Explain what an agent is in a way that distinguishes it from a chatbot with tools.

Thrumcastle Apparel sells clothes through an app. Weftline is the feature inside it that answers a shopper's question with real matches from the catalog. Idris Bramante owns it. When the marketing team renamed it "your AI stylist" and promised it would build a whole outfit and check it out for you, nobody changed what Weftline actually does under the hood. Three weeks later, that gap had a dollar figure attached to it.

How to answer this, stage by stage

Nobody is grading whether you can define "agent" like a glossary entry. They are grading whether you can hold the line between two words that sound almost the same, under a real, dated number.

Let's learn

Weftline is a feature inside Thrumcastle Apparel's shopping app. You type what you need, a wedding outfit, a rain coat, and it searches the catalog and shows you real matches you can buy.

One box waits for you. The other keeps feeding its own last answer into its next move. That's the whole difference, drawn.

For its first year, Weftline's own copy matched what it did: "AI powered search. Finds pieces that match what you asked for." Modest, accurate, and nobody complained, because nobody expected more than a search. Reviews mentioning confusion sat around two a week, background noise for an app this size.

Then a rival app started calling its own feature an "AI stylist," and it tested well. In a launch planning meeting, Thrumcastle's marketing lead, Piers Callendar, pushed to rename the campaign to match: "Your AI Stylist. Tell us the event. We build the whole look. We check it out for you." Idris signed off. Nobody changed what Weftline's system actually did. It still called one search tool per message, showed a handful of matches, and waited.

Same code underneath both panels. Only the promise changed, and the promise is the part a shopper actually reads.

The extra confusion wasn't really about Weftline getting worse at search. It was about what Thrumcastle let a shopper believe would happen next. A shopper who reads "we build the whole look and check it out for you" reasonably expects Weftline to also pick shoes and a bag, put them in a cart, and charge her card, no further typing needed. Weftline had never done any of that, not once, for anyone.

What it cost at its worst: sessions that came from the campaign ended after one search screen sixty one percent of the time, against a normal rate of thirty three percent for shoppers who never saw the "stylist" promise. Refunds and goodwill credits tied to "not what the AI promised" hit eighty six thousand dollars across three weeks. Piers pulled the remaining ad spend, forty thousand dollars already committed, once the pattern was undeniable.

What I would leave alone: the plain "similar items" strip at checkout, which shows one related item and has never claimed to do anything more. Nobody's ever filed a ticket about it, because its words match its behavior exactly.

The lesson: the word "agent," like "stylist," is a promise about how many decisions a system will make without you. It isn't a mood word for "has AI in it." Say the real mechanism, not the aspiration, and the mechanism will earn the mood word back on its own, later, once it's true.

Now here is the same thing as a story

The short version above is what you actually say out loud. Read this one for the meeting that started it, and the one review that ended it.

Idris Bramante had run Weftline for two years by the time any of this happened. He liked the feature precisely because it didn't overpromise: type what you need, get real matches from a real catalog, no magic. Return rates on items bought through Weftline sat lower than the app average, because the copy set the right expectation and shoppers who found a genuine match kept it.

The trouble started with a competitor. A rival shopping app rebranded its own search feature as an "AI stylist" that spring, and its download numbers jumped. Piers Callendar, who ran marketing for Thrumcastle, brought a slide to the quarterly planning meeting: "We keep calling this thing search. Everyone else calls theirs a stylist. Let's stop underselling what we built." Nobody in the room thought they were promising anything untrue. They thought they were finally describing Weftline the way it deserved.

Nobody decided, on any single day, to promise something Weftline couldn't do. A better sentence in a slide deck just never got checked against the actual code.

"Your AI Stylist" launched on a Monday. The new copy read: "Tell us the event. We build the whole look. We check it out for you." Weftline's code changed not at all. It still called one search tool per message, returned a handful of matching items, and stopped.

For the first week, the numbers looked fine, because nobody was watching the right one. Sessions that came in through the campaign were ending fast, but a fast session had always read as a good session on Thrumcastle's dashboard. One star reviews mentioning confusion crept from two a week to nine. Nobody had a number in mind that would have made nine feel alarming.

Then, on a Tuesday in the third week, Odalys Marchetti opened Weftline looking for something to wear to her sister's outdoor wedding in June. She typed exactly what the ad told her to type. Weftline came back with four dresses that matched. She waited for it to also suggest shoes, a bag, and a way to buy the whole thing at once, the way "we build the whole look and check it out for you" had told her it would. Nothing else happened. It was just waiting for her to type again, same as it always had.

She left a review that afternoon: "This isn't a stylist. It's a search bar in a chat bubble." A shopping forum picked it up by evening. By Wednesday morning, forty one reviews that week used some version of the same complaint, and Piers was the one who found Odalys's post, forwarded to him by a friend outside the company who had no idea he'd had anything to do with the campaign.

The decision Idris would take back sits in that quarterly planning meeting. Renaming Weftline felt like giving it the credit it deserved. Nobody asked the one question that actually mattered: does the system take more than one action on its own, or does it wait for a person every single time? Weftline waited every time. It had always waited every time. The new name didn't change that. It just changed what a shopper was allowed to expect.

