Consulting has always thrived at the edge of transformation. Today, AI is reshaping how knowledge work itself is done. Once again, consulting stands at an inflection point. The difference is that this time, the transformation does not only change clients — it goes straight after the consultants' own work.

When factories replaced workshops, early "efficiency engineers" showed industrialists how to discipline and organize labor. When corporations became too complex to steer by intuition, strategy firms built frameworks to rationalize decisions. When computers arrived, IT consultants wired the modern organization. Each technological shock created new complexity; consultants turned that complexity into billable projects.

I. What Consulting Has Historically Sold

If you look at the last century, you can read consulting as a sequence of "products" packaged around each wave of economic change: scientific management with Taylor, organizational design and budgeting, corporate strategy and portfolio management, business process reengineering, digital transformation.

Underneath the buzzwords, consultants were selling three things:

• Cognitive capacity. Smart people with time to analyze messy situations, pull data, build models, and structure decisions.
• Patterns and benchmarks. Knowledge of what works in other firms and industries; reusable frameworks to compress that experience.
• Legitimacy. A way for leaders to justify hard decisions ("McKinsey recommends…").

In other words: externalized knowledge work, wrapped in process and prestige.

II. The Quiet Rise of In-House Consulting and Embedded Teams

Long before AI, many companies started to internalize part of that capability. Large groups built internal consulting units and "strategy & transformation" teams that operate like mini consulting firms but serve internal clients only.

In the United States, many Fortune 500 companies have long operated in-house "mini-firms" that look and behave like strategy consultancies. A similar wave is now visible in other countries. In France, the Association Française du Conseil Interne (AFCI) brings together internal consulting entities across large organizations; by mid-2024 it counted 34 member teams, up 48% in three years, mostly inside banks, energy groups and public institutions.

These internal teams exist precisely because there has always been a tradeoff:

• External consultancies bring some of the best-trained talent and broad cross-industry pattern recognition, but they arrive cold, stay for a few months, then leave.
• Internal consulting units are embedded and always on — they understand the company deeply and can follow through over years, but they rarely have the same density of experience, training, or external benchmarking as top-tier firms.

At the same time, a new category of tech-enabled services has emerged: companies like Palantir that combine software platforms with forward-deployed engineers embedded for years inside client organizations. They don't just write reports; they ship and operate systems that become part of the client's infrastructure.

Both trends point in the same direction: clients want continuity and ownership, not a new team of strangers every time. The value is moving from episodic advice to permanent capability.

AI takes this logic to its conclusion.

III. AI Labs Are Explicitly Targeting Knowledge Work

If you ask AI lab leaders what comes next, all they talk about is knowledge work: PowerPoints, legal briefs, financial models, codebases, research reports…

Anthropic's CEO Dario Amodei has argued that AI could wipe out roughly 50% of all entry-level white-collar jobs within one to five years, and push unemployment into the 10–20% range if societies do not adapt.

OpenAI has begun to evaluate its models on GDPval, a benchmark designed around real professional tasks across 44 knowledge work occupations — from software developers and lawyers to nurses and mechanical engineers. On this benchmark, GPT-5.2 Thinking now beats or ties industry professionals in about 70.9% of comparisons, and the Pro variant reaches 74.1%, while operating at a tiny fraction of human cost and time. A 35% jump compared to GPT-5 Thinking a few months earlier.

GDPval uses real work products like spreadsheets, contracts, slide decks, support tickets and asks models to produce outputs that domain experts then judge against human work. It is exactly the kind of material junior consultants spend their days on.

If a model can already match or beat a mid-career professional on a substantial fraction of these tasks, it is hard to argue that the bottom of the consulting pyramid is structurally safe long term.

IV. From Smart Tool to Autonomous Project Team

The other important change is autonomy.

Code at big tech companies is now mainly written by AI. At Anthropic, Boris Cherny, Claude Code creator, recently wrote:

"The last month was my first month as an engineer that I didn't open an IDE at all. Opus 4.5 wrote around 200 PRs, every single line. Software engineering is radically changing, and the hardest part even for early adopters and practitioners like us is to continue to re-adjust our expectations. And this is still just the beginning."

Andrej Karpathy summarized the same shift:

"I've never felt this much behind as a programmer."

Not only do models get more efficient but they also get better on long-horizon tasks. The research organization METR has introduced a metric called time horizon: how long a task LLMs can carry to completion with a reasonable success rate. Looking across frontier models since 2019, they find that this time horizon has been doubling roughly every seven months.

In recent evaluations, GPT-5.2 (high) achieved a 50% time horizon of about 6.6 hours — meaning it can autonomously complete multi-hour tasks about half the time without human step-by-step micromanagement.

For consulting, that matters more than another +3 points on some academic benchmark. Most real consulting work is multi-hour task chains: read 200 pages of internal documentation, extract the structure of a process, reconcile it with data from a few systems, draft a first-pass diagnosis and a short note for the client sponsor.

The fact that we now have models that can carry such chains for several hours — and that this horizon is on an exponential curve — means today's capabilities are a floor, not a ceiling.

V. The Cost of Intelligence Is Collapsing

Capability is only half the story. The other half is cost.

Results on ARC-AGI-1, a reasoning benchmark, suggest that in just one year the cost of intelligence has been divided by roughly 400×.

On GDPval, OpenAI reports that GPT-5.2 Thinking and Pro deliver expert-level performance at less than 1% of the cost of human experts, and more than 10× faster on average.

In a business built on leveraged pyramids of junior labor, like consulting, that pressure is impossible to ignore.

