave the gate. They load the luggage, drive the tugs that push back the plane and use the orange wands to marshal airplanes. They are out on the ramp whether it’s 100 degrees in Phoenix or -40 degrees in Nome.

But does that mean I could be one? Nope.

Though I do know not to load bags into an engine. That won’t end well. Note to Alaska: don’t hire Gemini to load bags.

On to LinkedIn…

Imagine my surprise when LinkedIn’s AI told me I was a great fit… for a ramp service agent role. LinkedIn has a feature for premium members that will tell you how good a match you are for a role.

Here’s my rating for a Ramp Service Agent role.

That is not a job I remotely qualify for. Location doesn’t match. Comp doesn’t match. Skills don’t match. But that rating is one of benefits I get for $40 a month.

This role at Alaska Airlines is a great fit. I could fly this at 40,000 feet. Fortunately, I’m also a “high” fit for this.

The only way these answers help is if candidates are doing spray-and-pray applications. That’s a waste of time for applicants.

Now imagine a recruiter looking at this same view. (I don’t have access to LinkedIn Recruiter, so I don’t know its sort order.) It wouldn’t help them either. If anything, it would destroy credibility for LinkedIn as a recruiting tool. And when a company owned by Microsoft (one of the biggest backers of AI) ships something this sloppy, it casts a shadow over all their other products.

But what about that “BETA” label, Rakesh? For those not familiar with tech talk, it means they’re testing the product for release.

I’ve designed and launched search products for much of my career. I’d never put this out beyond a closed internal beta, much less as a premium feature. As it stands, this isn’t a product — it’s unpaid labor for LinkedIn’s AI. The thumbs up/thumbs down will train their model. Even better: just hire a large RLHF team in India.

Amy Miller, a recruiter at Amazon, hates AI for “scoring.” This is a good reason why.

Rampies don’t need AI scores — they get planes moving. LinkedIn’s “match” feature should aspire to that kind of utility: useful, reliable, and grounded in reality. Until then, it feels more like unpaid labor for their AI than a benefit for members.

And if you’re looking for a product executive who knows AI — and knows when not to trust it — send me a message.

Parting shot: Here’s what WordPress generated. The NTSB will want a conversation.

Written by me, lightly edited by ChatGPT, illustrated by ChatGPT & Gemini. Unlike LinkedIn’s AI, none of them tried to send me to Nome at 40 below.

Alaska provides another glimpse of the future

Rakesh Agrawal

Alaska Airlines is offering another time-saving feature for passengers in Portland and Seattle: bag check with facial recognition to avoid having to wait in line for someone to verify your ID.

Although this feature isn’t an industry first, it’s an important improvement. Alaska has had some notable firsts.

From ChatGPT:

First Airline to Sell a Ticket Online (1995)

This is their most famous claim to fame. In 1995, Alaska Airlines became the first airline in the world to allow customers to book and purchase a flight over the internet. This innovation fundamentally changed how travel is sold and paved the way for the modern e-commerce and online travel agency landscape we know today.

First Airline with Online Check-In (1999)

Building on their web-based success, Alaska was the first airline in the world to introduce online check-in. This allowed passengers to check in for their flight and print their own boarding passes from their home or office, a revolutionary convenience that helped passengers bypass long lines at the airport.

Air travel is one of the rare cases where interests are aligned between customers and businesses.

(The information below is based on having flown 2 million mile across dozens of carriers. As with all things travel, there are exceptions and airlines that handle things differently.)

We’ve already seen a lot of innovation in the time I’ve been flying.

Old wayNew way
Carry paper ticketE-ticketing
Wait in line to check in Mobile boarding passes, online check in, kiosks
Wait in line to drop my bag (in the rare cases I do) Face recognition to check bags (this is what Alaska just announced)
Wait in longer lines at security for people who rarely travel fumbling for things TSA Pre✓ and Clear Plus
Wait for a refundE-tickets mean you don’t have to return paper tickets
Wait in line for an hour for flight changes Many airlines automatically re-book; if those options don’t work, you can use a mobile app to find alternatives
Stay over a Saturday night to get the best faresThe rise of low-cost carriers all but eliminated Saturday night stay requirements

These changes weren’t just cost cutting — they removed friction for travelers and streamlined operations for airlines.

