Treat It as an All-Knowing Blind Man — Seven Opening Moves for AI Agents

Maybe you think your work environment simply does not apply, or that you are already using AI as far as it goes. What follows are some reference steps for widening the situations you use it in.

They are all very basic, but that is the point: most people are stuck not on advanced technique, but on the first two steps never having been done.

Five things first.

⚠️ Before you start

1. You are responsible for the consequences and the output of the AI.
2. Back up your data properly.
3. On using AI at work and on handling sensitive data, look up the relevant rules and tools yourself, for example local models¹, guardrails, compliance gates, sandboxes, hooks² and so on. Your company very likely has a written AI policy; read it first, or ask the relevant department.
4. Do not let this feel tedious or difficult: you can start by using AI to raise productivity within a safe range, and you will naturally find the time and the interest to look into these unfamiliar-sounding things.
5. Once more: you are responsible for the consequences and the output of the AI.

The seven

#

1. Digitise the material you need. Plain text, slides, email, photos, anything works; as long as the information is digital, the AI can read it. Even material you consider unimportant helps a lot; things we take for granted are very meaningful background for it. If the data is sensitive, first confirm your account is set not to share, and that this complies with company rules.

2. Put the material in the agent’s working directory. Following on from the above, for similar tasks put all the material into the folder the agent is running in, and tell it in the conversation: “If you need anything, ask me, or interview me.”

3. Don’t draw boundaries; try asking the AI to do anything. Imagine that, apart from typing to the agent, you cannot use the keyboard or mouse at all. Even tasks you consider simple and quick, let the agent try, and watch it as it goes.

4. Imagine you are about to leave the job. How would your successor take over this work? Give the agent that same information; it will not waste the effort.

5. Let it “see” first. If your data is scattered, say across cloud drives, calendars, a CRM, databases and so on, and you want the agent to read or operate several systems at once, ask it: can these systems be reached over SSH or with a token? Is there an API? Can the data be exported?

   At the start, treat it as an all-knowing blind man, and ask it how these things could be made visible and reachable to it. If you have no idea what any of this means, copy this whole paragraph and paste it to the agent.

6. Get the snowball rolling. All of the above snowballs; once you start trying, it grows quickly, and before long the sense of achievement and the dopamine that come with higher productivity will push you to keep refining the agent’s workflow.

7. Practise somewhere harmless first. If you are afraid of using AI in the real work environment at first, of it deleting your data, say, or sending a wrong email, that worry is reasonable: in unfamiliar situations AI really does not run one hundred percent the way we intend. Find an area you are interested in where mistakes do not matter and try there. I did side projects after work, which is what made me fairly familiar with where current AI’s limits are.

The fifth is the one that matters

#

Most people can manage the first four; the fifth is the watershed.

I think “an all-knowing blind man” is the most accurate description of agents at this stage. It understands everything, but by default it sees nothing: not your ticketing system, not your calendar, not a directory that only exists on one machine.

Most people are disappointed by AI because they evaluated it while it was still blind, pasted a question in, got a generic answer, and concluded the thing is not much use.

The problem is usually not that the agent is not smart enough, but that you have not yet given it eyes. Giving it eyes means letting it connect, read an API, or obtain exported data. That is usually not a prompting problem but an engineering one, which is why the real threshold tends to sit on the IT side.

The wall after the opening moves

#

To be honest, following these seven, most people hit the same wall within a few months: you know there is something that could be automated, but you are not sure at which level to cut.

Abstract too shallow and the agent cannot help; abstract too deep and you spend three months building a framework nobody uses. This judgement cannot currently be outsourced to AI; it is one of the few parts of the whole thing that still belongs entirely to humans.

Further reading: before starting, it is also worth looking at the bill once. My Abilities Are Shrinking is about the three costs I still have not solved →

If you have walked through these seven and are stuck in front of that wall, that is exactly what I do.

The work behind that wall is something I can take on: scope, pricing, how to start →