Every Monday at 08:00, before I open my laptop, an agent has already done a chunk of my job.
It reads the live portfolio of business needs and ideas, the ones already validated with a business unit. It scans a database of startups. Then it proposes matches: this startup could solve that project. And it goes past a one-line “synergy.”
Each match arrives with a fit thesis, a simulated expert panel where the model argues both sides in character, external context pulled from the web with sources, a draft proof-of-concept with scope and KPIs, and a short list of risks and open questions.
I approve a match by typing one comment: @Match Scout approve. The agent links the startup to the project and never flags that pair again.
This used to be manual work. Matching the startups you meet against what the business actually needs means reading the portfolio, scanning the startup database, and writing a first-pass analysis for every plausible pair. It is slow, it scales badly, and it is the first thing to get dropped when the week gets busy. Now the agent does the first pass and I spend my time on the decision. How that is governed matters, and I come back to it below.
Of course not everything is gold. Not everything it produces is worth reading. Some matches are obvious, some are wrong in ways that take me a minute to name, and a few are confident nonsense wearing the same tidy format as the good ones.
Nobody was replaced. The agent took the reading, not the deciding, and the deciding is the job. On top of that I cannot share more details about the hit rate or what came of any of it, which is the deal when you build inside a company, so discount everything I say about output quality accordingly.
→ What I can give you is my POV and my high-level tips to build an Innovation OS like this.
I am going to start from the tool, which is the exact mistake I tell you not to make three sections from now. If you already have a workspace you like, skip ahead.
Why Notion?
I built this Innovation OS from scratch inside Notion.
Most people file Notion under pretty databases and to-do lists. That was fair until late 2025.
Quick version for anyone who has never opened it: Notion is one workspace where your documents, your databases and your wikis live together. A meeting note links to a project, which links to a task, which links to a person. No app-switching. One source of truth.
Then the agents arrived. So now it became “The AI workspace that works for you”
In September 2025, Notion shipped 3.0. Notion AI became an agent that can do what you can do in the workspace: build pages, build databases, run multi-step work across hundreds of pages, up to 20 minutes on its own.
In February 2026, Custom Agents went live. You build a small team of them, each with its own job, running on a schedule or a trigger, shared across the team. My Monday scout is one of these. They went generally available on 4 May, and this is the part most write-ups skip: they now run on Notion Credits, an add-on at ten dollars per thousand credits, and only on Business and Enterprise plans. Agents are not a free feature you switch on. They are a line item.
The same release added the controls that make them survivable in a real company: per-agent and workspace spending caps, automatic pausing when spend spikes, permissions over who is allowed to create an agent, and a dashboard showing what each one costs. If you are the person who has to answer for this internally, that list matters more than the demo.
In June 2026, Notion opened the workspace to outside agents. Claude Code, Cursor and OpenAI’s Codex can be invited in as tracked participants, assigned work from a shared board, and monitored through their own activity feed and audit trail. Still early, but the direction is clear. Notion is positioning itself as the coordination layer where humans and several AI models work side by side.
Adoption says this is not a demo. Notion reports more than one million Custom Agents created during the beta alone. One early adopter it put forward, the HR platform Remote, says its IT agents triage tickets with over 95% accuracy and resolve more than a quarter of them without a human, saving around twenty hours a week. Vendor-sourced numbers, so treat them as directional rather than proven.
You also get to pick the model. Claude, GPT, Gemini, Grok (…) all run inside the same agent, so you can put the expensive one on the reasoning and the cheap one on the tidying.
What I actually built
An innovation hub in Notion. Intake, scoring, scouting, governance and pilots in one place, with agents running on top of it. Months of work, most of it spent on the design rather than the software.
That is the last I will say about my setup, because it is the least useful part of this. Yours should look different. What transfers is the order of questions I would work through if I started again, and I got that order wrong the first time by thinking about tools too early.
Start with the tasks, not the tool. Write down what the team actually does in a week. Not the mandate, the tasks. Screen inbound ideas. Run a needs conversation with a business unit. Find startups against a need. Chase a pilot that has gone quiet. Report upward. The list is usually shorter than people expect and duller than the org chart implies, and most of it turns out to be remembering and routing rather than having ideas. That is worth sitting with, because remembering and routing is exactly what software is good at.
Clarify the boundaries and then ask which of those tasks need to be remembered. Some work only needs to happen. Some work needs to still be findable in eight months, when someone asks why that startup was dropped and everyone who was in the room has moved on. This question produces the architecture, and it is the one most teams skip.
Then look at what is already in the building. Before anyone opens a procurement process, find out what the company already pays for and what IT has already cleared. A licence you already hold skips vendor selection and procurement, often the difference between having a system this quarter and having one next year. (It does not skip the AI and data protection review). Budget through the back door is survivable. Data governance through the back door is not.
