Tuomas Piippo
CTO, Founding Partner
Rework
When AI has done substantial intellectual work, that work should leave behind an understanding that people can examine, acquire and continue.
There’s a specific kind of pleasure in presenting something you’ve really spent time on and put your mind into. For me, it’s about the only kind of public speaking I can say I enjoy.
Think of it. You’ve spent weeks trying to understand something. You’ve followed leads that went nowhere, changed your mind about a few things, and gradually begun to see how the pieces fit together. By the time you’re presenting your findings, you probably understand quite a lot more than you’ve managed to put on the slides.
The hard work is done, and now the task left is getting the people in the room to see what you’ve come to see, as well as you can.
You’re looking forward to the questions, too. Someone asks something you hadn’t quite considered, you take a few mental steps back, and you can begin to see where it connects to the rest of what you’ve learned.
It’s not so much about being right or having all the answers, I think, but about having done enough work to earn a familiarity with the problem. When you’ve done that, even an “I don’t know” doesn’t feel like a defeat. You can say it with confidence.
The other kind of presentation
Now imagine presenting something you’ve spent considerably less time thinking about.
You’ve given Copilot a collection of interview transcripts, spreadsheets and process descriptions. It’s done a rather impressive job of analyzing them, complete with references to the original material. You’ve gone through the findings, corrected a few things, and asked it to investigate some further.
Eventually, you’ve got a rather professional-looking presentation. You take it to the client and start walking them through the findings. The story seems to hang together, you’re familiar enough with the material to explain the main points, and everything is going rather well.
And then comes the first difficult question.
You recognize the finding, of course. You remember going through it, perhaps even correcting some of the wording. But as you try to follow the question back through the reasoning, there’s surprisingly little to hold on to. You reach for the understanding you expected to be there, and come back rather empty-handed.
Suddenly, you’re not really thinking about the client’s problem anymore. You’re thinking about how to get through the question without revealing how little you understand about the conclusion you’re presenting. You start looking for something plausible to say, hoping you’ll find something solid along the way.
Even an “I don’t know” feels quite different now. In the first situation, you knew where the limits of your understanding were. Here, you’re not even sure whether the question has been properly investigated. Perhaps Copilot had a perfectly good reason for the conclusion. Perhaps it simply sounded like the sort of thing that belonged on the slide.
You don’t know which, and there’s one answer you really don’t want to give: “I don’t know. We did this with AI.”
I don’t think it’s because using AI is something to be ashamed of per se. The client may well assume you used it, and often, in a way, you’d be stupid not to. The problem is that the client is not asking who made the presentation, but rather why they should believe what’s in it.
Why Is AI-Generated Analysis Hard to Defend?
That doesn’t mean we should stop delegating analysis to AI. Quite the opposite. The interesting question is how we could delegate much more of the work without ending up in this situation.
A person who understands a problem well is not simply someone who knows many things about it. They have also acquired some understanding of the boundaries of what they know.
With the AI-generated presentation, you may have more answers at your disposal, but far less understanding of their limits.
You don’t necessarily know whether a difficult question challenges something well established or something the AI inferred rather casually.
That’s what I mean by defensible understanding: being able to trace any AI-generated conclusion back to its evidence, the alternatives that were considered, and what remains uncertain.
Identical Deliverables, Different Accomplishments
In our Copilot example, the consultant has a perfectly good analysis, complete with references to the original material. But when a difficult question comes up, asking Copilot for an explanation may mean asking it to work out the reasoning again. The answer could be excellent. What the consultant cannot be sure of is whether it’s the reasoning that led to the original findings.
What an AI system should, in my opinion, be able to do is maintain an evolving understanding of the case as it investigates it. It should retain not just findings and their supporting sources, but the relationships between them: how statements were interpreted, what conclusions depend on which evidence, which alternative explanations were considered, and what remains uncertain.
The consultant could then explore the case with the AI, following conclusions back to their grounds, examining alternatives and asking questions as their own understanding develops. Rather than reconstructing the investigation from its finished documents, they could engage with the investigation itself.
And that difficult question they’d have dreaded in the Copilot scenario? Perhaps now they could say, “That’s a great question. Let’s ask the AI to show us what that conclusion depends on.”
Here, the AI wouldn’t be conducting a new analysis and hoping its answer happened to agree with what went into Friday’s slides. It would be exploring the same investigation that produced those findings. And if something had become clearer or changed since then, it could show what had changed, and why.
Two systems might produce identical analyses, but leave the consultant with very different opportunities to understand the work behind them.
You Can’t Ask the Same Consultant Twice
A consultant who has spent weeks investigating a problem isn’t quite the same consultant who began. They can answer questions they couldn’t have answered before, and they know more about the limits of what they’ve established.
Why shouldn’t the same be true of the AI that did the investigation?
That’s what I mean by subjective AI. The thinking that went into Friday’s slides should have left the AI changed. It shouldn’t emerge from the engagement with nothing more than a finished presentation file. It should have acquired an understanding of the client’s situation through doing the work, and retained that understanding afterwards.
This is where I find neurosymbolic AI particularly interesting. The idea is to combine the interpretive abilities of language models with more explicit computational structures. The language model can work through interviews, documents and data, while the structures keep track of what was established, how the findings depend on the evidence, and what remains uncertain.
It’s not necessarily the language model itself that changes, but the AI system built around it. When the consultant asks about a finding, that system can continue the same investigation rather than reconstructing an explanation from the original documents.
The emphasis then shifts from recreating explanations to helping the consultant follow the line of inquiry: what was discovered, how the conclusions were reached, which alternatives were considered, and what remains uncertain. The AI can help the consultant, and ultimately the client, develop enough understanding to question the findings and recognize their limits.
The interesting test comes after the slides are finished, when someone asks a question nobody anticipated.
I wouldn’t want the AI’s contribution to be something we feel we have to hide. Its work should be attributed openly, just as we’d acknowledge the work of another team member.
And when we turn to the AI with a further question, it shouldn’t have to investigate the case all over again. We should be able to carry on with the investigation it had already done.