I see a lot of innovation ahead, some which are starting to take shape.

AI

This is an interesting one because the definition has been so vague and misused. Airlines have been using machine learning and predictive modeling to set price and fleet management for decades.

When I talk about AI, I mean new tools like LLMs and user agents.

The most obvious use for AI for airlines is better answering user customer service inquiries. Unfortunately, most of the tools out there now are dumb bots. They don’t incorporate things like my status, itinerary and previous interactions. The best LLM-linked bots will also incorporate the reference tools airline agents access.

Not much further down the AI path is user agents. Instead of booking through the airline Web site, agents will do the searches on my behalf. (What you see now is dumb AI, which is just a fancy screen scraper. They may be marketed as agents, but they’re not.) Airlines need to be paying attention to agents that use MCP and testing ways to integrate with their systems.

The upside for airlines: ability to cut out OTAs and integrate their direct-connect capabilities. It eliminates fees that GDS’s charge and makes it easier to sell ancillaries like premium seats and lounge access.

One key to AI is to not lose trust. If an AI gives a response, it should be treated just like an agent gave that response.

Air Canada refused to honor a response from a chat bot, saying the bot was inaccurate. This led to a social media firestorm and a lawsuit that Air Canada lost. It’s much better in the long run to maintain trust.

This is just on the customer side. There is a lot of room for using LLMs in maintenance, pilot resources, irrops strategy and ad targeting.

Mobile

Airlines have been making big leaps using mobile for all areas of the travel experience. On the consumer side, you can manage new bookings, change flights, flight status and more.

  • Hotel vouchers sent straight to our phones for irrops: Don’t make me stand in a long line for a paper slip after a cancellation. A mobile voucher also tells me the quality of the hotel being offered — no more roulette with substandard properties.
  • Meal vouchers: Sent to our phones. United already sends you virtual debit cards, though they should be Apple Pay or Google Pay provisioned. (Many merchants don’t know how to type in credit card numbers.)
  • Delayed baggage: When the airline knows a bag has been left behind at the previous station, send a pre-filled delayed baggage form. This saves the passenger and the baggage service staff time. Vouchers can also be sent. It’s a great time to be looking at this while airlines integrate IATA 753, a mandate on how baggage must be tracked.

CTI (Computer Telephony Integration)

We’re already seen integration between your actions online and the experience you get when you call. With better CTI, that process will become more seamless. This is one of the best cases of aligning interests. Less time on the phone is good for me and the airline.

If you have a phone that is registered to your account:

  • The agent should have my next reservation already on screen.
  • If I was recently on the Web site, skip all the prompts and go to an agent who is prepped with what you were trying to do. (I don’t need to hear that you have a Web site, because I was just there.)
  • Take user state into account. if I just got off a flight and had checked bags, my call may should be handled differently

The caveat here is that caller ID and ANI can be faked. Airlines need to do a basic level of authentication before jumping into the conversation. Increasingly voice authentication can be defeated using AI. Pull up the record, but ask a question that only I know the answer to.

Less time in line or on hold is good for passengers and airlines. The same technology that cuts friction also cuts costs — if airlines use it to build trust instead of deflect service.

Bottom line: Less time in line or on hold is good for passengers and airlines. The same technology that cuts friction also cuts costs — if airlines use it to build trust instead of deflect service.

Agentic Commerce or Agentic Con Job? Whose Side Will Your AI Really Be On?

Rakesh Agrawal
Image of someone searching for flights

There’s a lot of talk in the AI world about “agentic commerce.” My big question: whose agent?

We have “agents” today, but most of them aren’t working for you. They aren’t fiduciaries.

Travel agents? They steer you to whoever gives them the largest commission. Despite all their targeting data, the default sort order is Recommended. I usually want Price or Rating, but Expedia wants to steer my booking—so it shows Recommended. (Expedia takes a 15–30 percent commission on hotels.)

Real-estate buyer’s agents? Same story. When I bought my place, my agent urged me to pay full asking price. I bid lower. He wanted the higher price because his commission would be higher.

