Data first: why we assess your data before we build an agent

We assess your data before we build an agent because an agent is only as good as the system of record it reads and writes. The hard part of enterprise agents is not the prompt. It is getting a clean, permissioned, current view of SAP, Oracle, the core or the CRM, and being allowed to write back. That is what the data assessment in Blueprint checks, in 1 to 2 weeks, before we sign a KPI.

Why this is the step everyone skips

MIT NANDA found that 95% of enterprise GenAI pilots produce no measurable P&L impact (Aug 2025). In our experience the most common reason is upstream of the model: the pilot was built on an export, a sample or a sandbox, and production data looked nothing like it. Menlo Ventures reported that only 16% of enterprise GenAI deployments in 2025 were real agents; the rest could not act, often because nobody had solved the write path to the system of record.

What the data assessment covers

• Systems of record: which system holds the truth for this process (ERP, CRM, core, WMS, TMS, HRIS) and which ones only hold copies.
• Read access: APIs, database views, events or files. Latency, rate limits, field coverage. What the agent will see and how stale it will be.
• Write access: which actions the agent needs to perform, through which interface, with which permissions and which approval step.
• Data quality on the fields that matter: duplicates, empty fields, free text where a code should be, values that changed meaning two years ago.
• Identity and PII: how customers, drivers or candidates are identified across systems, and what needs scrubbing before a model call.
• Residency and regulation: where the data may live, which cloud and region, and what the regulator expects for this process (CNBV, SAT, LFPDPPP, HIPAA).
• Observability: can we reconstruct, six months from now, what the agent saw when it decided?

What we find, typically

The same four things, in different orders. The ERP has the field but nobody trusts it. The API exists but is rate-limited to a number that cannot survive a campaign. The write path is a screen, not an interface. And the process has a shadow version in spreadsheets that the ERP never sees. None of these stops a project. All of them change the architecture, and it is much cheaper to know in week one than in week nine.

What we do with the answer

The assessment ends in one of three decisions. Build the agent as planned. Build the data layer first: an integration, a warehouse layer on Snowflake, an event stream, a clean view. Or pick a different process, because this one cannot be measured yet. We have proposed all three. The KPI Contract is signed only when the data can support the KPI.

The context layer starts here

The data assessment is also the seed of the context layer: the model of how your business actually works. Entities, relationships, policies, exceptions. It is built from your systems, not from a workshop, and it is what lets an agent tell the difference between a delayed order and a cancelled one, or between a customer who is late and one who is disputing an invoice.

The stack we bring

Twelve years of integrations across SAP, Oracle, NetSuite, Odoo, Salesforce, HubSpot and the core systems of banks and fintechs in México, a Snowflake partnership for the data layer, and a runtime that treats MCP as the interface between agents and tools. In 2026 we ran data discoveries for groups in logistics, real estate and manufacturing before any agent was scoped. That order is deliberate.

What to bring to the first session

• The name of the system where this process lives, and the person who administers it.
• One month of real records from that process, with PII removed if you prefer. An export is fine.
• The policy document the team actually uses, not the official one.
• The number you would like to move, and how it is measured today.

With those four things a Discovery Express, two sessions at no cost, is enough to tell you whether the process is ready, what the data layer would need, and which two or three quick wins are available before any agent is built.