Your Digital Workers Do Not Need a Data Lake: Agentic Data Mesh

TL;DR

Centralising your data is not a prerequisite for automating your work. It is a multi-year programme that delays every use case behind it. The agentic data mesh removes it from the critical path.

Our agentic data mesh is a dynamic query layer over the sources you already have. Ask a question in natural language. The agent works out which columns it needs, which systems hold them, and how to retrieve them. Your data never moves.

Discovery happens once, not every run. Every validated path to a value gets written to memory and reused. When a schema changes and the memorised route breaks, the agent rediscovers it and updates itself.

Reliability is the whole point. In our benchmarks, accuracy on a range of questions moved from 30 to 50 percent up to over 90 percent, and run-to-run consistency from around 20 percent to 75 to 85 percent, beating what we would expect from human experts.

Every value carries its provenance. Any number in the output traces back to its original source and every transformation applied to it, with an explainability report available on demand.

Our Argument:

Every enterprise AI programme hits the same wall. Before the agents can do useful work, someone decides the data has to be consolidated first. Master the entities. Build the warehouse. Agree the semantic layer. Eighteen months later there is a governance forum, a migration backlog, and still no automated workflow.

The cost of that decision is not the platform spend. It is the queue that forms behind it. Every use case waits for an infrastructure project that was scoped before anyone knew what the agents would actually ask for.

The industry data says the same thing from a different angle. In the 2026 State of AI Agents Report, 46 percent of organisations named integration with existing systems as a primary obstacle to agents in production, and 42 percent pointed to data access and quality. These are not model problems. They are plumbing problems, and enterprises keep trying to solve them by building more plumbing.

Consider how a competent new analyst handles the same situation. Nobody rebuilds the warehouse for them. They ask around, find out that finance owns the authoritative revenue figures, learn that the CRM export is two days stale, work out which join actually holds, and write it down so they never have to work it out again. Within a month they are productive on exactly the fragmented estate that supposedly blocks automation.

That is the behaviour we built into the Agentic Data Mesh, now shipping as part of the Digital Worker Factory.

How it works

The mesh sits above your public and proprietary sources and turns a natural language question into a query plan.

First it translates the question into a dynamic schema: the set of columns needed to answer it. For each column it consults its domain knowledge of each source, along with any relevant long term memories, and shortlists the sources worth testing. It then probes those sources to confirm the data is reachable and how to reach it, and writes every successful route to memory. Once a validated source exists for every element, it runs the full selection, including whatever filtering the question requires. Provenance is tracked at every step.

Three things follow from that design.

You do not have to map your estate upfront. There is no relationship graph to author, no ontology to ratify. The mesh builds its own picture of your sources as it is used, and if you already have documentation, a lightweight ingestion step gives it a head start.

Schema drift stops being a fire drill. When a memorised route stops working, the agent detects the failure inside the run, repeats discovery and validation, and updates the memory for next time. No re-engineering ticket.

The output is auditable. Every field can be traced to its source and its derivations, and the explainability tool reads the full query trace to produce a plain language account of how a number was arrived at. In regulated workflows that is the difference between an interesting result and a usable one.

None of this argues against warehouses. If you have Snowflake or Databricks, the mesh queries them like any other source and the Digital Worker deploys alongside. The argument is narrower and firmer: centralisation should be something you do because it earns its keep, not something your automation roadmap is held hostage to.

Start with the data you have, where it already is.