What rebuilding Guru for the AI era taught us about enterprise AI accuracy, platform shifts, brand risk, and why per-seat pricing breaks down for AI agents.
Traditional software is being compressed into “tool calls” for Agents. Capabilities that used to justify an entire company now become a checkbox inside a model or a platform, and the gap between "this is our product" and "this is a prompt" keeps shrinking.
Every software company felt some version of this in 2023 and 2024. The most common response was to add AI to what already existed. A summarize button, a chat box in the corner, a sparkle icon on the toolbar.
Not long ago Mike Gadsby, who co-founded O3, told me what stuck with him watching Guru these last couple of years was that we reinvented the business and the team from the ground up. It left him asking a bigger question: what does it actually mean to be AI native, and where do we go next?
That question is why O3 is putting on the 1682 conference this fall, and why we were glad to sign on as presenting sponsor. This post is my early attempt at an answer.
The bet we made at the end of 2023
We decided the durable problem in enterprise AI was accuracy.
Accuracy is a strange problem because it stays invisible until it isn't. You connect a model to your company's SharePoint site, ask it something a new hire would ask, and get back a confident, well written answer sourced from a document from 2021 that three people on the team have quietly known was wrong for years. The model did its job. The knowledge failed it.
Nobody discovers this in a demo. Everybody discovers it in production. Getting to information stopped being the hard part a while ago. Trusting what comes back is the hard part, and that turns out to be a knowledge problem rather than a model problem.
Two platform shifts, one answer
I have built through two of these now. At Boomi we bet on the move from on premises software to the cloud. At Guru we are betting on the move from SaaS to AI and agents.
Both times the tempting answer was the cheap one. At Boomi, take the software you have and host it somewhere. At Guru, put AI features on top of the product you already sell. Both times the right answer was to rebuild for the new world, which is expensive, slow relative to how fast the market is moving, and terrifying when your existing product is paying the bills.
That is the part that does not get easier the second time. These markets move at a frenetic pace, and the pressure to ship something quick and call it a strategy is enormous. The right answer is still the hard answer.
Being early is a tradeoff, not a free advantage
We named the accuracy problem before our buyers were asking about it. In 2024 the question in most companies was whether AI could be used for work at all. Accuracy is not a question you can ask at that stage, because you only hit it after you have pointed something at your own messy, contradictory knowledge and watched what happens.
Getting there early gave us a real head start, and not because we wrote more code. Code generation is close to instant now, and a year of engineering output is not the moat anyone thinks it is. The head start is in domain knowledge. Which workflows actually matter to a company. Where the knowledge really lives. How an agent has to collaborate with employees for any of it to work in production. That takes real time and there is no way to compress it.
The other side of the tradeoff is that the stakes go up. Build years ahead of the market on the wrong problem and you have spent years on the wrong problem.
What I would tell another founder is to get very good at finding your ICP before your problem goes mainstream. When the market is not yet asking the question, the small number of companies who are asking it are the whole signal, and they are worth far more of your attention than any analyst report. It is part of why we were glad to support 1682. Our friends at O3 World are curating a room of thoughtful people who are asking lots of questions and care about real AI outcomes, and that is the crowd to be paying attention to right now.
Your vision can outrun the models
The hype consistently oversells what models can do today. The counter muscle is unglamorous but critical: test the workflows that matter to you, on your own data, and find out for yourself where they break.
The harder half of that discipline is retesting. Capability moves quickly, and a workflow that failed badly six months ago may work now. In one example we wanted an agent that could sit in Slack and in meetings and pull out the knowledge worth keeping, and for a long stretch the models could not do it in any reliable manner. That changed this year. It would have been easy to miss, because our own last test said it was not possible.
Your test results go stale faster than you expect them to. This world is changing faster than we realize!
A good brand can work against you
The standard advice is that selling to your existing customers is easier than selling to new ones. Our experience says not always.
Our brand with our installed base was a wiki. A good one, one people genuinely liked, and one that is a different product from what Guru is today. That association was earned over a decade and it does not update just because we built a new product. Meanwhile someone meeting us for the first time attaches our name to what we actually do now, with nothing to unlearn.
The lesson is that marketing to your existing customers matters as much right now as acquiring new ones. Same effort, same rigor, same campaigns. If your product has genuinely changed, the people already paying you are a new audience and you have to treat them like one.
