Building AI agents inside a retail ERP
What I learned connecting Claude to NetSuite via MCP servers, and why context engineering matters more than prompt engineering.
Read →3 min read
Topic
AI-Augmented Operations
AI-augmented operations is the practice of integrating AI agents and large language models into enterprise workflows to improve human decision-making rather than replace it.
AI-augmented operations integrates AI agents and large language models into enterprise workflows to improve human decisions rather than replace them. The goal is giving people better information faster, not removing people from the process.
There are two ways to deploy AI in an enterprise. The first is to automate a process end to end and hope the model is right. The second is to put the model where a person is already making a decision, and give them better information than they had. The second one works.
The technical work is less about prompting than most people expect. Connecting a model to live business data: an ERP, a warehouse, an order system: is a context engineering problem: deciding what the model can see, how that data is structured, what it is explicitly told it does not know, and what happens when a tool call fails.
I build these systems against production data: a language model connected to the ERP through structured tool servers, running read-only for four months before any write path was enabled. The lessons that mattered were about schema documentation and error paths, not about model choice.
Key concepts
The vocabulary that comes up most in this area, defined.
| Term | Definition |
|---|---|
| Context engineering | Designing the information environment an AI agent operates in: what data, instructions, and tools it sees at inference time. |
| Agentic AI | A model that plans and executes multi-step work using tools, rather than answering a single prompt in one turn. |
| Tool interfaces | A standard contract between a model and an external data source, replacing bespoke API wrappers written per query pattern. |
| Read-only first | Granting an agent query access long before write access, so it demonstrates it understands the data before it can change it. |
| Negative context | Explicitly telling a model what data it does not have. Prevents confident answers built on absent inputs. |
Articles on ai-augmented operations
Frequently asked questions
AI-augmented operations is the practice of integrating AI agents, large language models, and intelligent automation into enterprise workflows to enhance human decision-making and operational efficiency without replacing human judgment.
Context engineering is the discipline of designing, managing, and optimising the information environment that AI agents operate within. It involves curating what data, instructions, and tools an AI model sees during inference to produce reliable, accurate outputs.
Prompt engineering shapes the instruction. Context engineering shapes everything else the model sees: the data, the schema documentation, the tool definitions, and the ordering. In production systems, context determines output quality far more than phrasing does.
Yes, if you start read-only. Give the agent query access first, verify it interprets the schema correctly over months rather than weeks, and add write paths one operation at a time with explicit error handling.
A standard interface replaces building a bespoke wrapper for every query pattern. Each new capability becomes a separate deployable server rather than a change to an existing one, so the surface grows without the integrations coupling to each other.
Other topics