Guide · Legacy modernization
An agent that cannot write to your system of record is a demo
The constraint on most AI work is rarely the model. It is the integration layer between the model and the record, and the reason nobody has touched that layer is usually that a failed change there has to be disclosed.
Three routes teams take, and the mechanism that stops each one
The AI gets built beside the record
It reads a replica and writes nowhere. That demonstrates well and it never becomes the behavior of the system, so it has no path into a release. The work gets done twice, and the second time under a deadline.
The upgrade defers one more cycle
Modernization carries no deadline of its own, so it loses to features until an end-of-life date supplies one. By then the window is short enough that the migration happens under time pressure, which is how the expensive version gets built.
Data quality is assumed
No profiling has been run. The model then answers confidently from records with duplicate keys, stale joins and constraints nobody enforces. A wrong answer delivered with confidence costs more credibility than shipping nothing.
82 percent of IT leaders name data integration as one of the biggest challenges they face when using AI. It is the layer, and it is where the work is.
Salesforce 2026 Connectivity Benchmark Report. 1,050 IT leaders at organizations of 1,000 or more employees across 9 countries, fielded October to November 2025.
The path runs into the system of record
Three things change the odds, and none of them are about the model.
- Modernize the integration layer first. The path runs into the system of record, because that is the only version of the work that survives a release review. Read paths and write paths both, with the write path tested before anything depends on it.
- Anchor the work to a date that already exists. An end-of-life, a committed migration, an audit cycle. Those supply the forcing event that modernization never generates on its own, and we work backwards from them.
- Put engineers on it who have carried reportable systems through audit. 20 years on systems where a failed change gets disclosed. That experience is the reason a rollback path and a test suite exist before the first commit, and not after the first incident.
When a modernization is not the answer
- A rewrite with no business event behind it. With no date and no blocked initiative, it defers again the moment features arrive. That is a planning problem and we cannot solve it from outside.
- A data-readiness program you want priced before it is scoped. We profile first. Quoting an unknown is the mechanism by which these overrun.
- No engineering function on staff. Embedded engineers work to a team’s standards. With no team, there is nothing to embed into.
Bring the system nobody has changed in three years and the date you are working to. You will leave with a read on the path, including the parts we would decline.