The useful architecture is not an AI layer that can do anything. It is an operational core that lets AI prepare work while permissions, approvals and evidence remain authoritative.
Let AI interpret and prepare. Let the operational core authorise and commit.
Start with the lifecycle, not the chatbot
A CRM or ERP becomes valuable when it represents the real lifecycle of a lead, customer, case, order or asset. Before adding an agent, map the states that matter, who owns each decision, what evidence is required and which transitions are irreversible.
This exposes the places where AI is genuinely useful: understanding incoming requests, recovering context, proposing the next action, drafting a response or identifying an exception. It also exposes where AI should not be the authority.
Separate reasoning from commitment
We treat the model as a replaceable reasoning runtime. The business system remains responsible for identity, permissions, approvals, idempotency, audit history and reconciliation with external providers.
- AI reads a scoped, permission-filtered view of the record.
- Every proposed action is expressed as a typed business operation.
- Rules and approvals decide whether that operation may proceed.
- The resulting state and evidence are written back to one authoritative record.
Design for retries and disagreement
Real operations contain duplicate emails, late responses, concurrent edits and provider timeouts. An agent workflow must expect the same request to arrive twice and the same action to be retried after an uncertain result.
Give every consequential operation a stable key, record its input digest and read back the resulting state. When data sources disagree, preserve the disagreement and route it for review rather than asking the model to silently choose a winner.
Measure an operating outcome
A strong pilot follows one bounded workflow end to end. Measure elapsed time, rework, exception rate and evidence completeness—not merely how often the model produced a plausible answer.
The acceptance test is simple: can the team understand what happened, recover from failure and remain accountable for the final decision?
