Build it yourself · note 5 of 6
You could build this yourself. Here is what to plan for.
The thing product people are not supposed to say. Plus the five things that decide whether your own version survives its first year.
Now the thing product people are not supposed to say. You can build this yourself. The pieces are available, the models are good enough, and with one capable engineer who is curious about agents, a first working version is a matter of weeks rather than months. If you want to build it, build it. You will learn a lot, and you will understand your own market better simply for having thought hard about how to watch it.
What we would tell you, having done it, is that the build is the easy half. Plan for the rest.
Someone has to own the quality bar. The difference between a report people read and one they archive unread sits almost entirely in the instructions, the structure and the standard you hold it to. That is not a one time setup. It is an ongoing editorial job, and it needs a named owner.
Source curation is the actual work. Ten thousand sources is a spam problem. Sixty right ones is intelligence. Choosing, scoring and retiring sources is what makes the output worth reading, and it never finishes.
Cost needs a fuse. AI research is easy to run, and easy to run expensively. Set budgets per run, per month and across the whole system, and meter real usage instead of estimating it. Otherwise you find out at invoice time.
It fails quietly. When a source changes its page layout you do not get an error. You get a calm report saying nothing much happened, which looks exactly like a quiet week. Telling those two apart takes deliberate design.
Then the question worth asking early. Who else can run it. Systems like this tend to live in the heads of the one or two people who built them. If that person is on holiday when the board pack is due, or takes another job, you discover how little of it was written down.
And the unglamorous rest. Email deliverability, unsubscribes, data retention, GDPR, and what happens to all that accumulated memory when someone asks you to delete their data.
If that list looks fine to you, go and build it. Genuinely. And if you would rather have the result than the project, that is what we sell.