When we started building Wato, we assumed the interesting problems would be about intelligence: how to make an agent reason more carefully, plan further ahead, and recover from its own mistakes. Most of what we actually learned turned out to be about companies instead. How they really work, how much change they can absorb, and how wide the gap is between the way AI is sold and the way it gets adopted on the ground.
The tools on the market today are mostly built for two kinds of buyer. There is the AI-native startup that will happily reorganize itself around any promising new system, and there is the Fortune 100 that can afford a dedicated team to make one work. The large middle is left holding products it cannot deploy and advice it cannot act on. We built Wato for that middle, and the lessons below are the ones that shaped it.
Deployment Is The Hard Part
The difficulty we kept running into had very little to do with the model and almost everything to do with the environment it had to live in. A company arrives with its own tools, its data scattered across all of them, and years of accumulated habits about how work actually gets done. Dropping an agent into that setting and expecting it to be useful is a far harder problem than improving the agent in isolation.
Many AI products quietly sidestep this by assuming the company will reorganize around them, adopting a new system of record and migrating their work into it over time. The companies we care about will never do that, and they are right not to. We designed Wato to sit on top of the tools a company already runs, so that adopting it asks for no migration and no new place to keep everything. The shortest path to value was to leave the existing stack where it stood and reach the work where it already happens.
People Want To Talk To Their Tools
When we talk about agents here, we mean the AI tools people are already using to get work done: Claude Code, Codex, Cowork, ChatGPT, and the others like them. The people who get the most out of these tools are rarely the ones willing to sit down and wire up integrations by hand. On a normal team, someone has a question or a task, wants to express it in plain language inside whichever tool they already have open, and wants to see it handled. That instinct is the entire appeal of working this way, and a surprising number of products manage to bury it under configuration screens and setup steps.
We leaned in the opposite direction and exposed everything through MCP, the open standard these tools already speak, so that a company's systems become available right inside the tools people already use and reachable through ordinary language. Someone can ask for what they need and the relevant systems respond, without anyone first learning where a setting lives or how a connector is supposed to be assembled.
Indexing Everything Is A Trap
The instinct across the industry is to ingest a company's entire data footprint and build one comprehensive index over it, on the theory that an agent can only help if it can see everything. We came to believe this is a mistake, and on two counts. The first is security. Once a system has ingested everything, it will eventually surface material that was never meant to travel, and the blast radius when something goes wrong is the whole company at once. The second is more mundane but just as real. Comprehensive indexing carries a permanent maintenance burden around freshness, duplication, access control, and drift, and those problems tend to compound as the corpus grows rather than settle down.
We built memory to be deliberate instead. It retains what has been reviewed and is genuinely relevant to the work, which keeps the exposed surface small and, as a welcome side effect, tends to make the answers better, because the signal that matters is not buried under everything the system happened to be able to reach.
Governance Becomes The Whole Game
Below a certain size, governance feels like overhead a team can defer. Above it, governance quietly becomes the substance of the problem. Once an organization is large enough, the operative question stops being whether an agent can perform a task and becomes who is permitted to perform it, against which systems, and whether the company can later prove what happened.
This only gets harder because people never settle on a single tool. They use several at once, each with its own notion of identity and access, and consolidating all of that into something a company can reason about and enforce is genuinely difficult. It is also work the model labs are not well positioned to take on, since their domain is the model itself, while governance lives in the space between the model and a company's own systems. That space is where we chose to operate.
Most People Are Not Power Users
Adoption inside a company is far less even than the marketing around AI suggests. On almost every team, a few people develop real fluency and quietly build workflows that save themselves hours, while everyone around them keeps using AI in shallow, occasional ways or avoids it altogether. The value those few create tends to stay locked inside their own accounts and personal habits, invisible to the rest of the organization.
We came to see the central opportunity as one of extraction and distribution: noticing what the fluent few have worked out, capturing it in a form other people can reuse, and spreading it across the company so the majority inherit the benefit without having to become experts themselves. Improving the underlying model does very little for a team whose real bottleneck is that its best practices live in three people's heads.
The Mid-Market Is Underserved
All of this pointed us toward a part of the market that nearly everyone else overlooks. The pressure to act is not lost on these companies. In conversations with business and industry leaders, the same realization keeps surfacing: becoming AI-native has stopped feeling optional, and there is a shared sense that the companies which fail to make the shift will be left behind by the ones that do. The urgency is genuine and widely felt.
What has been missing is a path the middle of the market can realistically take. Mid-sized companies cannot stand up an internal AI team, and they are in no position to undertake the top-to-bottom overhaul that becoming AI-native is usually taken to mean. The retraining, the organizational friction, and the cost do not justify themselves at that scale, and the disruption of rebuilding around AI is a price these companies cannot reasonably pay. Yet they are often the ones with the most to gain and the least equipped to capture it on their own.
The rest of the industry has organized itself around the two ends of the spectrum, serving the startups that are AI-native by birth and the largest enterprises that can buy their way there, which leaves the middle with tools it cannot deploy and counsel it cannot use. That middle is our wedge, and in hindsight every other lesson here is really an answer to a single question: how do you bring agents to a company that cannot reinvent itself to receive them?