AI adoption is as hyped as everything else
Put a general purpose chat box in front of a company and prepare to be underwhelmed. A small group of people who already tinker with AI get enormous leverage from it. A larger group uses it occasionally and not very well. Most people never really start, because a blank text field asks you to invent your own use case, and that is not something most people have time to do in the middle of their actual job.
The dashboards say the rollout worked. But nothing really moves in the business.
Our learnings led to multiplayer agents. A small group who genuinely understands the work builds a curated, purpose built agent, and everyone else just uses it. No inventing a use case, no learning to prompt, no wondering whether AI applies to their job. The expertise gets built once and consumed by many.
The design constraint that matters is not making your frontier users more powerful. It is making everyone else productive without asking them to become frontier users.
Pricing has to follow whoever is doing the work
Per seat pricing is clean and simple when a human operates the tool. It stops making sense when the system is doing the work, mining sources and verifying knowledge while nobody is logged in.
Two things we learned moving to consumption. The first is that every billable event has to be explainable in one plain english sentence. If you cannot say what the customer got for it, do not charge for it.
The second is that the education effort is much larger than you plan for. Enterprise software buyers already know what a seat is and how it behaves. Consumption units are specific to your product, which means you are teaching a new concept to every buyer, every finance team, and every renewal conversation. That teaching has to be clear and completely transparent, and it is not a one time exercise.
What I keep coming back to
Almost none of what we learned was about the models.
The models were the easy part, and they got better on their own while we worked. The hard parts were the brand we had built, the buyers who were not ready, the users who needed a reason to start, and the pricing conversation that had to be taught from scratch. Every platform shift has been like this. The technology arrives first and everything around it readapts. Regardless of how fast technology moves, it always comes back to humans ability to adapt to it.
At 1682
Almost everything in this post is unfinished. The pricing conversation is ever evolving. Certain calls are still in progress. That's the stuff that doesn't fit into a blog post.
That's why we were glad to sign on as presenting sponsor of 1682, a one day conference our friends at O3 World put on. It's about the business of innovation…what actually works, what doesn't, and why.
The theme this year is AI for all of us, which happens to be the problem I care most about. The hard part was never making your best people more powerful. It's making everyone else productive at scale. I'll be there talking through where we are today, and where this whole world of AI and agents is heading next.
See you there!
Traditional software is being compressed into “tool calls” for Agents. Capabilities that used to justify an entire company now become a checkbox inside a model or a platform, and the gap between "this is our product" and "this is a prompt" keeps shrinking.
Every software company felt some version of this in 2023 and 2024. The most common response was to add AI to what already existed. A summarize button, a chat box in the corner, a sparkle icon on the toolbar.
Not long ago Mike Gadsby, who co-founded O3, told me what stuck with him watching Guru these last couple of years was that we reinvented the business and the team from the ground up. It left him asking a bigger question: what does it actually mean to be AI native, and where do we go next?
That question is why O3 is putting on the 1682 conference this fall, and why we were glad to sign on as presenting sponsor. This post is my early attempt at an answer.
The bet we made at the end of 2023
We decided the durable problem in enterprise AI was accuracy.
Accuracy is a strange problem because it stays invisible until it isn't. You connect a model to your company's SharePoint site, ask it something a new hire would ask, and get back a confident, well written answer sourced from a document from 2021 that three people on the team have quietly known was wrong for years. The model did its job. The knowledge failed it.
Nobody discovers this in a demo. Everybody discovers it in production. Getting to information stopped being the hard part a while ago. Trusting what comes back is the hard part, and that turns out to be a knowledge problem rather than a model problem.
Two platform shifts, one answer
I have built through two of these now. At Boomi we bet on the move from on premises software to the cloud. At Guru we are betting on the move from SaaS to AI and agents.
Both times the tempting answer was the cheap one. At Boomi, take the software you have and host it somewhere. At Guru, put AI features on top of the product you already sell. Both times the right answer was to rebuild for the new world, which is expensive, slow relative to how fast the market is moving, and terrifying when your existing product is paying the bills.
That is the part that does not get easier the second time. These markets move at a frenetic pace, and the pressure to ship something quick and call it a strategy is enormous. The right answer is still the hard answer.
Being early is a tradeoff, not a free advantage
We named the accuracy problem before our buyers were asking about it. In 2024 the question in most companies was whether AI could be used for work at all. Accuracy is not a question you can ask at that stage, because you only hit it after you have pointed something at your own messy, contradictory knowledge and watched what happens.
