Why do enterprise AI projects fail on data rather than models?
Because foundation models are already good enough for most enterprise tasks, while master data frequently is not. If the same customer exists three times, or the material master holds free text where it should hold attributes, the model inherits that and presents it fluently. Bad data in a spreadsheet looks bad; in a generated paragraph it looks authoritative. So the honest first step is usually data work, not model work.
When is RAG the right approach, and when is it not?
RAG suits questions answerable from documents you control and keep current — policies, procedures, product and contract content. It is wrong when the corpus is out of date, because retrieval does not fix stale content. It makes a stale answer sound authoritative and attaches a citation, which is worse than no answer. Audit the corpus and assign owners before building the pipeline.
Can an AI agent use our existing integration service account?
It should not. An agent on a shared account inherits every permission that account has accumulated, so one reasoning error or injected instruction has an unbounded blast radius. Each agent gets its own scoped token for only the actions it needs, every action is written to a replayable audit log, and anything irreversible passes a human approval gate.
How do you decide whether an AI system is good enough to deploy?
Against a test set built before the build: real cases with agreed correct outcomes, a scoring method, and a threshold signed off by the process owner in advance. Fixing the threshold before anyone sees the model perform is deliberate, because afterwards it drifts to wherever the model landed.
What drives the running cost of enterprise AI?
Three things: tokens per interaction, which retrieved context length and conversation depth dominate; vector storage plus re-indexing as content changes; and inference capacity if you host open-weight models. All of it is usage-based, so it is lowest during the pilot. Budget against second-year volume with realistic adoption, not against the pilot invoice.
Where does AI genuinely pay off in an enterprise today?
Document extraction and classification into ERP, demand forecasting measured against a recorded planner baseline, code and test generation, service-desk deflection on questions the knowledge base really answers, and summarisation inside a workflow people already use. Fully autonomous decision-making in finance or compliance is not there yet for most organisations, and we will say so rather than sell it.
What does AI governance need to contain to be usable?
Four concrete things rather than a policy statement: an approved-tools register, a data classification saying which class of data may go to which model, retention and logging rules for prompts and outputs, and named ownership per system. From those you can answer a specific question at a specific moment, which is what governance is for.
How do the EU AI Act and the DPDP Act affect what we build?
The EU AI Act classifies systems by risk, so obligations follow the use case — employment, credit and biometric uses carry materially more than a drafting assistant. India's DPDP Act bites on personal data in prompts, outputs and logs: purpose limitation, notice and retention apply to a prompt log as much as to a database.
