Twenty-Three Asks, Five Machines

Read the twenty-three replies as a set and a shape appears that nobody in the room asked to learn. Five were outreach in some form: cold calls, cold messages, prospecting and qualification, the follow-up email. Five were admin, timesheets, meetings, an inbox, accountancy, appointment reminders. Four were documents produced from scratch, client decks, presentations and spreadsheets, proposals, quotes. Three were content, the weekly post, a backlog that never fills, retail copy. Two were planning and data, a project plan with uncertain dates, a database. One was screening a hundred CVs, one was chasing deadlines across teams, one was automating the testing of a chatbot, and one reply was empty. When we sorted the twenty-three, they collapsed into five recurring patterns, and eighteen of them were answered directly by one of the five. The other five were the same five patterns wearing a different hat.

The mismatch between the two lists is not a failure of the room. It is how requests for technology always arrive. People ask for the thing they have heard named, and the name doing the most work this year is agent. As we put it on the twentieth of August, the word currently means a chatbot with a name, a workflow with ambition, and a genuinely autonomous system, three different things with three different price tags. The survey carried the same signal from another direction. Eight of the sixteen respondents use three AI assistants, one uses four, and the average is 2.4 tools per person, yet eleven of the sixteen still describe themselves as daily chat users and only three as people who build. Seven said they had built something with AI already. Tool count is not capability. A person can run three assistants and still be at chat, and most of this room was.

Exactly one of the twenty-three asks was intrinsically an automation job, the request to automate test runs on a chatbot, and even that one came back from us reframed as an evaluation problem: write the rubric by hand first, and make the grader cite the line it graded against. The remaining twenty-two did not need an agent in any of the word's three meanings. They needed a delegation brief and a check. A brief says what the task is, what it may touch, what good looks like and what a wrong answer looks like. A check is the point at which a person confirms the output before it leaves. The sentence we taught the room to separate the two categories is worth keeping: an automation repeats a decision you already made, an agent makes the decision again, every time, inside rules you wrote. Almost nothing on the second list wanted a decision remade. It wanted a decision made once and then carried.

So we built five machines instead of teaching agents. The Pipeline Engine answered the five outreach asks. The Proposal and Quote Factory answered the four document asks. The Morning Desk answered four of the admin asks. The Content Engine answered the three content asks. The Project Pulse answered the two planning and status asks. All five ran on one fictional four-person Dubai consultancy, with data we invented badly on purpose, typos, numbers buried mid-sentence, one arithmetic mistake left in, a client who quietly stops replying, a pipeline file nineteen days stale, because a machine that only works on clean data is a demo. The outreach asks taught us something too. Every one framed lead generation as a volume problem, more prospects, more messages. The fictional pipeline showed the opposite shape: the value was not in the prospects nobody had contacted but in the ones who had gone quiet and nobody had noticed.

The scarce thing in that room was never the tooling. It was the translation from what people said they wanted to what they needed gone, and that translation is a judgment, not a feature. The evidence is in the survey's last question, where thirteen people finished the sentence I wish AI could just. Three of them asked for the machine to tell them when it is wrong, to be transparent, to raise an early warning. That is a verification wish, and it is the most sophisticated request on either list, because it comes from people who have delegated something, watched it come back plausible and wrong, and understood that the check is the product. An organisation that hears its staff ask for agents and buys agents has skipped that step. The request arrives in vocabulary. The need arrives in the diary. Someone has to read the diary.