Instinct, Town and Grok Bot Are All Task Assistants. Liaison Is the Digital Version of You.

So 20VC put out an episode yesterday with the founder of Town, and the title of it is "The $100 Billion AI Assistant Race: Town vs Instinct vs GrokBot." And I listened to it because those are the three names everybody's talking about right now, and about halfway through I'm sitting there going, okay, but what are they actually racing on? Because when you go look at what each one of these things does, in their own words, not the press coverage, they're doing the same job.

So let me just walk through them.

Town raised $55 million from a16z on June 3. Every user gets what they call a Townie, and it plugs into your email, your calendar, Slack, your docs, and it learns how you work and then it drafts your replies and your meeting prep. And the CEO told Platformer in August that you can only use Town if you connect your calendar and your email, which, I mean, that tells you where it lives, right? It lives in your inbox.

Grok Bot is xAI's, it came out August 11, and the way they describe it is "your team of always-on agents. They have their own computer, work inside tools and apps like you do, and keep working 24/7." So each bot gets a cloud computer, it signs into the tools you already use, and it comes back to you when something needs your approval. And then two weeks later they put it in every SuperGrok tier and the Cursor Pro and Teams plans.

And then Instinct, which is the one that's kind of blown up. It's still invite-only, and on August 26 the founder, Noah Shinn, put out the post introducing it, and he said "there are no new interfaces. You can text or call it. It's trained to use a phone and a computer in the same way that humans do." Same day TechCrunch had them at $250 million raised at a $2.5 billion valuation, and the examples in the article are people having it plan a road trip, buy groceries, cancel subscriptions, that kind of thing.

So you've got three companies, right, and all three of them are: connect it to your inbox and your calendar, text it a task, it goes and uses a browser or a phone like a person would, and it comes back when it's done or when it needs a yes from you. That's the product, and it's the product for all three of them. Does that make sense? Like, I'm not knocking any of them, I think they're all going to get a lot of use, but that is one job description.

Everybody agrees the moat is context. Nobody's building the context.

So the investor that led Town's round said it straight out in the funding announcement: "The moat for consumer applications isn't the model. It's accumulated context alongside a magical product experience." And I agree with him completely. We wrote back in December that the models and the agents are commodities and the moat has to be somewhere else.

But then go look at what context any of the three is actually accumulating. Instinct is sitting on your phone, so your texts, your calls, your screen, your location. Town is sitting on your inbox and your calendar, and by design it drafts and doesn't send. Grok Bot is sitting on a cloud computer that's logged into whatever you handed it.

Every one of those is somebody else's surface that they're renting. Your inbox already existed. Your phone already existed. Reading them is not building a context layer, it's plugging into one that Google and Apple built. And the thing about your inbox is it's a record of what people said to you, not a record of what your business did.

Because a context layer over a business is a different animal, and I want to be specific about what it actually takes, because this is the part that's invisible from the outside.

You've got to pull from systems that flat-out disagree with each other, and then decide which one is right. We wrote about that in April, where a customer's own inventory system and their warehouse's report never matched, and one unrecorded return turned out to be 72% of the gap. You've got to work out that Bob Smith in the CRM and Robert Smith in the books are the same person, without merging anything, because both systems keep operating and both keep being right about different things. You've got to know that a payment marked pending is not cash, which is the kind of thing that overstates a month by six figures if you get it wrong. And you've got to carry the record behind every single number, so that when the answer looks off you can click into the actual invoice instead of arguing with a chatbot.

None of that is prompting. It's plumbing, it takes years, and it is the entire reason an agent can say something true about your business instead of something plausible. That's the layer we've been building, and it's the one thing in this whole category that nobody can catch up on in a sprint, because it isn't a model release, it's a thousand decisions about your data.

And Grok Bot has a second problem: it can only ever be Grok

So this is the one I'd actually push on if I were evaluating them, and it's specific to xAI.

Grok Bot runs Grok. That's the whole point of it, right? It's xAI's shop window. And if what you're buying is a smarter chat box, honestly, who cares. But if what you're buying is an outcome, if you're paying somebody to actually collect the invoice, prep the deal, work the schedule, then being tied to one lab's model is a real exposure, and I don't think people are pricing it.

Here's why. The best model for a given job changes every couple of months, and it doesn't change uniformly. One model is better at long reasoning, another one is better and ten times cheaper at the boring mechanical work like classifying and extracting and ranking, another one is better at speech. If you're locked to one lab, you take whatever that lab ships, on their schedule, at their price, and when somebody else jumps ahead on the exact task your outcome depends on, you can't move. You just wait.

We built the opposite on purpose. Model IDs live in one file in our system and nowhere else, every call declares what kind of work it is, and the actual model that serves it gets resolved at runtime from a routing table per customer. So switching which brain does a job is a configuration change, not a deploy. Today we run Anthropic models across the reasoning tiers, we've got a direct adapter to xAI so we can point a job at Grok when Grok is the right answer, we can go through OpenRouter to reach almost anything else, and we run an open model on our own hardware for the highest-volume mechanical lane where paying frontier prices would be silly.

And we use it. Earlier this month we retired our small-model tier entirely because we watched it answer worse on real work than a big model run at low effort, so that lane moved. When we picked the speech model for our voice features we ran the candidates against real recordings and switched to the one that got people's names right. That's not a roadmap, it's just Tuesday, and none of it required shipping code to a customer.

So the asymmetry is this. We can use Grok. Grok Bot cannot use anything but Grok. When you sell outcomes rather than a subscription to one model, that difference compounds every time the frontier moves.

What we're doing differently

So what we built with Liaison starts from the other end of all this. It's not sitting on your phone, it's sitting on the business, so the inbox and the calendar, yes, but also the CRM, the books, the job system, payroll, whatever you actually run on. All of that normalizes into one place, and then instead of waiting for you to give it a task, it works your role.

So overnight it's going through what came in, what went out, who still owes you. It's drafting the reply to the customer who emailed you at 9 PM. It's flagging the customer that went quiet. And then in the morning you've got the list of what it did, what it found, and what it needs you for, and every one of those answers traces back to the record it came from, so checking it takes you a few seconds instead of the twenty minutes you were already spending.

That's the difference between an assistant and what I'd call the digital version of you. The way I put it to our team this week is, it's not the task. It's not the AI that does one thing really well. It's you, the person who knows the business, running the business from the role. The assistant waits for you to ask, and it's good at the thing you asked for. Liaison doesn't wait, because it's sitting on the same information you'd be sitting on if you were the one up at 2 AM, and it works it the way you would.

And even Town's own founder puts a ceiling on the assistant version. He told Platformer that what's out there today gets at "10 or 20 percent of the toil," and that's his number, not mine. You can see the same thing at scale. NBER surveyed almost 6,000 executives in the US, the UK, Germany and Australia over the winter, and about 70% of their companies had adopted AI while 89% said it had made no difference to their labor productivity. Everybody's using it and almost nobody's business runs any differently, because a faster task is still the same task. Somebody still has to think of it, hand it over, and check what comes back.

Now the cloud computer that xAI is selling as the headline, where the bot gets its own machine and signs into your tools, we run that too, and it's genuinely useful for a tool that has no clean API. But it's a lane, not a product. What decides whether it's worth anything is what's waiting on the other side of that login, and which brain you're allowed to point at it.

So the question I'd ask isn't which of the three you'd pick. It's, after a month with any of them, does your business actually run differently, or did you just get a few tasks done a little faster? If you want to see what the other answer looks like on your own numbers, come sit down with me for thirty minutes on your business. I don't want to waste your time, so bring the thing that eats your evenings and we'll look at it together.