The AI cannot see your data. That is the real problem
Independent survey after independent survey names the same blocker. Not model capability, not skills, not budget: whether the system can reach the right information, with the right permissions, in a form it can use.
What we actually build
Entitlement-aware retrieval
The hard part of enterprise AI is not retrieval, it is retrieval that respects who is allowed to see what. We model permissions first and index second.
Context and semantic layers
A stable vocabulary between your systems and the model, so the answer does not change because someone renamed a column.
MCP integration to systems of record
CRM, ERP, case management, document stores and internal APIs. MCP has won as a connector standard; the unsolved parts are identity, audit and gateway policy, which is where the work is.
Evaluation and observability
Instrumented from day one on open standards, so you are never locked into whoever is monitoring you. This is also what makes the run contract possible.
Lineage and provenance
Where an answer came from, which documents, which version, at what time. Your auditor will ask, and so will your customer.
Legacy enablement
Long-lived systems that resist change. AI coding gains are large on greenfield and close to nothing on complex legacy, which is precisely why this stays human work.
A shape we can defend in a technical review
Systems of record
- Salesforce
- SAP / NetSuite
- SharePoint
- Case management
- Internal APIs
Access layer · What we build
- Identity & entitlements
- MCP connectors
- Context & semantic layer
- Audit log
Application
- Workflow
- Model / agent layer
- Deterministic steps
- Tool calls
- Human review point
- Output to system of record
Evaluation & ops
- Trace collection
- Golden datasets
- Judge pipeline, sampled
- Drift detection
- Cost telemetry
- Incident alerting
- Monthly report
Regression gate
Open standards throughout, so the observability backend can be swapped without re-instrumenting anything.
Illustrative reference shape. Every engagement produces an architecture specific to your estate.
Questions worth answering
What is the biggest barrier to getting value from enterprise AI?
Data readiness and access, not model capability. 58% name data readiness and access as their number one barrier (KPMG 2026), 72% cite data quality as the top obstacle (Deloitte Private 2026), and three in four UK professional services firms are unready on data, orchestration and monitoring (GOV.UK 2026).
What is entitlement-aware retrieval?
Retrieval that respects who is allowed to see what. Permissions are modelled first and the index is built second, so a model can reach the information it needs and nothing it should not. It is the part of enterprise RAG that most often stops a rollout.
Do you use MCP to connect to systems of record?
Yes. MCP has effectively won as a connector standard for CRM, ERP, case management, document stores and internal APIs. The unsolved parts are identity, audit and gateway policy, which is where the engineering work actually sits.
Start with what is actually blocking you
The baseline usually finds that the problem is not the model. It is the four systems that will not talk to each other.