Almost every service you buy, no matter how it’s marketed, isn’t designed in your best interest. “Free” stock trades? Robinhood makes its money on the back end through payment for order flow—$677 million of it in 2023. The finance industry even lobbied successfully against a rule that would have required brokers to act as fiduciaries.

Google is the ultimate example, even though Google’s founders were skeptical of the advertising model.

“We expect that advertising-funded search engines will be inherently biased toward the advertisers and away from the needs of the consumers.” — Sergey Brin and Larry Page, 1998 (The Anatomy of a Large-Scale Hypertextual Web Search Engine)

In the U.S., the only mainstream professions actually required to put your interests first are lawyers and fee-only financial advisers. (Yes, lawyers still have a conflict: the more hours they bill, the more you pay.)

What AI could change

With a fundamental shift like AI, there’s a possibility of turning that model on its head.

Take Amazon Marketplace: merchants compete for visibility, and Amazon takes 30–50 percent for the privilege. If my agent could talk directly to their agents, we could split that spread—I’d pay less, the merchant would earn more.

My agent could negotiate on my behalf:

“Hey airlines, I need to fly SFO → JFK next Tuesday. Give me your best bid.”

Instead of spending 20 minutes sifting through Expedia, my agent could strike a deal in milliseconds. I could even place limit orders: “When someone offers this trip for $300, buy it.”

Airlines and intermediaries would hate that, but it’s closer to an efficient market. Intermediaries wouldn’t skim 30 percent just for hosting a platform. (Yes, platforms claim they provide protection and customer service, but in reality most offer little of either.)

Today’s model is expensive because there are many mouths to feed along the way.

Will it happen?

I’d happily pay my agent a subscription fee if it truly worked for me. Will it happen? I’d love it—but history says no.

We’re addicted to “free,” even though “free” often costs us more.

(Written by me, lightly edited by ChatGPT.)

Disclaimer: As with all things AI, the industry moves at a rapid pace. Models evolve, tools update, and behaviors shift—sometimes overnight. By the time an author hits ‘publish,’ the example they’re using may already be obsolete. It’s not that the writer was wrong. It’s that the system changed while their post was still rendering. Disclaimer 2: The previous disclaimer (only) was written by AI. Disclaimer 3: Any future attempts to update Disclaimer 1 may invalidate Disclaimer 2.

2 + 2 = 5: Why AI can’t do math

Rakesh Agrawal

Since the dawn of calculators, we’ve trusted computers to do math. Unless there’s a bug in the code, computers have been great at math. Better and faster than the fastest humans. We use them in everything from stock trading to calculating change at the register. Some grocery stores even have automated coin dispensers; the register calculates the coins due and tells another machine to spit out the right coins.

What are marketed as the most powerful supercomputers ever… can’t do math. They can write sonnets, pass the bar exam, summarize Tolstoy… and botch 17 + 5. When I gave AIs a spreadsheet with places I’ve traveled and asked it to do a simple count, they kept coming back with the wrong numbers. (The spreadsheet I imported did an easy and accurate count: the number of rows matched the number of places I’ve visited.) ChatGPT can get something as basic as less than or greater than wrong. When it comes to math, AI might not be smarter than a 5th grader.

It has to do with the way AI works. Unlike spreadsheets and other tools we’re used to, it’s just predicting what will come next. It’s not actually doing math. It predicts based on inferences. The name is a tell for people in the industry: LLM stands for Large Language Model. It’s not a Large Math Machine.

But people who’ve been told that AI is a magic machine don’t know that. Like everything else, the AI confidently spits out an answer. The better implementations will do the inferences, recognize that might be wrong (“hey, I just did math!”), go back and rewrite Python code, execute it and spit back an accurate, calculated answer.

Yes, there is a small print disclosure that says “AI can make mistakes.” It should add “My math might not be great.”

If you do use AI to do math, ask it to double check the answer as a math problem. Or, you can just use a spreadsheet until the technology gets better.

Disclaimer: As with all things AI, the industry moves at a rapid pace. Models evolve, tools update, and behaviors shift—sometimes overnight. By the time an author hits ‘publish,’ the example they’re using may already be obsolete. It’s not that the writer was wrong. It’s that the system changed while their post was still rendering. Disclaimer 2: The previous disclaimer (only) was written by AI. Disclaimer 3: Any future attempts to update Disclaimer 2 may invalidate Disclaimer 1.

