“Consultancies are done. Partners are leaving because they think their firms are a zero.”
I’ve heard this enough times now from Silicon Valley insiders that I want to address it.
At this point, the received wisdom in tech is that the consulting industry is dead. Consultancies sell intelligence-as-a-service, and AI is getting smarter than them, and cheaper. And if (ex-McKinsey!) Anthropic engineers are to be believed, “95% of computer facing jobs” will be automatable by AI “by 2028”.
I do not agree that consultancy is dead. But it’s worth acknowledging what the Silicon Valley people are right about: the models will be smarter than us.
Consultants often tell me “we will use AI for the drudgery of implementation, and we will be paid for guiding the bots with our expertise and judgment.”
No. You won’t be hired for your intellectual judgment.
AI systems are already strong enough to solve open problems in mathematics. Improvements are not slowing down. Intellectually speaking, AI won’t be overseen by humans; it will be the other way round. Soon, it should probably be a fireable offence to answer any business problem without consulting AI. Long-term, the chances of human expertise adding real intellectual value over models are, as the Bay Area crowd would say, “zero”.
If your selling point was being “the smartest person in the room”, then you’re in for a rough time. You need to change fast. Luckily, you’ve got options.
The complements of intelligence
Consultants are in the problem-solving business, and intelligence alone does not solve most problems. You can’t do a post-merger integration or an operating model redesign just by thinking about them from first principles.
Intelligence needs sidekicks. These are its complements.
Consultants sell at least three powerful complements to intelligence:
- Proprietary Information. There is an information asymmetry at the heart of consulting. A client is seeing a problem for the first time; the consultancies have seen it a thousand times. Today, the consultants package this asymmetry as experienced partners and internal expert networks. In the future, it will be proprietary data for AI.
- The Squishy Stuff. Business problems are just as often social as intellectual. When initiatives fail, often it’s because good ideas were undermined by the people who were meant to make them happen. You don’t solve this with more IQ. You solve it by helping clients with the squishy stuff - listening, story-telling, incentivising, deal-making, reorganising - all uniquely human undertakings.
- Independence (and liability). A firm is not allowed to audit itself. A fund shouldn’t mark its own valuation homework. Even if the superintelligent AI is neutral, someone independent needs to write its prompts. At the soft end, independence means being a neutral arbiter; at the hard end, it means taking legal liability.
Economically speaking, if intelligence is powerful but also commoditised (‘intelligence too cheap to meter’) then the very best thing you can be is its complement. As demand for the commodity explodes, so does demand for your complement - except you can charge a higher margin. It’s already happening - BCG reported 25% YoY growth in AI services. As the value of AI rises, so does the market for its complements; in this case, human implementation. I expect the growth of complements to accelerate.
The one that compounds
I don’t worry that Tracelight’s customers are going to miss out on selling the Squishy Stuff, or independence. Those are business as usual. I worry that they are missing the boat on proprietary information. This is bad because, of the three complements, proprietary information is the one that compounds.
Consultants accumulate knowledge over a career, become partners, and then sell that expertise. This knowledge is private. Clients don’t have it, and neither do OpenAI or Anthropic. The models have read everything ever published about post-merger integrations. But they have no idea what happened on the inside; the consultants do.
As cognition shifts from humans to AI, the most urgent task for every consultancy is to rebuild this learning loop in digital form.
In practice, this means that every interaction is recorded and learned from: how the analyses and narratives were built, what the feedback was, where the project ended up - and even the Holy Grail: what the actual business outcomes were. None of this would have to be written down; agents record the work as a byproduct of doing it. The dataset is stripped of all client specifics - tedious work that agents, unlike humans, will happily complete - and becomes knowledge for all subsequent projects. This knowledge can be used either to update skills, or to be searched, or to be baked into the weights of the consultancy’s own LLMs. An unredacted version of the data could be retained under strict control, to be revisited only by future projects with the same client. The resulting system learns more efficiently, and in a more verifiably compliant way, than the system it replaces, where the learning lived primarily in partners’ heads.
The labs
This is one of the main reasons why customers choose us instead of the Big Labs.
You must assume two things: first, that all your clients are buying the labs’ horizontal tools; second, that the underlying models will continue to get smarter. Under those circumstances, why should a client continue to pay for your intelligence, when they can get it directly?
Your intellectual margin will come from a self-reinforcing loop between models and your proprietary knowledge, as described above. The design of this loop is specific to your organisation. But the economics of the big labs depend upon selling a standardised, low-friction product to everyone. You can hope they build some new primitives that allow you to configure your exact requirements in their product. But that’s hard. And if they’re going to prioritise anyone’s needs, it will be your clients’. A world without consultancies is just fine for them.
What is not fine for them is a world where you can easily switch between models. They lock you into their own LLMs. If the right system turns out to require orchestrating between AI models, including, perhaps, your own post-trained open source models, you are in trouble.
The choice
It’s a true fork in the road. The consultancies that build the right operating system for AI will entrench themselves further with each new project. The ones that fail will quickly be outcompeted.