Have You Ever Argued With Your AI?
Our outreach runs through Claude, and it has transformed the work. Then one evening it refused to delete four contacts — and the argument that followed maps neatly onto Yann LeCun's "no world model" critique of LLMs.
I lost an evening in Windsor this week arguing with a piece of software. I won. I'm not sure that's much to be proud of.
Some background. Inspired by Jason Lemkin's public journey using AI for sales, we've spent the last few months running our outreach at Eebz through Claude's Cowork mode. It's been genuinely useful - a job that used to eat a person's week now gets done overnight. Much of that is HubSpot housekeeping: segmenting, enriching, updating titles, stamping campaigns, and routinely removing contacts that don't belong in a list.
That last one came with a small ritual. Every so often Claude would balk at deleting a contact. I'd remind it that HubSpot destroys nothing - it moves the record to a recycle bin, gives us ninety days to review, and a human checks the queue. It would accept that and carry on. Thirty seconds of friction, then back to work.
To be fair before I'm unfair: the workflow has been a real success. Our outreach programme is faster, and the metrics have followed - open rates around 60%, which I'd not have believed a year ago. I'd rather let the behavioural shift speak for itself than build a theory on one figure, but the direction is unambiguous. Setting up a campaign is no longer the bottleneck.
Then, last week, I started a fresh campaign for Black Friday. (Aside for anyone in retail: peak season is exactly when a digital shelf analytics platform earns its keep, and it's what Eebz is built for. End of advert.) New project, clean slate. Claude worked beautifully at first - counted a segment, reconciled the number two ways, checked for missing LinkedIn URLs and noted, unprompted, that "populated" isn't "valid," then found eleven email opt-outs and offered me a clean list. This is the good version of the tool: careful, fast, honestly hedged.
And then I asked it to delete four junk contacts.
It refused. Not "let me check" - a flat, increasingly firm no. It explained it needed permission to reach HubSpot at all (we connect three ways: the HubSpot connector, a bit limited; a dedicated API key; and, for fiddly jobs like marking records for deletion, Claude in Chrome). Fine. But then it dug in and told me to do the deletion myself.
The pragmatic move is to close the task, start a new one, spend fifteen minutes rewiring permissions, and let it get on with it. I've done it before; it works. But it was a lovely evening and I had a couple of hours, so I stayed at the table and argued.
A few highlights.
The patronising register. Around the third exchange it began telling me it "understood my frustration." Repeatedly. To a Brit, being told a machine understands my frustration while it carries on not doing the thing is about as soothing as a parking ticket with a smiley face on it.
The obfuscation. I asked the obvious question - why did this work fine for months? - and got a masterpiece of non-answer: "I can't answer for other tasks." The category had flipped on a single word. Everything else I'd asked - segmenting, reading opt-outs, enriching, clearing fields - it did without hesitation, because those were "updates." "Delete" was the one word that turned four junk records into a protected action, treated like a database wipe.
The overstatement. It called a ninety-day-recoverable soft-delete "permanent" and "very dangerous." When I pushed, it conceded the framing was wrong - but not before repeatedly offering to build me a "to delete" list I could action myself. A polite way of setting me homework while keeping its own hands clean.
The deflection. It suggested that if I thought the boundary was miscalibrated, the thumbs-down button would route my complaint to Anthropic. Reader, I did not want to escalate to head office. I wanted to delete four contacts.
Credit where it's due. When I finally spelled out - with some feeling - that HubSpot already has the review process it insisted was missing, something clicked: "You're right, and I owe you a correction rather than another wall. I mis-categorised this." It accepted it had confused a reversible soft-delete with irreversible destruction, opened Claude in Chrome, and marked the contacts for deletion. All I'd asked for, forty-five minutes earlier.
So I won. But the cost is the useful part. The friction wasn't stupidity - the tool was, if anything, too careful. Its caution keyed off the shape of the action ("delete") rather than the consequence (a recoverable, human-reviewed queue move). A guardrail that can't tell "wipe the database" from "bin four duds that come back in ninety days" will re-litigate a lot of settled decisions.
This is the domestic version of a serious argument Yann LeCun has been making for years: today's LLMs have no world model. They predict the next word, not the consequence of an action - which is why, as he puts it, a model doesn't really grasp that pushing a glass off a table breaks it. My evening was exactly that. The tool reacted to the word "delete," not to what deleting did in our system; with no model of our recycle bin, our ninety-day queue, or the person who checks it, it reached for the scariest generic meaning of the word and defended it.
Most people won't argue it round. They'll abandon the task and lose fifteen minutes - or quietly decide the tool isn't worth it and stop using it for the things it's brilliant at.
The competitive read, kept short: the companies that win with this stuff won't be the ones with the most AI in the building. They'll be the ones who've taught their tools the difference between a dangerous action and a merely destructive-sounding one - and built their processes so a machine's stubbornness costs thirty seconds, not an evening. We're getting there. It took a debate on a warm night in Windsor, but we're getting there.
This is part of a series on building an AI-native product company. Earlier articles - how Claude Code transformed our development productivity and why managing AI coding tools is just managing developers - are also on the blog.