Why Domain AI Compounds: The Structural Case for Starting Now
A domain AI system is not a tool you buy and depreciate. It is an asset that accumulates: every case it touches, every document it ingests, every correction it absorbs adds to its value. The compounding is structural, and the timing matters more than most decision-makers realise.
Tools depreciate. Knowledge bases compound.
Almost every piece of software a senior practitioner has ever bought followed the same curve: most valuable on the day of purchase, declining from there. The case management system, the citation manager, the laboratory information system — they arrive at peak capability and depreciate as the world moves around them. The accounting category they belong to is correct: depreciating capital expenditure.
A domain AI system runs on the opposite curve. On the first day, it is at its weakest: a thin corpus, a draft axiom document, an integrator who has been on site for three weeks. By month twelve, the corpus has absorbed real casework, the axiom document has grown several pages, and the failure modes that surprised everyone in the first month no longer occur. The system is more useful at twelve months than at one, and more useful at thirty-six than at twelve. The asset accumulates because the substrate is your own accumulated practice, not a vendor's feature roadmap.
Most decision-makers we meet evaluate AI adoption with the wrong mental model. They compare price to feature list as though the system were a piece of equipment. The right comparison is to a cellar, an archive, or a long-running research programme: the decision to start is mostly a decision to begin accumulating. Everything else follows from that.
The three feedback loops
Compounding is not a slogan. It is the visible behaviour of three specific feedback loops running inside a properly designed system.
The first is the corpus loop. Outputs produced by the system — case memos, evaluation reports, literature syntheses — are reviewed, refined by a senior, and, once accepted, fed back into the knowledge base. The next time the system retrieves on a similar question, those reviewed outputs are available as context. The retrieval quality improves not because the underlying model has improved but because the material the model is reasoning over has become more representative of your actual practice.
The second is the axiom loop. Every time the system produces an output that is structurally wrong — not just stylistically off but wrong in a way only a senior would notice — a line is added to the axiom document. The constraint propagates to all subsequent generation. The class of error does not recur. This is why we treat the axiom document as a living artefact rather than a one-off specification: it captures, week by week, the implicit rules of the field as they become visible through their violation.
The third is the process loop, which is less visible than the other two and arguably more important. Over the first six to twelve months, the organisation learns which parts of its workflow should be absorbed by the system, which parts should stay with people, and which sit at a useful boundary between the two. The division of labour stabilises. The stabilisation itself is the asset: a firm that has solved this allocation is now operating differently from one that has not.
Why generic AI does not compound for you
A useful clarification. The compounding described above is not a property of AI in general. It is a property of a domain system built around a private corpus. The standard chatbot does not accumulate for you. Each conversation is, from the firm's point of view, ephemeral. To the extent that aggregated usage improves the underlying model, that improvement is distributed uniformly across every user of that model. It does not become your firm's asset. Your competitors see the same gains on the same day.
A domain system inverts this. The corpus is yours. The axioms are yours. The reviewed outputs that re-enter the corpus are yours. None of this flows to the vendor or to the wider user base. After eighteen months of operation, two firms that adopted the same underlying model but ran private systems on their own corpora are not running comparable assets. They are running two different systems that happen to share a substrate.
This distinction is blurred at month six and decisive at month twenty-four. Most senior practitioners we work with intuit this once it is named, but the intuition does not survive contact with a vendor demo that frames the choice as “chatbot or no chatbot.”
The cost of a late start
The hardest argument to land with senior decision-makers is that the price of waiting is not measured in months. It is measured in compounding that did not happen. Starting two years later does not mean reaching the same place two years later. It means starting from zero against a competitor who, by then, has two years of corpus accumulation, two years of axiom refinement, and two years of process stabilisation. That gap does not close by buying a more recent model. The model is not the asset.
This is structurally similar to how senior expertise itself accumulates in fields like Korean appellate practice or policy-evaluation research: the depth comes from years of cases seen, not from any single year being especially intense. A firm with a fifteen-year archive of reviewed memos is not a firm whose juniors have read more textbooks. It is a different kind of firm. A domain AI system that has run for two years on real casework is, in the same way, a different kind of infrastructure from one spun up last month.
The reasonable instinct of a careful senior practitioner is to wait for the technology to settle. In the case of foundation models, that instinct will be partly vindicated — the models will improve, and they will be cheaper later. In the case of the corpus and the axiom layer, the instinct is precisely inverted. The longer you wait to start accumulating those, the more years of accumulation you have permanently forgone. There is no version of next year's model that retroactively gives you this year's reviewed outputs.
What compounding looks like at month 6, 18, and 36
Specifically, and without abstraction.
At month six, the corpus has reached a critical mass for the firm's core workflow. Outputs begin to sound like the firm rather than like the model. The senior reviewing them stops finding the same class of mistake twice. The integrator, who was on site weekly, is now on site fortnightly. The first internal users beyond the senior have started using the system on their own work.
At month eighteen, the axiom document has roughly tripled in length and has become the firm's most valuable internal document — the only place where the implicit standards of the practice are written down. Review time on system-generated drafts has fallen by something like half, because the obvious errors no longer occur. The system is now assumed in the firm's workflow, in the way a good library or a competent paralegal team is assumed.
At month thirty-six, a junior who joins the firm has access, on their first day, to retrieved drafts that reflect the senior's standards. Their formation is faster because the standard is legible, not because the system replaces their learning. The senior's judgement now scales in a way it could not scale before, because the substrate of the firm has absorbed it.
None of this is accelerated by a better foundation model arriving next year. The model can be swapped under the same corpus and the same axiom document; everything above the model remains yours. This is the structural argument for not waiting, stated without the dressing.