Tuomas Piippo
CTO, Founding Partner
Rework
Anyone who has ever had, or been, a kid in the back seat will recognize the question:
Are we there yet?
I’d like to think I was a reasonable kid. I wouldn’t start asking straight away after we left home, but would wait until we had been on the road long enough for the question to feel justified. By then, we had clearly made progress, but from the back seat it was surprisingly hard to tell how much of the journey was left. The questions could start. Are we there yet? What about now?
What does this have to do with AI and consulting productivity?
I think consulting has now been on this road long enough for the question to feel justified, too: have we actually arrived at the kind of productivity gains we have been told AI would bring?
In one sense, the answer is clearly yes. Generally available AI tools already let us do many parts of our jobs much faster than before. Research, drafting, analysis and presentation work are obvious examples.
But as a whole, I don’t know. Many individual tasks are already much faster, but it is less clear what that has added up to at the level of the job itself. Didn’t we only just pass the gas station that was supposed to be somewhere around the midpoint? What will it look like when we are there? What would actually have had to change along the way for consulting productivity as a whole to reach something like 5× today’s level?
I increasingly think the answer has less to do with making individual tasks faster, and more to do with how much of a consulting engagement still has to be continuously carried by a person.
What if “AI does the routine work” is the wrong starting point?
This is a question we have been forced to think about while building Rethink, our AI product for management consulting.
Our initial instinct was fairly conventional: find recurring work with limited need for judgment, automate it, and leave the difficult parts to the consultant.
AI does the routine work. Humans do the hard thinking.
There is nothing obviously wrong with that division. But I am no longer convinced it gets you very far towards 5×.
AI would have to do more than help with the work while the consultant is there. It would have to carry enough of the engagement that useful work can continue while the consultant’s attention is somewhere else.
In our case, that would mean Rethink becoming the best-informed member of the team on where the inquiry stands and where it is going. It should know what has changed, what remains unresolved, what the evidence now suggests, and what is worth asking next. And if it can also do the routine work of asking those questions, all the better.
That is a very different ambition from automating routine work. The goal is not for AI to make the final judgments or replace the consultant, but for the machine to take responsibility for maintaining the state of the inquiry.
Software development offers an interesting comparison
Something like this shift is already visible in software development.
Early AI coding tools mostly made developers faster at things they were already doing. Complete this function. Explain that code. Write a test. Generate some boilerplate.
Certainly useful, and even more so as AI has learned to write more complex and difficult code over time. But the human still has to be there for the work to continue.
The more interesting shift is that an agent can now be given a larger piece of work and keep going. It can inspect the codebase, make changes, run tests, see what failed, revise the implementation and continue, while the developer’s attention is somewhere else.
Software development does, of course, give AI some important advantages. The thing being worked on is machine-readable. The agent can act on it directly. Tests tell it quite quickly whether many of its actions worked, and mistakes can often be reversed.
That means the agent can stay inside a loop of observe, act, check, correct without requiring a person to hold the loop together.
In many other AI workflows, the person is still doing exactly that. They gather the information, give it to the model, check the output, take the next action somewhere else, observe what happened and come back.
The AI may be doing sophisticated work in the middle, but the person is still keeping the process alive.
What would this look like in consulting?
Imagine a fairly ordinary engagement. There have been a dozen stakeholder interviews, several versions of a process map, a growing list of hypotheses, a data request that is still only partly answered, and a steering meeting next week.
The familiar use of AI is to make each piece faster. Summarize the interviews. Draft the process map. Analyze the data. Prepare the slides.
In the more ambitious version, as new interviews come in, the system would update its understanding of how the process actually works. It would notice that three people describe the same handover differently, that a new data point weakens one of the working hypotheses, and that an apparently minor exception may explain a problem seen elsewhere. It would keep track of what is still unknown and which questions are now worth asking.
When the latest dataset arrives, the work would not begin with somebody remembering why it was requested. The system would already know which questions the data was supposed to answer, run the relevant analysis, connect the findings back to the inquiry and identify what changed as a result.
And before the steering meeting, the consultant would not ask AI to “prepare a deck.” The more useful starting point would be that the system already knows what has happened since the previous meeting, which conclusions have become stronger or weaker, where the client still disagrees, and which decisions now need human attention.
The consultant would still decide what they believe. They would still challenge the client, read the room, make trade-offs and take responsibility for the recommendation.
But they would be returning to an inquiry that had remained alive and progressed while they were doing something else. That would allow the consultant’s attention to move between more engagements without each one effectively pausing in their absence.
There is an important condition to all of this, though. If the consultant has to go back through every interview, dataset and intermediate conclusion to check what the system has done, not much has been gained. For this kind of delegation to work, it has to be possible to see where conclusions came from, what changed them and where uncertainty remains, without doing the work all over again.
That, to me, is a much more plausible picture of 5× productivity than making every individual consulting task five times faster.
We may not be quite there yet, but we’re getting closer.