Vision
Why Most AI Tools Fail to Capture How You Think
The knowledge trap
Most AI tools compete on how much they know. More data, more documents, longer context.
Knowledge is not what separates one expert from another. Two fractional finance directors can read the same management accounts. One tells the founder to cut a product line this quarter. The other says to give it six months. Same facts, opposite advice, and both defensible.
The difference lies in how each of them decides. That is what clients pay for, and it is what most tools miss.
What "how you think" means in practice
It is less mysterious than it sounds. It comes down to a few things you do repeatedly.
Trade-offs. When speed and quality conflict, which gives way? When a client relationship and a point of principle conflict?
Thresholds. How sure do you need to be before you act? Some people move at 60% confidence and correct as they go. Others wait for 80%.
Conditions. This is the one that matters most. Almost nobody has a flat rule. You move fast when a decision is reversible and slow down when trust is at stake. You are generous on scope with a new client and strict with one who has taken advantage before.
Experience. The cases that taught you those rules: the client you kept a year too long, the hire you rushed.
Three ways tools get this wrong
They average
A general model produces the most likely answer across everything it has read. It is the average of many people's writing, which makes it nobody's judgement in particular. Fine for general questions, poor for standing in for a specific person.
They ask you to describe yourself
Many tools build a profile from a questionnaire. The trouble is that people are unreliable narrators of their own judgement. Ask an agency founder whether they would fire a difficult client and they will say "only as a last resort". Put a real case in front of them, with a client who has been rude to a junior twice, and they say "end it this week". The case gets the truth. The questionnaire gets the self-image.
They reduce you to scores
Another approach rates you on a set of traits: risk tolerance, decisiveness, directness. We went some way down this road ourselves and abandoned it. A single risk score cannot hold "fast when reversible, slow when trust is at stake". It averages the two into "moderate", which is wrong in both situations.
What works better: decisions as evidence
The approach we settled on is to collect decisions and keep the conditions.
On Imora you answer real scenarios in your own words. Each answer is stored as evidence. Imora distils the evidence into principles, each with the conditions it applies under and the exceptions, and links them into your reasoning graph. Every principle points back to the words it came from, so you can check it.
When someone asks your twin a question, it reasons from the principles and evidence that fit that kind of decision. The reply opens with the twin's read, and "why this answer" shows what it used and how confident it could be. The writing is done by Anthropic's Claude. The graph, and how it retrieves your decisions, is the part that is Imora's own. Introducing the reasoning graph explains it fully.
Staying honest about the limits
No model of a person is complete, so what a twin does at the edges matters as much as what it does in the middle.
When you read a reply and it is wrong, mark it "Not what I'd say" and give your version. That correction outranks everything else.
Only you teach your twin. Nothing a visitor says changes how it thinks.
When a shared twin has no recorded view on a question, it says so. It offers an approach, clearly framed as an approach and not your position, and the question lands in your inbox. Over time that inbox becomes a list of exactly the judgement your clients and team wanted from you and could not get.
A useful exercise, whether or not you use Imora
Write down five decisions you made in the last month where a reasonable peer would have gone the other way. For each, note what you chose, why, and what would have changed your mind.
That page tells a new hire more about how you think than your CV, your website and your slide decks combined. It is also roughly what Imora's quick start asks of you: five scenarios, about 10 minutes. The product page shows what happens next, and what makes a good AI twin has tests for judging the result.