The first thing worth automating is not the work you can see. It is the sorting that happens before the work, when something arrives and you decide whether it deserves you at all. Almost every AI tool sold to marketers does the opposite, and that is why so many of them add to the pile instead of clearing it.
There is no product marketing team where I work. There is me, and over the past year, a growing number of agents. The two that gave back the most time make nothing at all. No deck, no post, no draft, no email. Their whole job is working out what reaches me. The ones that build things took several times longer to get right and returned less.
This is about which of those to build first, and why the order is not obvious. The short version: automating the deciding removes work permanently, and automating the doing gives you more work to check.
Two kinds of agent
The distinction carries the rest of this, so both sides get a definition.
A sorting agent takes a stream of things arriving at you and decides which ones matter, to whom, and how much. Its output is a judgment about attention.
A producing agent takes an instruction and returns an artifact. A deck, a post, an article, a reply.
Nearly everything being demonstrated to marketers right now is the second kind, because an artifact appears on screen and a judgment does not.
The one I would build first
If I were starting over, the first thing I would build is the agent that watches competitors.
It runs on a schedule, every day, across every comparator. It reads their company pages on LinkedIn and it reads their websites, and for each one it answers four questions: what is new, what has changed, how that change lands on my side product by product, and what the change actually means. The impact carries a severity rating. Where a competitor announced something through the press, I get a link to the announcement. Where they wrote it up themselves, I get a link to their own page.
Then the decision that does most of the work: it ignores their blogs.
Competitors publish constantly and almost none of it is a change in what the company is doing. Ordinary marketing output is the loudest thing they emit and the least informative, so the agent is instructed not to look at it. What it looks for instead is corporate. The company, the product, the vision, the positioning. Things that move rather than things that get posted.
Most competitive intelligence tooling does the reverse, and the reason is not mysterious. Content is the easiest thing on the internet to collect, and volume is easy to mistake for value. What comes back is a feed nobody opens by week three.
The severity rating is the second decision, and it is the one that took the longest to get useful. The agent does not report a change. It reports my exposure to that change, product by product. A competitor repositioning can matter enormously against one product in the portfolio and be irrelevant against another, and that difference is the entire thing I need. A report that tells me what happened has handed the work back. A report that tells me where it hurts has done the work.
The second one, and what it took
The other agent that makes nothing reads the day's email and works out which messages are mine to answer. It sorts twice, first by which product a message concerns and then by what is being asked for, and it hands back a summary with the action attached. I have written elsewhere about how its approval design works, so what matters here is the sequencing. It was the first thing I automated, and it is still the one I would not give up.
Being honest about the cost: it was poor early on. Recognising the type of request came easily, since the types repeat. Attaching a thread to the correct product did not, and the misses were frequent enough to be a genuine annoyance in the first weeks. Two months of retraining took most of them out. Not all of them, which is why random messages still get opened by hand.
So sorting is not free. It is cheaper, which is a different claim, and the comparison is the point.
Why the sorting agents were cheaper
Three reasons, and they are worth separating because only the third one explains the compounding.
A sorting agent chooses from a closed set. There are six things people ask me for by email, and the agent picks one of them. A producing agent chooses from an unbounded space: every sentence, every slide order, every image. Fewer available answers means fewer ways to be wrong, which means fewer rounds of correction.
A sorting failure is cheap and visible. If the inbox agent routes something wrongly, the cost is that I open one message I did not need to open, and I find out immediately. If a producing agent gets a deck wrong, the cost lands in front of a customer, and I find out from a rep afterwards.
And the returns run in opposite directions. A sorting agent removes work from your day and it stays removed. A producing agent adds output, and output needs checking, so the amount of work requiring your judgment goes up rather than down. Automate production and you have not bought time, you have bought volume, and volume arrives with a review queue attached.
Go one level up, because that third point is about where the scarce thing sits. Herbert Simon named it in 1971, writing about how to design organisations for a world that had too much information in it. What information consumes, he said, is rather obvious: it consumes the attention of its recipients, so a wealth of information creates a poverty of attention. Computers, Communications, and the Public Interest, Johns Hopkins Press, 1971 He meant reports and memos rather than agents -- but an agent that makes things is an information source you installed on purpose, and it spends exactly the same currency.
For one person covering a portfolio, the ability to produce was never the constraint. Knowing what is worth producing was. Point an agent at production and you have automated the part that was not the bottleneck.
What the producing agents cost me
I still run them, so this is not an argument against building them. It is an argument about order.
The agent that builds custom pitch decks works because the job it does is narrow: reordering slides that already exist to fit the story a particular buyer needs. That narrowness is what made it tractable.
The agent that handles social was a different experience entirely. Its brief was to reach several distinct audiences at once through one post, which turned out to be a genuinely hard instruction for it to hold. Agreeing the look of an asset, and the rules governing it, consumed round after round, while the mechanics of publishing were comparatively trivial.
The content system was the most expensive of the lot. It produced writing in the wrong voice for months, and the fix was not a prompt. It needed the process a human writer follows to be written out step by step before the agent could follow it, and even after that its own quality score was measuring the wrong thing.
Every one of those is a producing agent. Every one of them cost multiples of what the sorting agents cost. That is five systems, all mine, which is enough to be a pattern in my own work and not enough to be a finding about anybody else's. I do not know whether it holds for a team with more hands than judgment to spare, where the arithmetic might genuinely run the other way.
Four questions before you build anything
Run these against whatever you are about to automate.
- Does this agent hand me a judgment or an artifact? If it is an artifact, ask what decision has to be made before that artifact is worth making, and check whether anything is automating that decision instead.
- What is the closed set it chooses from? Write the list. If you cannot write it, the output space is unbounded and you should expect months rather than weeks.
- What does a failure cost, and when do I learn about it? Cheap and immediate is a good place to start. Expensive and delayed needs a gate before it needs a build.
- After this runs for a month, is there less work in my day or more? An agent producing five things a week where you produced one has increased your review load. That can still be the right trade, but make it deliberately.
The one that catches people is the fourth. It is easy to measure output and hard to notice that your own queue got longer.
Where to start on Monday
Take whatever arrives at you in a stream. Email, requests, competitor noise, inbound questions, a shared channel. Pick the stream you triage most often and least willingly, and write down the closed set of things it contains. That list is most of the specification.
Then build the thing that reads the stream and tells you which items are yours. Give it nothing else to do. Do not let it draft replies, and do not let it act.
You will get less to show for it than a demo would give you. What you get instead is a shorter day, and that is the thing you were short of.
