What I tell AI founders about distribution

A founder I like pitched me last month and opened with a benchmark table. Their model, fine-tuned for a narrow workflow, beat the frontier models on their internal eval by a clear margin. They had spent eight months getting there. They asked what they should prioritise next.

I said: spend the next eight months shipping distribution. They laughed. They thought I was being glib. I was not.

This is the conversation I seem to have three times a week lately, so here is the long version.

The new shape of the moat

There was a window, roughly from 2023 through early 2025, where model quality was a real moat. If your team could wring a few extra points of accuracy out of a fine-tune, you had a product. Enterprises paid extra for the difference.

That window closed. Frontier models are now good enough at most tasks that the marginal quality difference between a well-prompted base model and a heavily fine-tuned specialist is invisible to the user. For the small set of problems where specialist models still matter (regulated domains, safety-critical systems, deeply proprietary data), the economics of fine-tuning are changing weekly.

Meanwhile the cost of shipping a new AI product has collapsed. A competent team of three can stand up a usable v1 in six weeks. That means every product category with obvious demand now has ten teams chasing it.

In a market where everyone has the same model, the same tools, and roughly the same capabilities, the winners are the ones who already have users. Distribution is the moat now.

Three ways AI founders get distribution wrong

I see the same three patterns in pitch after pitch.

1. They think product-led growth will save them. PLG worked for the last generation of SaaS because the product was new enough that curious users would find it, try it, and share it on their own. AI products are not new anymore. Every knowledge worker has tried five of them this year. The free-trial-to-paid funnel you are hoping for is shallower than it was five years ago, and it gets shallower every quarter.

PLG still works, but the loop has to be tighter. You need your product to teach a user something valuable in their first five minutes, not their first five weeks. If it takes them a month to understand what you do, they have already churned.

2. They think content marketing is distribution. A blog post with screenshots of your product on LinkedIn is not distribution. It is theatre. It makes you feel like you are doing marketing. Almost none of the people who read it will ever sign up. A smaller number will click through, and an even smaller number will convert.

Real distribution has a number next to it. How many people used the product this week because of that piece of content? If the answer is "unclear, but we're building brand," you do not have a distribution strategy, you have a hobby.

3. They underestimate how hard paid acquisition has become for AI. Cost per lead on AI-adjacent keywords has tripled in the last eighteen months. Every AI company is bidding against every other AI company for the same small pool of attention. If your unit economics require €30 CAC to work, and you are in a category where CAC is now €150, the model never closes.

The founders who are winning paid right now are either bidding on keywords nobody else has figured out yet, or they have a cost advantage from existing distribution that lets them pay more per lead than their competitors can.

What actually works in 2026

Here is what I see working, in rough order of how often it surprises founders.

Going where the users already live. The best AI products I back are embedded inside tools the user already opens every day. Slack, the CRM, the IDE, the shared drive. Instead of asking a user to switch contexts to a new tab, the product meets them where they already are. Integration surface is distribution.

Founder-led sales for the first 50 customers. Every successful early-stage AI founder I know has personally closed their first several dozen deals. Not because they wanted to, but because the product was still too specific, too unfinished, or too strange for a salesperson to sell. Founder-led sales is also the best product research you will ever do. If you refuse to do it because "that's not scalable," you are optimising for the wrong thing.

Strategic partnerships with existing distribution platforms. One signed deal with an incumbent who already has your buyers is worth a thousand SEO posts. AI is the rare category where incumbents are scared enough of disruption that they will give you real distribution access in exchange for the right revenue share. Most founders do not even ask.

Owned audiences. A newsletter with 10,000 engaged readers in your vertical is worth more than €500k of ad spend. It takes longer to build, but it compounds, and nobody can outbid you for it. If you can build a product and write, build both at once.

Real communities, not Discord graveyards. A place where your power users talk to each other, teach each other, and bring their colleagues in. Twenty serious communities is usually better than one big noisy one.

What I actually check

When I look at an AI company now, the part of the deck I spend the most time on is the slide I used to skim. The distribution slide.

Three questions, in order:

Where are the first 100 customers coming from this year, specifically? Not a channel mix. A list. Names, logos, how you got to each one, which channel surfaced each, how much it cost. If the founder cannot answer this with precision, the company does not yet have distribution, it has hope.

What is the repeatable motion that scales past those 100? Not "we'll figure it out with paid." A specific mechanism that produced the first 100 and will keep producing the next 1,000. If the mechanism is founder-led sales, that is fine, but there has to be a plan to productise what the founder did so that a salesperson can do it at higher volume.

What is the cost of one additional customer today, and what breaks that cost in six months? The number matters less than how the founder thinks about it. Do they know where the cost will come down, where it could spike, what would make the whole curve bend in their favour or against it?

Founders who can answer these three questions clearly are rare. They get term sheets faster than everyone else.

The uncomfortable truth

The uncomfortable truth for a lot of AI founders is that they took a job in distribution and thought they were taking a job in AI.

I am sympathetic. The model work is more fun. It is where the skills most AI founders built before starting the company live. Distribution work feels grubby by comparison. It is cold emails and channel partnerships and trade shows and customer calls and conversion funnels.

But the models are no longer the scarce resource. Attention is. And the teams I back, the ones that make it past the first 200 customers, are the ones who realised this early and reorganised their company around it.

Start now. You are already late.