“When will AI be able to replace a product thinker like me?”

Rakesh Agrawal

Whenever I meet people, they tell me that AI will get rid of most jobs—except theirs.

I know that there is a good chance that at some point AI will develop skills that will help it get closer to my skill level.

I’ve written about how AI can replace PMs with up to 7 years of experience. So I asked ChatGPT “When will AI be able to replace a product thinker like me?”

The TL;DR: Keep building the future. AI can (currently) only do what is in its training set. If you do things that haven’t been done before, you can stay ahead of the tsunami. If you can manage complex organizations to ship. If you’re a system-level thinker.

Is this good advice? That’s TBD. But it’s also the best answer that I’ve been able to come up with when talking with many long-term tech execs.

The following is a transcript of ChatGPT’s answer. When drafting this, I told ChatGPT that I’d put my own assessment in italics. It misunderstood and put its assessment of what my assessment would be. They were spot on, so I left them in.

Executive Summary

AI can already outperform junior and mid-level PMs at execution. It writes PRDs, prioritizes roadmaps, handles tradeoff matrices, and synthesizes user feedback—all faster, cheaper, and more thoroughly than most human PMs. But it still struggles with judgment, taste, and navigating political, temporal, and organizational complexity. Those are the layers where top product thinkers still thrive.

I’m watching this happen in real time. What used to take days of back-and-forth now takes a well-structured prompt and 30 seconds. But the AI still doesn’t know when to say no—or when to break the rules.

This memo breaks the question into three timeframes: what AI can do today, what it will likely master soon, and what may remain uniquely human.

1. What AI Can Already Do (Today)

AI now matches or outpaces $150K+/year PMs at:

  • Feature scoping & prioritization
  • Drafting PRDs and user stories
  • Competitive teardown and synthesis
  • UX tradeoff analysis
  • Roadmap justification and market segmentation
  • Edge-case anticipation
  • Implementation feasibility checks (tooling, latency, systems integration)

Key insight: For structured thinking and repeatable processes, AI is already a high-functioning, tireless, no-ego PM.

In side-by-side comparisons, I’ve seen AI outperform real PMs in breadth, speed, and edge-case awareness. It’s humbling—and a little unnerving.

2. What AI Will Likely Master Soon (1–3 Years)

  • Modeling political/organizational dynamics: understanding who needs buy-in and where blockers live
  • Multi-system reasoning: how product changes ripple through infrastructure, marketing, and user trust
  • Value-based prioritization: balancing revenue, delight, risk, and cost with contextual nuance
  • Taste mimicry: copying the instincts of great product thinkers based on exposure to their prior work

Key risk: If you stay in the realm of optimization and extrapolation, you’ll get outpaced by a machine that never sleeps and never second-guesses.

It’s already good at seeing the chessboard. It’s learning how the players behave. Soon it will start asking which games are worth playing.

3. What’s Hard to Replace (Maybe Never)

This is the true job of senior product leadership:

a. Knowing when to ignore the data

“The metrics say it’s fine, but I know it’s not.”

b. Calling organizational bullshit

“This isn’t a feature request. It’s a political workaround.”

c. Pattern recognition across domains

“This reminds me of something that failed five years ago in a different vertical.”

d. Taste

You can’t A/B test your way to Pixar. You don’t user research your way to vision. You feel when it’s right.

Key moat: Judgment, taste, and courage remain hard to automate—especially when success requires going against the grain.

These are the moments where senior PMs still earn their title. When to kill a darling. When to greenlight a moonshot. When to say “trust me.” AI doesn’t have that instinct—yet.

Final Assessment

If your job is to execute, AI is already competitive.

If your job is to bet, guide, interpret, and influence—it’s still your game to lose.

The PMs who will survive and thrive are those who:

  • Integrate systems thinking and storytelling
  • Understand power and incentives
  • Apply historical and cross-domain reasoning
  • Develop and defend taste

AI isn’t coming for your job. It’s coming for the parts of your job that weren’t your differentiators anyway.