One alternative got seriously discussed before the copy went back to something honest, and it's worth naming, because on paper it looked like the braver fix: build Weftline into an actual agent, one that really does assemble a full outfit and check it out, so the ad would finally be true. It lost, for now. A system that decides on its own to charge a shopper's card needs real guardrails first, a required check before it ever spends money, a hard cap on what it can spend without asking. Thrumcastle hadn't built either one. Shipping a fast version to match a marketing line was a worse bet than telling the truth immediately and building the real thing properly, later, with the safety checks in place before the first autonomous charge, not after.

Run the three weeks again, with the copy left honest from day one: "Smart search, picked for your event." Sessions end after one search screen at the normal thirty three percent, because nobody was promised a second act. One star reviews mentioning confusion hold near two a week. Odalys still gets four good dress options. She still has to pick her own shoes. She never has a reason to write the line that found its way to Piers's inbox.

What Idris would tell himself, back in that planning meeting: a name isn't marketing polish sitting on top of a product. It's a claim about how much work the product will do without you. We changed the claim and left the product exactly where it was, and the gap between the two was never going to stay invisible for long.

PICK, argued from both sides of one word

Not a way to decide whether "agent" is a good word to use. PICK is what forces you to test the word against the actual mechanism, twice, once for a team that oversold a chatbot, and once for a team that undersupervised a real agent.

One mistake is small enough to redraw by hand. The other one needed a chart to hold its own number.

One question, two branches, no third option. Weftline sits on the left branch no matter what its ad said.

Three things worth naming directly, since the real judgment sits here. The alternative worth naming and rejecting isn't only "keep the flashy name," it's also "build the real agent fast to make the ad true," and that one loses for a specific reason: a system that autonomously charges a shopper's card needs a confirm before charge checkpoint and a hard spend cap before it ever gets to act unsupervised, and Thrumcastle hadn't built either. The AI specific failure mode worth naming by name is an under scoped autonomous agent taking a real world, hard to reverse action, spending money, with nobody approving that specific step. The guardrail is exactly those two things: a required human confirmation before any autonomous charge, and a cap on what it can spend in a session without asking. And the trade off is real and accepted on purpose: honest, modest copy gives up some of the urgency that drives clicks, in exchange for never promising a capability the system doesn't have yet.

And if you want to be sure it really works, try it somewhere else

Same four letters, invoices instead of outfits, and this time the honest label was never the problem.

Tallymere Bookkeeping runs small business bookkeeping software. Its tool Duesend chases overdue invoices: it sends a reminder, and if nobody replies within five days, it sends a follow up, applies a late fee, and escalates to a collections referral that cc's the client's own finance team, each step decided from what happened at the last one, with no person approving any single step in between. Casimir Oyelaran, who owns Duesend, built it to be exactly what its name claims. It genuinely is an agent.

Every arrow in this chain fires without anyone's approval. That's what makes Duesend a real agent, and also what made it dangerous once, without a single word changing.

In its first quarter, Duesend ran this full chain, unattended, across 1,140 overdue accounts. Ninety six percent of the time it was exactly right, the client genuinely still owed the money at every step. But in forty six accounts, four percent, a fee or an escalation fired after the client had already paid, because the payment posted to the bank feed a day or two after Duesend had already queued its next step, and nothing in the chain ever paused to check again before acting. One of the forty six was a thirty eight thousand dollar enterprise account that nearly canceled after its own finance team got cc'd on a collections notice for an invoice it had settled four days earlier.

Run the same test on both products and it splits them cleanly. Weftline fails the test and shouldn't be called an agent. Duesend passes it and should be, but it still needed a guard Weftline never did.

Mapped onto PICK: the position doesn't change, Duesend really does chain more than one action from its own last result, with no person between steps, so calling it an agent was accurate the whole time. The impact is the opposite shape from Weftline's: the mistake here isn't a false promise, it's a true promise with no guard on the one step that could embarrass a paying customer. The cost asymmetry lands just as sharp: adding a pause and recheck before the fee and escalation steps costs a few engineering hours and a slightly slower average collection cycle, cheap and visible. Skipping that pause cost a thirty eight thousand dollar account nearly walking, hidden until the exact week it happened. And the kill criteria still answers it: yes, it takes more than one action in a row on its own, so the fix was never "stop calling it an agent." It was "put a human check before the two steps that are hardest to take back."

Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to it: run the one test, more than one action in a row, deciding the next from the last, no approval in between, that's an agent, anything else is a chatbot with tools, no matter what the landing page says.
Cost: no time to rebuild the guardrail this sprint. Fine, but pause the riskiest step manually until the check exists. Don't let a real agent's most expensive action run unsupervised just because the fix isn't built yet.
The model got better, for real: say Weftline's underlying search model genuinely got sharper at matching. Still doesn't make it an agent. A better single answer to one question is not the same claim as deciding what to do next on its own.

Where people run it wrong.
They let a marketing team pick the word before an engineer checks it against what the system actually does.
They assume "it uses AI" and "it's an agent" are the same claim, when the second one is a specific, checkable claim about who approves each step.
They build a real agent, get the definition right, and stop there, without asking which of its own steps is too expensive to run unsupervised.

How to use it live. Before you answer, ask yourself one thing: does this system take a second action on its own, deciding it from the result of the first, with nobody approving each step? If you can't point to that second, unapproved action, you haven't found an agent yet, you've found a chatbot with a tool and a good name.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0