Getting there early gave us a real head start, and not because we wrote more code. Code generation is close to instant now, and a year of engineering output is not the moat anyone thinks it is. The head start is in domain knowledge. Which workflows actually matter to a company. Where the knowledge really lives. How an agent has to collaborate with employees for any of it to work in production. That takes real time and there is no way to compress it.
The other side of the tradeoff is that the stakes go up. Build years ahead of the market on the wrong problem and you have spent years on the wrong problem.
What I would tell another founder is to get very good at finding your ICP before your problem goes mainstream. When the market is not yet asking the question, the small number of companies who are asking it are the whole signal, and they are worth far more of your attention than any analyst report. It is part of why we were glad to support 1682. Our friends at O3 World are curating a room of thoughtful people who are asking lots of questions and care about real AI outcomes, and that is the crowd to be paying attention to right now.
Your vision can outrun the models
The hype consistently oversells what models can do today. The counter muscle is unglamorous but critical: test the workflows that matter to you, on your own data, and find out for yourself where they break.
The harder half of that discipline is retesting. Capability moves quickly, and a workflow that failed badly six months ago may work now. In one example we wanted an agent that could sit in Slack and in meetings and pull out the knowledge worth keeping, and for a long stretch the models could not do it in any reliable manner. That changed this year. It would have been easy to miss, because our own last test said it was not possible.
Your test results go stale faster than you expect them to. This world is changing faster than we realize!
A good brand can work against you
The standard advice is that selling to your existing customers is easier than selling to new ones. Our experience says not always.
Our brand with our installed base was a wiki. A good one, one people genuinely liked, and one that is a different product from what Guru is today. That association was earned over a decade and it does not update just because we built a new product. Meanwhile someone meeting us for the first time attaches our name to what we actually do now, with nothing to unlearn.
The lesson is that marketing to your existing customers matters as much right now as acquiring new ones. Same effort, same rigor, same campaigns. If your product has genuinely changed, the people already paying you are a new audience and you have to treat them like one.
AI adoption is as hyped as everything else
Put a general purpose chat box in front of a company and prepare to be underwhelmed. A small group of people who already tinker with AI get enormous leverage from it. A larger group uses it occasionally and not very well. Most people never really start, because a blank text field asks you to invent your own use case, and that is not something most people have time to do in the middle of their actual job.
The dashboards say the rollout worked. But nothing really moves in the business.
Our learnings led to multiplayer agents. A small group who genuinely understands the work builds a curated, purpose built agent, and everyone else just uses it. No inventing a use case, no learning to prompt, no wondering whether AI applies to their job. The expertise gets built once and consumed by many.
The design constraint that matters is not making your frontier users more powerful. It is making everyone else productive without asking them to become frontier users.
Pricing has to follow whoever is doing the work
Per seat pricing is clean and simple when a human operates the tool. It stops making sense when the system is doing the work, mining sources and verifying knowledge while nobody is logged in.
Two things we learned moving to consumption. The first is that every billable event has to be explainable in one plain english sentence. If you cannot say what the customer got for it, do not charge for it.
The second is that the education effort is much larger than you plan for. Enterprise software buyers already know what a seat is and how it behaves. Consumption units are specific to your product, which means you are teaching a new concept to every buyer, every finance team, and every renewal conversation. That teaching has to be clear and completely transparent, and it is not a one time exercise.
What I keep coming back to
Almost none of what we learned was about the models.
The models were the easy part, and they got better on their own while we worked. The hard parts were the brand we had built, the buyers who were not ready, the users who needed a reason to start, and the pricing conversation that had to be taught from scratch. Every platform shift has been like this. The technology arrives first and everything around it readapts. Regardless of how fast technology moves, it always comes back to humans ability to adapt to it.
At 1682
Almost everything in this post is unfinished. The pricing conversation is ever evolving. Certain calls are still in progress. That's the stuff that doesn't fit into a blog post.
That's why we were glad to sign on as presenting sponsor of 1682, a one day conference our friends at O3 World put on. It's about the business of innovation…what actually works, what doesn't, and why.
The theme this year is AI for all of us, which happens to be the problem I care most about. The hard part was never making your best people more powerful. It's making everyone else productive at scale. I'll be there talking through where we are today, and where this whole world of AI and agents is heading next.
See you there!
Vous voulez apprendre comment Guru peut fonctionner pour vous et votre équipe ?
