Your AI agrees with you? That's a bad sign

In late September, on the sidelines of the UN General Assembly, Rwanda's ICT minister summed up the next step: "We need to move from small-scale pilots to solutions that can scale and actually respond to the needs of our citizens." Scaling takes skills — one of the pillars of the national AI policy.

Inside an organisation, that shift looks like this: AI stops being an experiment on one curious person's laptop and becomes part of everyday work. The skill most often missing, then, is knowing how to set up your tool so that it helps you think, instead of thinking for you.

An AI that applauds

From time to time I suggest ideas to clients on projects they have already started — for instance, launching a service offer for individual customers, and the marketing strategy to grow it. I know the subject, and most of those ideas hold up. But a handful of good ideas does not make a complete strategy.

When I run those ideas past an AI that hasn't been set up, the answer is almost always the same. It congratulates me, approves the whole thing and offers to move straight to production. No question about what is missing. No suggestion to simplify the method, or to check that we are not launching something far bigger than the need.

It feels good. It is also exactly the moment you stop checking.

It knows a lot — but nothing about you

An AI you have just opened knows an enormous amount. It knows nothing about you: your field, your standards, how you make decisions. So its answers are generic, and often too polite.

The temptation is to personalise it until it sounds like you. That is useful, on one condition: it must not end up agreeing with you by default. An AI that always says you are right is not an assistant, it is a mirror. It dulls exactly what we expect from a professional: critical thinking and the ability to analyse.

What you want is the opposite: an AI adapted to you, that helps you check what it produces and makes you defend your ideas — or drop the weak ones.

Three settings for a demanding AI

1. Write your instructions once

The main AI assistants offer a personalisation space in the account settings: instructions that apply to every conversation. It is the simplest setting and the most underused. Here is a short version of the one I use:

You are my thinking partner. Your job is to find the truth and tell it to me plainly, while staying within the scope of what I ask.

Before I decide: never agree just to be pleasant. If I'm wrong, say so. Point out weaknesses and blind spots in a few lines. No flattery. If you're unsure, say so, and source factual claims. Make me defend my ideas or drop the bad ones.

Once I've decided: do exactly what I ask, nothing more. If you think something extra would help, suggest it in one line and wait for my go-ahead.

If I seem to be looking for validation rather than the truth, point it out.

Two phases, two behaviours. While you are thinking, the AI pushes back. Once you have decided, it executes without reopening the debate. That avoids both traps: the yes-man tool and the tool that argues about everything.

2. Give it the methods of a profession

AI assistants can now be extended: project spaces with their own documents, skill modules, connections to your tools. Chosen well, these add-ons hold the AI to the recognised methods and standards of a field — designing a training course, analysing a budget, writing a specification.

In practice: create a project, upload your framework, your specification template or your quality checklists, and ask the AI to follow them. It stops improvising a method where a proven one already exists, and makes fewer mistakes on subjects it handles poorly.

3. Turn on memory — and choose what it keeps

Most assistants can remember things from one conversation to the next; the option usually sits in the personalisation settings, and you often have to switch it on. An important nuance: the model itself stays the same — a memory simply keeps notes about you. The better that memory is kept, the better the AI understands your subjects.

That memory raises a question people often forget: what are you willing to let it keep? Rwanda has just launched its national cybersecurity and data protection awareness campaign, and Law No. 058/2021 already governs how personal data is handled. Information about your clients, staff or beneficiaries has no place in the memory of a personal account. A simple routine: once a month, read what it has kept and delete anything about other people.

Stay in charge

A well-configured AI keeps your thinking in shape. Two habits help me.

Separate flagging from producing. At the start of an important topic, I ask for the most rigorous approach and checked sources. After that, by default, the AI flags what bothers it. It only produces when I ask, which spares me documents I never wanted.

Keep track. Over time, the real challenge is knowing where you stand on each subject and in what order to deal with them. For that I use a folder of notes and a dashboard of my priorities, which the AI reads at the start of each working session and helps me keep up to date. The decisions stay mine; the dashboard simply reminds me of the ones I have already made.

With these settings, the AI stops saying "great idea" about ordinary ideas. It asks the question that was missing, and you keep thinking.

In an organisation, the setting becomes a team rule

In Kigali, what I see in companies is still little day-to-day use of AI, or very basic use with no method behind it. It is the same pattern as with much older tools: a calendar managed day by day, an inbox where conversations get lost, files followed from memory. The result is familiar: long days and a constant rush not to fall behind. Solid organisation and a proven method beat any amount of rushing. AI is no exception.

The first setting takes ten minutes, on your own. In an organisation, if each person sets up their tool their own way — or not at all — you get very different standards from one office to the next, and data moving around with no rules.

Three decisions are enough to start: shared instructions, adapted to the team's work; a clear list of what can be given to the AI and what never is; one person who checks that these rules hold over time. It is the natural follow-up to the question we asked in July: who is running AI in your organisation?

---

If you want your teams to use an AI that makes them more demanding rather than lazier, let's take 30 minutes. No pitch — just a conversation.