How to use AI without hollowing out your company

This guide will keep you in the small cohort where your institutional wisdom grows, and your message, context, and understanding become your moat – not things you feed to the machine, and slowly stop understanding.

The first law

AI has made it trivial to sound like you’ve mastered a market you’ve never touched. Sounding competent and being competent have come apart – and for us, telling them apart is literally the job: separating the founder who lives their market from the one running a convincing impression of it. The test we hold founders to, we hold ourselves to first.

Before you ship something that’s passed through the model, ask:

Could the person who owns this output defend it, off-script, to someone who knows the subject?

Most people use AI to fill gaps in their knowledge or ability. It’s why most AI output is value-less – because it didn’t start with them.

Your reader – a customer who lives the problem or an investor who’s seen a hundred pitches – knows the subject better than your document.

Push one question past where the output stops. If you can’t answer it, you’ve got the illusion of understanding. It will crumble under the slightest pressure, and your credibility is gone. AI-evangelist founders are falling here because they forgot Claude is just a copilot.

If you can answer it, that means you understand the thing, AI earned its keep, and you gain a level of credibility few people are bothering to go to.

Every principle below this should abide by this law. Only when you can do it yourself should you let AI do it.

The eight principles

01 · Decide where AI runs by who reads it next

Gating AI by how important or sensitive the topic is is the wrong axis.

Gate it by who reads the output next.

If it’s something only you’ll see, let AI go at it. If it’s something your team relies on, AI’s maximum involvement should be drafting. If it’s something a customer or investor sees, AI can help, but you must own and understand every word.

As the audience widens, AI autonomy should contract. “Human in the loop” isn’t one setting – it’s a dial, and the reader sets it.

Pressure to ship will always threaten to skew your resolve. Return to this matrix, and adjust the deliverable, not the technology. Out of time on a Tier 1 piece? Ship a shorter version you wrote, not a longer one AI padded out.

Stanford and BetterUp found 41% of workers have been handed AI work that looked fine but lacked the substance to move the task on, and rated the sender less trustworthy (42%) for it. In a small firm that runs on trust, that’s the real cost.

Do

Let AI run wild on throwaway work only you’ll see

Don’t

Ship AI’s voice to a customer because the topic felt low-stakes

Ask

Who reads this next? Their answer dictates how far you trust AI with it.

02 · You shipped it, you own it

Everything that leaves the company needs one person accountable for the words.

Auto emails, chatbot responses, the sequence that fires at 2am – someone should take responsibility for everything “said”.

And as soon as it hits something that needs judgement, it should concede and hand over to a human.

Nobody gets to blame the bot. “The AI wrote it” isn’t a defence – you shipped it, you own it. Accountability doesn’t transfer to the tool just because the tool did the typing.

Do

Route auto-replies through a person who makes the final edit/call – every time

Don’t

Let anyone hide behind “the AI did it”

Ask

Whose name is on this, and could they stand behind it?

03 · Agents are processes (not people)

Talking about their “AI employee” has become a flex for some founders – and a whole industry is pushing it. We’re partially to blame. Founders have raised untold sums on the promise that enterprises can shed headcount and run on agents instead, and to sell it, they dress the software up as people. Artisan put “Stop Hiring Humans” on billboards across San Francisco and New York to market an AI sales rep they named Ava. Drive the 101 in SF and the same move is everywhere: give the bot a human name, then sell it as the thing that ends human jobs.

There’s something pretty dystopian about trying to give humanity to software that is being marketed as a way to replace humans. Resist it – in how you build and how you describe it. The moment a “what” becomes a “who”, accountability slides to something that can’t be accountable.

An agent is a tool inside a process. A smart spreadsheet, not a staff member – calling a bundle of code “IBM Bob” doesn’t make it a colleague. “It” didn’t ship – a person shipped using it. Agents shouldn’t be tracked like other names in your org chart, but by cost and error rate.

