Twilight of the Ivory Tower

Original illustration for The Hard Thing.

There is always a name that comes up when a difficult question has travelled around the room without an answer. The architect who understands the old platform. The SAP expert who remembers why the strange process exists. The infrastructure specialist who knows what nobody dares to touch.

Everybody waits for the verdict. It is usually right and rarely challenged.

Stephen King's The Dark Tower has its Crimson King. The corporate version is less dramatic. Its ruler has no supernatural powers, only years of experience, documentation nobody else understands, and a calendar full of meetings.

For years, it was hard to tell how much of that authority came from being right and how much from being impossible to verify. Now something else has learned to climb.

The old price of expertise

Some professions have always had a high entrance fee. Architecture, infrastructure, security, law, finance, and complex SAP processes are obvious examples. To understand a large codebase, you had to learn the technology, read incomplete documentation, trace dependencies, and discover why strange decisions had been made ten years earlier. The work could take weeks.

This gave experts real value and, sometimes, unusual power. They could say that a change was impossible or too risky. The statement might have been correct, but few people in the room could verify it. Knowledge became private territory. The expert lived in the tower, and the rest of the organisation waited at the gate.

The Polish szara eminencja translates as grey eminence. It is not quite the same, but the two can overlap. Both gain influence from knowing what others do not while remaining difficult to challenge.

The model has entered the archive

Models such as GPT Sol or Claude Sonnet already hold more theoretical knowledge than any individual specialist can remember. More importantly, an agent can combine it with the materials of a real company.

Give it a large repository and it can map modules, follow calls, inspect history, and suggest where a change belongs. Give it infrastructure definitions and logs, and it can connect a failure with configuration and code. Give it process documentation and SAP examples, and it can start reconstructing how the business works.

Models still miss context, make confident mistakes, and accept false assumptions. Human experts also forget, guess, protect old decisions, and overlook details. The useful comparison is not between a perfect expert and an unreliable machine. It is between the expert alone and the expert working with a system that can read almost everything and propose several explanations.

The second version wins.

Weeks become hours

A codebase analysis that once required several days can now produce a useful first map in minutes. Reverse engineering a poorly documented integration may take an afternoon instead of a week. An agent can compare implementations, generate tests, and document what it found while the expert handles the difficult decisions.

This changes which problems are worth solving. A company could live with an awkward system because understanding it cost more than improving it. Today, many such problems can again be solved with money. This time the money buys tokens to inspect more evidence, try several approaches, and ask one model to challenge another.

The cost of a second opinion is also approaching zero. A product manager can test a technical claim. An engineer can explore a business process. A new employee can prepare before meeting the old guard.

The tower loses power when everybody can see through its windows.

Insider knowledge does not stay scarce

In my previous post about competitive advantage in AI, I argued that proprietary data and process knowledge are among the few defensible advantages a company has. Company knowledge can remain extremely valuable while the person who remembers it loses their monopoly over it.

Feed an agent good examples, decisions, exceptions, and outcomes, and insider knowledge becomes reusable. A rule explained once can be applied in hundreds of cases. The value moves from the person who guards the knowledge to the organisation that captures it.

Not all experience can be written down. Experts notice weak signals and remember failures that never reached the documentation. Tacit knowledge matters, but its half-life as a private advantage is getting shorter. Every correction and decision teaches the system more about local reality.

For some experts this will feel like a loss. For the organisation it is healthy. No company should depend on one person remembering why a critical process works the way it does.

AI can question the process, not only the code

Process mining can reconstruct what happens from event data: which path an order takes, where an invoice waits, and how often employees use a workaround. An LLM can connect this evidence with policies, system behaviour, and business goals.

The interesting question is no longer only, How can we automate this step? It is, Why does this step exist at all?

Some steps exist because of law, an old contract, safety requirements, or a system that cannot yet be replaced. AI does not make these constraints disappear. It makes them easier to examine.

Agents can separate an actual legal requirement from a sentence repeated for years, find where an exception became the default, and ask uncomfortable questions without caring who designed the process.

Permissions, fragmented data, regulation, and decades of system history will slow adoption. They will not stop it.

Starting is easier than the transformation programme suggests

There is a beautiful part to this change, although laggards may not enjoy hearing it: working effectively with agents is not hard to start.

You do not need a new department or a two-year roadmap. Pick one contained problem. Give the agent the relevant repository, examples, and a clear definition of success. Ask it to show its assumptions and verify the result. Then review it as you would review a capable but new colleague.

The first attempt may be clumsy. The second will show which context was missing. Soon, the hard part stops being the prompt. It becomes access to data, good examples, permissions, and the willingness to change how people work.

What should experts do?

If you live in an ivory tower, do not spend your energy defending the stairs.

Accept being number two

You may no longer be the fastest person in the room at recalling facts or reading code. Accept it. Being number two to a system trained on a large part of recorded human knowledge is still a good position.

The ranking is less important than the result. The model can produce options. Somebody still has to judge them and take responsibility. Being second best at recall is fine if you are first at judgement and ownership.

Scale your expertise

Turn your experience into instructions, examples, checks, and reusable workflows. Let agents do the first analysis, test assumptions, and document the outcome.

One infrastructure expert can review more systems. One architect can compare more options. One SAP specialist can investigate more process variants. The expert becomes a force multiplier rather than a queue.

Leave the tower and become more versatile

In 2023, I wrote that startup founders eventually need to move from being generalists to focused leaders. I still agree that focus matters. What has changed is the cost of crossing into another field.

Agents make it easier to work outside your original speciality. An architect can explore product economics. A product leader can inspect an implementation. A process expert can prototype an automation. They can ask better questions and carry an idea further before needing help.

This makes versatility more valuable. The future belongs less to the person who owns one narrow area and more to the person who connects technology, process, customer value, risk, and business reality. Deep expertise remains useful, but it needs a wider surface.

Own the outcome

AI is already good at producing advice. Experts should move closer to the outcome: decide what to optimise, verify the evidence, choose among trade-offs, and check whether the change worked. Accountability is harder to automate than analysis.

Open the door

The era of expertise is not over. The era of using expertise as a protected monopoly is.

Experts still have an important role, just not the old one. They can verify, connect, teach, question, and take responsibility. They can use AI to spread scarce knowledge instead of guarding it.

The ivory tower does not need to be demolished. Open the door, build the stairs, and come down. The interesting work is no longer at the top.