Concept Note · AI as Institutional Memory Instrument · June 2026
Related: SM-007 ·
TN-011
Conventional modernisation discourse treats a legacy system as a technical artefact to be translated, replaced, or refactored. This framing is partially wrong.
A legacy system is more accurately understood as a sedimentary record of institutional memory. Its internal structure contains:
A forty-year-old COBOL system is not merely a technical liability. It is a formal representation of how an organisation once modelled risk, payment flows, customer relationships, or product categories. That model may be wrong by current standards. It is not empty.
The central observation follows directly: the longer a system has been in operation, the more likely that its functional logic exists only within the system — no longer in any human's memory.
For the purposes of this memo, ontology refers to the operational mapping between system concepts and real-world decision objects — the structured set of entities, relationships, and constraints that govern how the system interprets inputs and produces outputs. This is distinct from the system's code (its implementation) and its documentation (its self-description, typically incomplete).
This is ontological loss: the system knows more than its operators can explain.
The significance of AI in legacy modernisation is not speed or code generation. Its primary capability is the synthesis of heterogeneous material into a coherent structural model.
AI can analyse source code across languages and decades, database schemas and their evolution, ticket system histories, operational and error logs, and documentation — typically incomplete, often contradictory. From this it can reconstruct the system's implicit ontology: what concepts the system uses, how they relate, and what constraints they impose.
This reconstruction is not complete. The critical limitation is the distinction between structure and intention:
Structural information is recoverable from data. The intentional layer is partly irrecoverable — it depends on historical decisions, organisational culture, and context that was never formalised.
This means AI does not replace the interpretive layer of modernisation; it relocates it. Previously, interpretation depended on individual experts' memories. Now it can be grounded in systematic reconstruction. The nature of modernisation shifts accordingly:
SM-007's convergence finding identifies a recurring failure pattern across domains: systems do not fail when capacity is exhausted but when the ability to make decisions under stress is broken. The failure mode is not shortage but coordination collapse.
The same logic applies to IT modernisation.
The recurring failure pattern in migration projects is not that the new system fails technically. It is that a portion of the functional logic disappears in transit:
This surfaces later as erroneous system behaviour, manual workarounds, business regressions, or unexplained entries in operational and financial metrics. It is a continuity failure, not a technical one.
AI's contribution is that this gap can be identified and modelled before the transition. The legacy system can be treated as a unified analytic object in which explicit logic (code), implicit logic (usage and data), and institutional logic (decision history) form a single interpretable system.
The goal is not to replace the system. It is to preserve functional continuity while replacing the technical implementation.
In this sense, AI is not a development tool. It is an instrument of institutional continuity — closer in function to the WEM or HEM instruments in ACI's own architecture than to a code compiler.
Naming the upside is insufficient. The specific risk introduced by AI-assisted modernisation is different from the risks it reduces.
Previously, the dominant risk was technical failure: the new system did not work. That risk is real and well-understood. AI can substantially reduce it.
The new dominant risk is wrong reconstruction: the system's structure is recovered accurately, but its functional meaning is interpreted incorrectly. The ontology is rebuilt — but rebuilt around the wrong model of what the organisation is actually doing.
This risk is more dangerous than technical failure because it is less visible. A system that crashes produces a clear signal. A system that operates coherently but on a subtly wrong ontological model produces local correctness with global drift: every individual decision is defensible within the system's internal logic, but the aggregate trajectory diverges from what the organisation is actually doing. This is the same failure pattern SM-007 identifies in institutional systems under stress — coordination collapse does not announce itself.
AI does not resolve the legacy problem. It makes the legacy problem explicitly analysable for the first time. That is a different thing.
There is a further limitation not yet named. Reconstructed ontology is static. It shows what is present in the system; it does not show which structures were errors, which were temporary compromises, and which were deliberate choices that subsequently became operational locks. Without the temporal dimension — why a structure appeared in a specific year, under what pressure, to solve what problem — the reconstruction is a map without elevation. It shows terrain but not history.
This means the instrument is most reliable when used with a trained interlocutor: someone whose tacit knowledge can distinguish ontological residue (historically contingent, no longer active) from operational core concept (still governing decisions). AI externalises the system's implicit model; the experienced practitioner adjudicates it. This makes the instrument dialogic rather than autonomous — closer in function to WEM or HEM than to a code generator.
The corollary is precise: an organisation that has already lost its tacit knowledge — key staff departed, institutional memory severed — does not benefit from this instrument. It receives an accurately reconstructed wrong map. The conditions under which reconstruction is not safe, even with the best available AI, are the subject of CN-023.
AI does not resolve the legacy problem. It makes the legacy problem explicitly analysable for the first time. That is a different thing.
The implication for modernisation practice: the interpretive question — "what does this system actually know about the world?" — cannot be delegated to the reconstruction instrument. It requires human judgement that understands both the domain and the limits of the model being built. AI handles the sediment. The tectonic question remains human.
Legacy modernisation's fundamental problem is epistemological, not technological: what is a system when its original design knowledge has been lost?
AI changes this in two ways. It makes systematic ontological reconstruction possible from distributed, incomplete sources. And it shifts modernisation's centre of gravity from code to conceptual model.
The risks do not disappear. They change character. Technical failure was the old dominant risk. Wrong reconstruction is the new one. Preparing for the second requires a different kind of discipline than preparing for the first — and most organisations do not yet have it.