Why AI projects need boring plumbing before they need agents
Most useful AI projects start with clean inputs, stable workflows, and reliable handoffs before anyone needs a complex agent.
TL;DR / Key Takeaways
- AI cannot fix a workflow that nobody understands.
- Source systems, permissions, validation, and logging matter before model selection.
- The best first automation is often deterministic, not generative.
- A reliable handoff beats a flashy demo.
Most failed AI projects do not fail because the model was not smart enough. They fail because the surrounding workflow was not ready.
Before choosing an agent framework, make the plumbing visible.
The demo problem
A demo can make a messy process look simple for five minutes. Production work is different.
Real business workflows have:
- Incomplete inputs
- Duplicate records
- Missing permissions
- Unclear ownership
- Edge cases nobody wrote down
- Manual review steps that still matter
AI can help, but it cannot make those problems disappear.
Start with the workflow
Write down what happens today. Include the human steps, the systems involved, the files copied by hand, and the decisions that happen in email or chat.
Once the workflow is visible, the implementation path gets clearer. Some steps become API integrations. Some become scheduled jobs. Some become validation rules. Some become AI-assisted drafts or summaries.
Clean inputs beat clever prompts
Prompt quality matters, but input quality matters more. If the source data is inconsistent, missing, or stale, the AI layer becomes a polite wrapper around bad information.
Before model selection, confirm:
- Where the data lives
- Who can access it
- How it changes
- What fields are required
- What validation should happen before AI sees it
Where AI actually belongs
AI belongs where deterministic code struggles:
- Reading messy documents
- Classifying inbound requests
- Drafting responses for review
- Summarizing long context
- Extracting fields from inconsistent text
It does not need to own the whole workflow.
What to build first
The best first build is usually the smallest reliable system:
- A form that captures clean inputs
- A database table instead of a fragile spreadsheet
- A scheduled job that moves data consistently
- A review queue for AI-drafted output
- A dashboard showing what ran and what failed
Practical checklist
- Map the current workflow before discussing models.
- Identify source systems and permissions.
- Add validation before automation.
- Log every automated action.
- Keep humans in review where judgment matters.
- Document the handoff so the team can operate the system.
Related Reading
Related practical notes
Building a practical alerting system for Fivetran and dbt failures with AWS Lambda, DynamoDB, Teams, and Jira
A useful alerting system does not just say something broke. It knows whether the failure is new, whether it already alerted, whether it recovered, and when to create a ticket.
Read articleCX Alloy API connector lessons when throughput is painfully slow
Some APIs are not hard because authentication is impossible. They are hard because the shape, pagination, nested resources, and throughput force you to design for patience.
Read articleWorkday SOAP APIs, custom connectors, and the joyless art of not getting rate-limited
SOAP APIs can still power serious enterprise integrations, but they demand patience, defensive engineering, and a connector design that respects pagination, retries, payload size, and rate limits.
Read article