The same goes outward, in your own product. If it’s a bot, let it be a bot openly – no persona pretending a person is typing, no name dressed up as a colleague who doesn’t exist. A customer will forgive a bot; they won’t forgive discovering “Sarah from support” was never real.

None of this bans a name. Ours is called Poppy (for the tall poppy, not a person FWIW) and she says she’s a bot in the first breath. The line isn’t whether a tool has a name; it’s whether the name fakes a human or takes the credit.

Do

Name a tool if you want – but let it introduce itself as a bot, keep a human’s name on the output, and write new agents up like onboarding a hire.

Say

“Sam used Poppy to draft the reply”

Don’t

Give an agent a human face it doesn’t have, put it in your headcount or org chart, talk about “AI employees” to look advanced, or hand it the credit: “Poppy’s been a lifesaver this quarter”

Ask

Does your team talk about AI like a tool, or like a person?

04 · Understand it, or it’s slop

AI works well when it has a million data points. Your business is just one data point.

So don’t trust it to know or advise on you, singular. Do use it to tutor you on markets, plural.

Ask it for follow-ups, challenge its answers, make it re-explain until you actually get it.

The trap is shipping what you don’t understand: pitching a market size you can’t defend, running a strategy you can’t iterate when it breaks. It’s a very confident impression of expertise, and it holds right up until the first question the page didn’t anticipate. Your investor can prompt the same model you did – being more than a wrapper around it is your whole job.

Do

Use AI to get smart on something fast, then explain it in your own words

Don’t

Present analysis you couldn’t discuss without reading off the page

Ask

Could you defend this with the document face down to someone who’d know if you were bluffing?

05 · Writing is thinking

Principle 04 is about understanding the subject. This one’s about the tool that gets you there – writing.

Some documents exist to carry a finished thought: a memo, an update, a reply. Others exist to produce one: a strategy doc, a spec, a retro. In the second kind, the writing is the thinking – the iterations, the cuts, the “actually, that’s wrong” halfway down page two. Hand that draft to AI and you haven’t saved the work, you’ve skipped it. The thinking was the point; the document was only ever the proof it happened.

Cal Newport frames it as cognitive fitness: the strain of working an idea through on the page is how the muscle grows, and offloading it to a model is the mental equivalent of taking an electric scooter through boot camp – easier today, useless when it counts.

So we split the document in two. When the goal is the thinking, we ask for a one-pager of someone’s raw thoughts – rough, unpolished, theirs – with the AI research, facts and figures pushed to an appendix.

Do

Write the raw version of a thinking document yourself first; let AI’s research and data sit in the appendix

Don’t

Outsource the draft of the doc that was supposed to make you think

Ask

Is this meant to report a thought, or to produce one? If it’s the second, write it yourself.

06 · Everyone still hires product marketers

If the model could handle the messaging, Anthropic and OpenAI would be first to skip the hires. But the builders of the models themselves are actively staffing product marketers (and Copywriters, and Heads of Narrative).

Because the same model and the same prompt gets you and your competitors to the same message. Everyone arrives at the same dangerous territory: the crowded middle.

AI can still only predict the most likely next words. It doesn’t know the outlying message that’s worth saying.

Launching with a take only you could have avoids two of AI’s tripwires: genericism and bloat. The model shoots for generic and pads it with more generic because that way it covers all bases.

A product marketer knows how to meaningfully connect your insight to a market need, communicate it in a customer’s own language, and iterate it to find cut through. Once you’ve cracked it, you can start feeding it to a model.

Do

Put a product marketer’s point of view on the message, and make the model work from their positioning

Don’t

Ship a model’s generic draft as your positioning or your voice

Ask

If your competitor prompted the same model, could they get the same thing? If so, it isn’t yours yet.

07 · If it’s not written down, it doesn’t exist

AI can only reason with the context it can reach.

In most startups, the inside knowledge that would compound their edge isn’t written down – it’s in the founding team’s heads and WhatsApps and endless meetings nobody recorded. That makes it invisible to future hires, and to every tool you point at the business.