Also, before anyone asks the obvious question, because it is the right question. This runs inside the workspace the company already uses, under the same information classification and data protection rules that govern any other system holding this kind of material, and with the AI review track applied to it like anything else. Material that cannot sit there does not sit there. If you are building something similar, the conversation with your information security and data people is not an obstacle to route around. It is the thing that makes the system allowed to exist, and it is the reason this one still does.
Build or buy, and how to tell which
There are good innovation management platforms. Whether one is right for you depends on things that are specific to you, so what follows is criteria, not a recommendation.
Build if the process is not settled. A vertical tool encodes somebody else’s model of how innovation should flow, and buying one before you know your own shape means inheriting their assumptions and calling it a decision. Building it yourself is slower and forces you to make every design choice on purpose, including the ones you would rather have avoided.
Two more (personal) reasons.
I simply like the interface. I spend hours a week in this thing, and a tool you enjoy opening is a tool that gets opened. That sounds soft until you remember that adoption is what kills these systems, and I am the first user who has to keep turning up.
The second is the shape of the work. Building with several agents feels less like configuring software and more like working next to a fast colleague: I describe what I want, one drafts the database, I correct it, another cleans up the navigation behind it. That loop is what made building from scratch realistic at all. A year ago my honest answer to build-or-buy would have been buy, because building would have cost a year of evenings.
Buy instead when you need a validated external startup database and a solid trend scouting network, because rebuilding those is not a side project. Buy when what you actually need is the consulting that comes with the platform. And buy when you are processing thousands of ideas a year, because at that volume the ergonomics of a general-purpose tool break down and you will spend your life maintaining formulas instead of running the function.
The agent is only as good as the system
Here is the lesson, and it is popular in some AI circles.
AI does not fix a broken process. It amplifies whatever is already there. I cannot give you my hit rate. I can tell you where the misses came from. Almost every bad match traced back to something missing in the record: a project with no problem statement, a startup tagged by sector and nothing else, a strategy that lived in a deck instead of a field. Point the same agent at a messy workspace and you get confident nonsense, faster.
Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, and the three reasons it gives are escalating costs, unclear business value and inadequate risk controls. Not one of them is a model problem. They are all failures to design the system around the agent.
The engineering side of the industry reports the same thing. Anthropic published its findings on long-running agents in November 2025, and the failure modes it documents are the ones I hit: the agent tries to do everything at once, declares victory too early, marks work as done without checking, and leaves nothing behind for the next run to pick up.
The point is that governance, processes, workflow design sit underneath all of it. Information classification, NDA flows, vendor security checks, procurement processes, a separate review track for anything involving AI. Boring, and the reason a system like this survives contact with a real company.
5 things I would tell anyone starting to build their AI-powered Innovation OS:
Start from the task you would pay to never do again. Gabriel Hubert, who runs the AI company Dust, calls it an anti-to-do list: after every annoying task, ask how you never have to do it again. That question picks better first agents than any list of use cases.
Build the record before you build the agent. If the underlying data is not structured and current, you do not have an agent project, you have a data project wearing an agent costume. That is a less exciting thing to put in a steering committee update and it is the work that decides whether any of the rest happens.
Give every agent a parent. One named human who owns keeping it accurate as the underlying data changes. An agent with no owner degrades quietly, and people stop trusting it long before anyone admits it is broken.
Decide early who is allowed to build one. This feels premature until the month somebody's private agent starts sending your business units a competing version of the truth, on a budget nobody approved. Permissions and spending caps are boring to set up on day one and impossible to retrofit politely.
Check what it can see before you check what it can do. An agent inherits the data access of the space it was built in, and keeps that access no matter who talks to it. In a company with information classification and NDAs, that is the failure that actually hurts.
So what?
In my own case, what made it work was connecting innovation strategy, intake, scoring, scouting and governance into one OS that does not break at the handoffs. Not one clever feature. Most innovation processes die in exactly those gaps, in the space between the tool where the idea was captured and the tool where someone was supposed to act on it. So, make sure to start from a strong system, before pretending that AI (or Notion) will solve everything.
Also, if you have been treating Notion as a place to park notes, look again. It quietly became a place to design how people and AI work together.
Start with the database quality and workflow, not the agent. Build one record you would trust an agent to read, and structure it as if one already were. If your plan has no agents, that is still the whole job. The agent is a switch you flip later, and the flip takes an afternoon. Give it one job, on a schedule, with a clear limit and a way to write its decision back. Then see what shows up Monday morning.
Thank you for reading me. See you next time,
Davide
Disclaimer: I write here in a personal capacity. Views are my own and do not represent my employer.
Bonus reads:
Did you enjoy this edition? Share it with anyone you like and give it a heart. Subscribe to receive Open Road Ventures’ posts every week.
The Monday 8am cadence is the underrated detail here. A lot of agent setups fail not because the matching logic is bad but because there's no forcing function to actually review the output, so it just piles up unread. Curious how you handle the case where the agent's match is wrong two weeks running, does it self-correct or do you have to manually retune the prompt?