Most companies pick their AI vendor the way they get married: slowly, once, hoping they got it right. You compare models, you watch demos, you sign. Then you start pouring in your documents, your processes, your customers.
A year later, a competitor ships something better and cheaper. You price out the switch. And you stay.
So the right question was never "who's best?". It was: how long would it take me to switch?
A vendor that tells you how to leave it
Town is a US startup selling AI assistants to teams. It raised $108 million in a Series B at a $1 billion valuation, co-led by Forerunner Ventures and Menlo Ventures with participation from a16z and First Round Capital - just weeks after coming out of beta. Co-founder and CEO Jean-Denis Greze is the former CTO of Plaid.
In an interview with Platformer, he boils his sales pitch down to a line that should make any executive pause: the data, memories and contacts created in the tool belong to the customer - "to the point that if a better solution ever came along, switching would take a day".
A vendor selling how easy it is to leave. That isn't generosity: Town makes it an explicit competitive position against the big AI labs. The bet: when everyone else locks you in, the vendor that doesn't becomes the obvious choice.
The analogy: your phone number
Before number portability, switching mobile carriers meant switching numbers. Telling your customers, reprinting your cards, losing calls from the people you forgot. The cost of switching had nothing to do with the price of the plan: it sat in other people's address books. Carriers knew it perfectly well.
The day portability became mandatory, competition moved to what actually mattered: price and network quality.
Your AI vendor sits in exactly the same place. The price per million tokens is negotiable, visible, comparable. That's not what holds you. What holds you is everything the tool has accumulated about you.
What actually holds you, precisely
Let's name it. What you may not be able to take with you is your company memory: four layers, none of which have anything to do with the model.
- The indexed corpus - your documents split into chunks and converted into vectors (lists of numbers representing the meaning of a text, so you can retrieve "what talks about the same thing"). Those vectors are computed by one specific model: switching vendors usually means recomputing everything.
- Agent memories - what your assistants have learned about your habits, your customers, your exceptions. Rarely exportable, usually in a proprietary schema.
- Connectors and permissions - who can see what, which agent touches which system. Months of fine-tuning.
- Evaluations - the test cases that prove it works in your context.
The model itself can be swapped in an afternoon: the APIs look alike, the instructions are just text. That's exactly why the AI model is a commodity and your edge is elsewhere. The lock-in is never in the engine. It's in the memory - that is, in your company brain.
The risk nobody looks at: your vendor is also your competitor
Greze raises another, more uncomfortable question. Town runs on models from the big labs, which also sell their own enterprise assistants. "It's risky to be permanently competing with your suppliers," he says. "Is that really the natural state of the world?"
Apply it to yourself. If your application layer, your model and your context storage all come from the same player, you don't depend on a tool: you depend on a strategy that isn't yours. The day that player ships an offer that overlaps your business, or changes its pricing, you have no plan B - your vectors live on their side.
Town's answer is structural: stay neutral, route each task to the best model of the moment, train no model on customer data by default, and contract that with the model providers so inputs and outputs never feed their proprietary models. It's the same logic that makes it worth running part of your AI on open source models: don't confuse the engine with the car.
Four questions to ask before you sign
- Export: can I get my documents, my memories and my history back in a format readable without you - and how long does it take?
- Training: is my data used to train a model, by default? And at your subcontractors?
- Routing: can I switch models without rebuilding anything?
- Real timeline: if I leave tomorrow, what exactly do I lose?
If nobody can answer plainly, you have your answer.
What we do
At Elephant in the Room, we build the company brain on your side: your documents, your agent memories, your access rules, in open and exportable formats, on your infrastructure. On top, multi-model orchestration that lets you switch vendors overnight without rebuilding anything, and business agents that plug in with tight permissions - humans in control of the decisions that matter.
The goal isn't to make you dependent on us. It's to turn your context into an asset that stays yours, whatever the model of the moment.
A vendor you can leave in a day is the only one you can commit to for ten years.
