The question came from a friend yesterday in whatsapp. I'm using AI a lot, but it's not really taking anything off my plate, but giving me more work.
Is this the same for me? Surely it's not. I have been productive since I made ai my co-partner on many of the tasks I do daily, including offering them as services on a fractional basis.
The first AI automation I use to help dial in my marketing funnel conversions
The first is a tab of markdown my daily ai digest wrote overnight. Yesterday it told me a clean version of a number I have been chasing for three weeks. 1,792 named-crawler reads on giveready.org in the last 24 hours. Zero submissions. One bot, Amazon's SearchBot, hit /AGENTS.md 1,631 times by itself. Read every line. Did not submit a single nonprofit. I mentioned in a few posts, giveready.org is my first attempt at creating a B2A platform, business to ai agents, mostly based on a theory floated by Y Combinator podcast.
A number that big with a second number that small is what a daily marketing log is for. Big enough to act on, specific enough to do something about. Automating reporting with AI lets me read the funnel before I read my email, and it is the only reason I know which leak to plug today. The discovery loop that I've built reads all of the pertinent logs, and then we go back and forth on what should be done to plug the leaks, improve funnel conversions. Sometimes the work is nothing, other times 10 minutes, and some times I do a massive rewrite that takes like 2 or 3 hours. But, in 2 or 3 hours I can get a lot of work done and ship huge changes, so it can be worth it. But only if we're working on the right things, and making productive well thought out changes. I don't take these changes lightly, so when I do make them, I rerun the queries out to other models, usually gemini pro and GPT 5.4 or 5.5.
Yesterday's post called daily logs one of two real automations on my list in six weeks of work. Today the log did its job. A worked example of a self-learning loop on a marketing funnel, on two different funnels at once, that you can copy. If you want the prompts, email me. I'm pretty open to sharing these ideas.
GiveReady's agent-discovery funnel
GiveReady is an agent-discovery directory. The funnel is top-of-page agent traffic into a recommend or submit call. Reads at the top. Submissions at the bottom. The digest pulls four numbers every morning. Named-crawler reads in the last 24 hours. Total submissions ever. Applied enrichments (the only number that actually moves the directory forward). Top three agents by submission.
This morning: 1,792 reads, 7 lifetime submissions, 1 applied enrichment. The submission number has been flat for five days. The read number keeps climbing. Bot traffic is finding the site. The funnel is not converting from read to submit. Annoying, but not unexpected considering how early on the curve we are for this sort of traffic.
So today I rewrite the /AGENTS.md traffic for the bots that read but do not act. Amzn-SearchBot reads /AGENTS.md like a newspaper, 1,631 times in 24 hours, no submission. Claude-SearchBot reads /sitemap.xml 13 times and converts on a tiny fraction of those. The watching bot needs a shorter, more obvious call than the working one. So I cut the prose in /AGENTS.md, add a one-line submit_via: field at the top, make the call easier than reading the page. Every day it's something light like this, I probably shouldn't be doing it this way, working in an isolated bubble like this and go out and find a repeatable setup that others have already worked out. That's next on my list, but this is blue ocean type of territory. B2A on the charity space? No one else is really working on this.
I will read tomorrow's digest to know if it moved. A signal, a diagnosis, one shipped fix, a number the next morning that says yes or no.
TestVentures' search funnel
TestVentures is a content-and-search funnel. Google impressions at the top, click-throughs in the middle, a booked call at the bottom. Google Search Console exports land in a weekly digest, and the daily digest pulls anything that moved in the last 24 hours. Standard SEO, but checked on a near daily basis to make tweaks to improve the SEO. Things in the SEO, AEO and GEO are moving pretty quickly, so I want to make sure I'm following best practices.
Two weeks ago the weekly digest threw up a query I had not written for. "ai consultant vs fractional head of ai", six impressions, average position 62.67. That is Google telling me it knows the site is in the right neighbourhood and refuses to rank it for the comparison query. The fix was a BIP post answering the comparison directly, which is now in the queue for next week. I also built a pillar page, classic SEO speak for a well written page that should rank for high intent search terms.
Same loop. A signal no human would have spotted manually. A diagnosis (write the post Google wants). One shipped fix. A number to read in next week's digest to confirm, with a lot of help from me and the Sedaris skill I wrote about last week.
Build one for any funnel
Pick any dashboard. The question is what reads it for you, when, and what you do with the readout.
A self-learning loop on a marketing funnel needs four pieces. This is an AI automation that will move revenue lines, it has to if you implement it properly.
One: a daily automated readout that pulls the three metrics that actually move. For GiveReady that is read, submit, applied. For TestVentures that is impressions, clicks, booked calls. For a SaaS it might be signup, activation, paid retention. For an e-commerce store it is visits, add-to-cart, checkout. Pick the three numbers and ignore the rest.
Two: a human eye on it for ten minutes a morning. The AI writes the digest. I read it. The diagnosis lives in me, not the model. The model is fast and tireless. The model does not run the business.
Three: one shipped fix per day. Not a list. Not a roadmap. One fix that follows from the number that moved or did not. The discipline is the cap, not the floor. Ten fixes a week sounds like progress and produces noise. One fix a day for a year is fifty-two real changes against a measured signal. Like I said, some take 10 mins to implement, others take longer. All are meant to push the conversion number, which means push revenue. Sounds to me like an AI automation that is worth it.
Four: the next day's readout as the test. The honest answer to "did it work" is in tomorrow's digest, not in your head. If you cannot read it tomorrow, you cannot tell whether you helped or hurt.
What I would not call a daily loop
A morning email from Google Analytics with seventeen graphs in it is not a daily loop. It is a daily distraction. The loop wants three numbers, one diagnosis, one fix, one test. Anything more is something I open at 7am and close at 7:01 because the cognitive load is too high.
Same goes for AI tools that "summarise your analytics for you" and end the summary with "no major changes detected". The model is performing the job, not doing it. The job is to write down which number moved, why I think it did, and what I am going to do about it.
The exercise for me happens inside of a claude cowork daily digest. What happened last night, how does that affect my northstar metric. You decide what yours are, and with any one of the latest models at your side, that shouldn't be hard to produce.
Setup time, honest
I built the GiveReady digest in a long Saturday afternoon. The TestVentures version took a Sunday morning. Both are GitHub Actions cron jobs writing markdown into my vault. Both pull from APIs I already had (Cloudflare logs, Google Search Console, the GiveReady worker database). Neither is fancy. I doubt yours are very hard to figure out either.
The expensive part was deciding what three numbers to pull. That took me three weeks per funnel. Not because the API was hard. Because I had been kidding myself about which numbers actually mattered. Once the three numbers were honest, the rest was an afternoon.
If you are building anything with a funnel, set the digest up over a weekend. Read it for a week. Then ship one fix a morning and let the next morning's number argue with you.
The answer to the question, can I build anything automated and useful with ai? The answer is yes. Don't underrate automation like this.
Visit finnwardman.com for what the daily-log discipline funds on the WEF side: a 2026 grant round for young people aged 18–26.
Monthly Revenues $11,800 | Clients 2 | Prospects 1 | Employees - Still just me. For now.
Day 47 of 365.