So default to “open”: every decision and the why behind it in shared channels, meetings recorded and transcribed, processes written up once they’re cracked so anyone, person or agent, can run it.

Some of the best software companies already run this as policy. Photoroom enforces a company-wide “No DMs” rule: 74% of their Slack messages land in public channels, against 7% in DMs. Zapier turned it into a game – a monthly leaderboard of whose messages happen in channels versus DMs. As their CEO puts it: “transparency’s a team sport and disinfectant.”

Context in the open is how a company’s intelligence compounds instead of the workload just multiplying – one of the truest moats an AI-native company can have.

Do

Cultivate an open culture with real mechanisms – a No-DMs default, core values people actually know, recorded and transcribed calls.

Don’t

Let your company’s memory die in private threads and unrecorded calls

Ask

Could a new hire ask the company AI “Where is this doc?” and have a good chance of getting “right here” back?

08 · Match the length to the argument

AI pads, and it can’t help it. A model billed by the token has every reason to keep going and no instinct for when to stop – ask it to shorten something and it’ll somehow hand it back longer.

You’ve been on the receiving end of it – the War and Peace of slop. A colleague’s “quick summary” that arrives at four thousand words marked “no rush”, which they’ve plainly not read either.

Two fixes.

The prompt: most bloat is a prompt failure, not a model one. A vague ask gets a hedged, cover-all-bases reply; give it a role, an audience, a format, and a stop condition and it tightens up.

And the mindset: length is not a proxy for effort – usually the reverse. Pascal apologised for it in 1657: “I have made this longer than usual because I have not had the time to make it shorter.” Cutting is the work, and it’s the part the model leaves undone.

One caution the other way: there’s no lossless rewrite of natural language. Every time you hand a draft back to the model to trim, it nudges the words toward its idea of your meaning, not yours – so do the cutting yourself, or check that what survived still says what you meant.

The line we hold to: if you wouldn’t have written it this long without AI, don’t ship it this long with AI.

Do

Prompt for the format and length you actually want, then cut what the model padded

Don’t

Mistake volume for rigour, or ship a bullet’s worth of point as a page

Ask

Would this be this long if you’d written it by hand?

What this can’t decide for you

How much AI help is too much in a customer or investor reply.

Rough rule: fine when the value is something they couldn’t have got from a stock model in five minutes. When the only value is that you had a better model than they did, you’re a wrapper.

How to add real insight without grinding to a halt.

Our default: AI drafts, a human rewrites the parts that need an actual point of view – sometimes the whole thing, sometimes a single sentence. The skill is telling which.

Where the line sits between too proud and too cheap.

Too proud to use AI well is a failure. Shipping output any stock model could have produced is also a failure. Aim for the middle, and expect to miss both ways sometimes.

Why it’s worth the effort

“Slop” was Merriam-Webster’s 2025 word of the year. People now recognise its texture on sight.

It’s eroding trust on a colossal scale. People don’t just discount the content, they discount you.

Even writing that sentence feels wrong, because “it’s not X, it’s Y” is exactly the tell we’ve trained ourselves to hate. Hunt them hard enough and you start flinching at your own voice.

The platforms are turning too: LinkedIn shipped a “seems like AI slop” report button, Snapchat cut fully-AI clips from its feed. And the more people learn about AI, the less they trust it.

When someone looks you up, what you’ve published is what they get. It’s what gets quoted back about you for years. If this body of content is authentic and owned, trust builds. If AI had too much autonomy over the years, trust hollows out – along with the base of knowledge you could’ve built.

Trust – from customers who buy and investors who back – is your scarcest asset. The part you can’t prompt and tokens can’t buy. Preserving it comes with added cost. It’s slower, output is lower, more human hours are needed per thing that ships.

It can only exist where someone has something to lose.

When you’re building AI-native, your prerogative shouldn’t be “how can we use AI to guide our judgement?”, but “how can we use our judgement to guide our use of AI?”.

Deleting an em dash doesn’t implant a point of view.