VI. Timelines: AGI in Less Than 10 Years

So far, we've looked at what today's models can already do on real work, how long they can run autonomously, and how fast the cost of intelligence is collapsing. The last ingredient is trajectory: not just what AI is today, but where it is likely to be over the next decade.

We are no longer really debating whether AI will reach or exceed human expert level — only when.

Demis Hassabis, CEO of Google DeepMind, who is generally known for being cautious rather than sensational about timelines, puts human-level AGI roughly 5–10 years out. Dario Amodei talks about systems capable of "Nobel-level" scientific research within about two years, with end-to-end software engineering arriving even sooner.

On that kind of timeline, most cognitive tasks that matter for consulting move squarely into scope for AI within the next decade. Assuming model progress will stall is, at this point, a very bad bet.

VII. Why the Traditional Consulting Model Struggles in an AI World

None of this means "consulting disappears." But it does mean that the current operating model is almost perfectly misaligned with how AI creates value.

Traditional firms are built around a simple pattern: a team of juniors arrives with zero context, they interview a small sample of employees, sit in a few meetings, pull some exports from key systems, then feed notes, recordings and slides into their own brains to produce analyses and recommendations. Even when AI enters the picture, it is usually as a tool at the end of a human pipeline: consultants record interviews on their phones, transcribe them, paste chunks into models to draft slides faster.

This helps on the margin, but the bottleneck hasn't moved: a handful of overworked people decide who to talk to, what to capture, what to ignore, and how much context the model gets.

That creates four structural problems:

• A permanent cold start. External consultants always begin from zero. They meet 20–40 people out of hundreds or thousands and try to reconstruct the whole company from that sliver. However smart they are, the underlying data is partial and biased by design.
• Static snapshots in a moving system. The output is a deck, a beautifully formatted but frozen view of an organization that is changing every week. By the time conclusions are socialized, people have moved teams, tools have changed, and new constraints have appeared. Leaders pay for an expensive Polaroid of something that no longer exists.
• No compounding memory. Each project lives and dies on its own. When the team rolls off, so does their understanding. The next engagement, even with the same firm, starts from scratch.
• Humans ration the most important input: context. Models are already good at the kind of knowledge work junior consultants do, but they are starved of information. The very thing AI needs most — exhaustive, detailed context — is exactly what the human operating model cannot economically provide.

In other words, today's consulting industry keeps junior humans at the center of the system precisely when AI is becoming best at junior work. As models improve and costs keep falling, betting that this bottlenecked, sample-based way of working will remain optimal is effectively betting against the trend lines described earlier.

VIII. Letting AI See and Think at Full Scale

If today's models already match or beat human experts on many professional tasks, and are getting better every few months, why are they not already running most of consulting?

Because the bottleneck is no longer intelligence. It is context.

Models do not know how your specific company works in practice. They do not see who really approves what, how many times a ticket bounces between teams, which tasks silently eat entire afternoons, or how processes drift over time. They only see the tiny slice of reality we choose to collect for them.

AI lets you flip this completely. Instead of using models only to polish slides, you can use them to collect and structure data at a scale no human team can match: hundreds of interviews transcribed and analyzed, thousands of pages of documentation synthesized, logs, tickets and workflows parsed into a map of tasks, actors, tools and bottlenecks.

Once this map exists, you can finally let models think at full length on top of it. They can run multi-hour chains of reasoning, explore many scenarios, simulate alternative designs, and search for subtle patterns that no individual brain could hold at once. All of this still costs less than a single traditional project.

Anthropic recently offered a concrete glimpse of this regime. They asked a team of 16 Claude agents to write a Rust-based C compiler from scratch, capable of compiling the Linux kernel. Then, in their own words, they "mostly walked away" while it ran for about two weeks across nearly two thousand Claude Code sessions.

At that point, even one of consulting's oldest services — benchmarking — starts to look different. Instead of vague best practices, you can compare work at the level of tasks, in aggregated and anonymized form. You can tell a client that on a specific task their teams spend several extra hours per week compared to similar organizations. This kind of precise gap analysis becomes possible once AI has done the heavy lifting of mapping work in detail across many companies.

The core shift is simple to state and hard to overstate. The limitation is no longer what models can do in the abstract. It is how much real context about your company you are willing and able to feed them.

Conclusion

Consulting was invented for a world where complexity grew faster than any leader's ability to follow it. For a century, the answer was to send in clever humans, let them sample a small part of reality, and compress it into decks. The method evolved, the buzzwords changed, but the core routine stayed surprisingly old-school.

AI breaks this pattern in two directions at once:

• It lets you see far more than any consulting team ever could. An AI system can listen to every interview, read every procedure, follow every ticket flow, and hold it all as structured data. It runs continuously, so it catches small drifts and rare edge cases humans will always miss.
• It lets models tackle more complex problems. Once you give them a rich map of how work happens, they will soon be able to think for hours on a single question, explore options and refine proposals, still at a tiny fraction of the cost of a traditional project.

Human consultants do not disappear from this picture. They still frame the right questions, read politics and culture, and carry responsibility for decisions and the story that moves people. What changes is the ground they stand on and the source of legitimacy.

As models keep proving that, given enough context, they can produce sharper diagnoses and more robust simulations than any small team, leaders will start treating their output with the same seriousness they once reserved for elite firms. In some areas it will feel less like asking for an opinion and more like querying an oracle that has actually seen everything.

At some point, having your own AI consulting team inside the company will feel as standard as having a finance function or a CRM. A system that knows your processes by heart, tracks how they evolve, flags issues early and suggests concrete fixes. Humans will still make the calls, but on top of an AI layer that sees deeper and farther than any individual ever could.

---