Proactive AI agents now run our client work: they start before anyone asks, and the biggest gap left between them and a person is judgement. Six months ago that was not really possible.
Everyone keeps saying the models are not the limit, and in September I made the same case about adoption inside other companies. What surprised me is seeing it happen in our own work: the limit is how well you harness the models, and proactive AI agents are what that looks like when it works. What I fear is that there is no clear line between work you can give a person and work you cannot give an agent.
The short version
- Six months ago, automation meant n8n workflows, proactive AI work was not really possible, and approving work meant me clicking in n8n.
- Today each client has a floor of AI agents that reads its Slack, mail and call notes, turns them into tasks and does the work, and a head agent reviews the results before my one-click approval.
- The optimistic read: I think tasks can now run without a person overseeing them, at a high level of accuracy. The pessimistic read: there is no clear line between work for a person and work for an agent.
- Judgement is the biggest gap: asking whether a solution is the right one, rather than whether it works. My bet is that the gap does not last.
Six months ago: n8n workflows and AI that waited to be asked
Six months ago we still sold cold email to clients, and automation meant n8n. I have built 200+ workflows in it, and each one is a fixed chain of steps that runs on a trigger and does exactly what it was drawn to do.
A workflow follows the path drawn for it: every branch is decided in advance by whoever builds it. When a workflow needed a yes, the yes was me, clicking in n8n.
Most people still see AI as reactive: it answers when asked. Six months ago proactive work was not really possible, because the models were too stupid and we did not have the tools to run that kind of work in the background.
Now: a floor of agents per client, and a Decide list
Today our client work runs on Etiennator, an in-house system of AI agents with one isolated floor per client. There are 5 floors, each with its own records, departments and head agent, and the work on them ranges from a creator-marketing client's portal automation to a B2B data client's enrichment pipelines.
A floor takes in that client's Slack, mail and call notes without being asked, turns them into tasks, and has its departments do the work in a sandbox. The floor works out the tasks itself, and that is the proactive work that was not possible six months ago.
The head agent reviews results before I see them, and I approve with one click, from a Decide list or from a morning push.
The same pattern runs our own small SaaS for local-business data. Signups arrive from ChatGPT and Google, an agent desk greets them, follows up and routes bug reports, and fixes I approve go live through a review step.
Six months ago
- Automation meant n8n: 200+ workflows, each a fixed chain of steps
- Proactive AI work was not really possible
- Approving work meant me, clicking in n8n
Today
- 5 client floors of AI agents in Etiennator
- Each floor reads Slack, mail and call notes and turns them into tasks unasked
- A head agent reviews results first, then I approve with one click from a Decide list
What proactive AI agents change, read both ways
When AI only answers, a person is in the loop by definition, because a person asked. Once proactive AI agents start work unprompted, someone has to decide where the person sits.
The optimistic read starts from the usual complaint that AI waits to be asked. Set it up as background work, connected to your mail and every client-facing channel, and it can tell you what you promised you would do and what you need to do.
I think tasks can now run that way without a person overseeing them, at a high level of accuracy, and that was not possible before. We still put a click at the end, so this is what I believe is possible; we have not switched the click off.
The pessimistic read is about what the agents still miss, and the biggest gap is judgement.
“The biggest discrepancy right now is judgement and critical thinking: being able to step out of what you are doing and see if it is indeed the right solution, rather than just checking whether your solution works.”
We have watched that gap in our own work, in a story told in full in make the AI show you the denominator. In August, an AI running a $2.32 data pilot on a client project capped its own sampling at 4% to stay on budget, then recommended we walk away from a paid data source.
Every number it reported was right. It never asked whether 816 results for 20 major cities could be real; when I asked, it went and checked, found the cap it had set itself, and voided its per-email figure.
It had checked that its solution worked. Whether it was the right one was a question only a person asked.
The two reads disagree on one question: whether that gap lasts.
The next 6 months: I do not see the judgement gap lasting
My bet is that judgement stops being the biggest discrepancy between an agent and a person. I do not see it lasting.
I will know it has closed when our head agent asks the 816 question itself, before a result like that reaches my Decide list.
For now, the line in our system sits in two places: the Decide list, where results wait for my click, and the Send button on our SaaS, where the desk writes the follow-up and nothing outbound leaves without me.
What could prove me wrong
Everything in the "Today" column is new, and new systems have not met their hard cases yet. The floors were built on 25 September, and on 9 October agents stopped sitting idle while results waited for my approval.
The accuracy I describe is what I see running them, in one operation with one approver: me. Only two steps have measured numbers so far, from blind tests of the agents' written checklists: task splitting went from 62% to 88.5% correct, and mail routing improved 9.9%. Neither measures how often a floor starts the right task unasked.
If I am still the one asking the 816 question, I was wrong.
Key takeaways
- Proactive AI agents work in the background on your mail, chat and call notes and start tasks before anyone asks, which a chat assistant waiting for a prompt cannot do.
- Approval can survive the move to agents: a head agent reviews results first, and the owner approves each one with a single click from one list.
- Checking that a solution works and asking whether it is the right solution are two different checks, and AI agents are still weak at the second one.
- There is no clear line yet between work for a person and work for an agent; ours, for now, is the Decide list for results and the Send button for anything outbound.
Common questions
What is a proactive AI agent?
An agent set up as background work and connected to your tools, so it reads your mail and client channels and starts the work without being asked. A reactive one answers only when someone asks.
What can AI agents still not do well?
Judgement: stepping out of a task to ask whether it is the right solution, rather than checking whether the solution works. It is the biggest gap today, and I do not see it lasting.
Who approves the work in an AI agent system?
In ours, a head agent reviews results before I see them, and I approve with one click from a Decide list or a morning push. On our own small SaaS, nothing outbound leaves without my click on Send.