10 lessons from building
AI agents that people
actually use.
Practical patterns from 18 months of production
multi-agent systems in the enterprise.
Hi, I'm Aman.
Sr. AI Engineering Architect. 18+ months building multi-agent AI systems over structured enterprise data. Thousands of daily users. Before that: distributed systems, cloud architecture, and a deep belief that the best technology disappears into the workflow.
Where these lessons come from
18 months building a multi-agent system over enterprise SQL data, used daily by thousands of business users. These 10 lessons shaped whether the system got adopted or ignored.
These will resonate most if you are...
- Building agents for business users, not analysts or developers
- Working with structured data (text-to-SQL, database-backed tools)
- Using multi-agent orchestration (LangGraph, CrewAI, etc.)
- Trying to move past demos into daily production use
- Operating with compliance, access control, and internal jargon
The common thread
None of these are about which model or framework to pick. They're about the patterns between the model and the user that determine trust, adoption, and retention.
Some are emerging across the industry (think OpenClaw-style ambient agents). Others are specific to enterprise data. All come from shipping, watching users, and iterating.
Views and opinions expressed in this presentation are my own and not those of my employer. Examples are generalized and do not disclose proprietary information.
03 / 15Your Users Speak a Language
the Model Doesn't.
Every org has internal shorthand. Acronyms that mean different things in different contexts. LLMs guess confidently. They guess wrong.
Thousands of internal acronyms per org, many overlapping across departments.
The pattern
Resolve jargon before agents execute. Use surrounding context for disambiguation, not just the term itself.
~50% of user queries benefit from disambiguation.
Assume Sensible Defaults.
Show Your Work.
Clarification questions add friction. LLM accuracy degrades up to 33% in extended multi-turn conversations.
Microsoft Research, ICLR 2026
AI: "Which region?"
User: "North America"
AI: "What time period?"
Three turns for one question. Early wrong turns compound.
AI: All regions, current quarter.
Region: All · Period: Q3 FY26 · Metric: Revenue
Users see the assumptions. They narrow from there.
Data filters transparency
Critical in text-to-SQL. Business users think in portal labels. The DB schema uses different names. "Active" = status='CURRENT'
Show which filters were applied, in language they understand. Lets them catch mismatches. Builds trust.
Intelligent reflection
Results sparse but close match exists? Auto-retry with corrected input. Resolve ambiguity behind the scenes.
Guide Discovery.
Don't Hope for Good Questions.
Most users don't know what your system can do. If the LLM suggests follow-ups, it will hallucinate capabilities you don't have.
~40% of follow-up interactions come from guided suggestions.
The architecture
Decouple from your main agent. Own trigger policies, capability grounding, novelty filtering. Run in parallel, zero added latency.
Open-sourced this as a dedicated LangGraph node:
Security Is Not an
LLM Problem.
Access control cannot live in a prompt. Not partially. Not with clever instructions. Not at all.
Might work for one query. Might not for the third. Can be jailbroken. You can't tell an auditor access control "usually works."
Empty Results Are Never
"No Data Found."
Three distinct reasons a query returns nothing. Each needs a different response.
Access-filtered
Data exists, user can't see it. Say so: "Records exist but aren't visible with your current access."
Filter mismatch
Entity exists, filters don't match. "No Q3 orders for Acme. Could mean they ordered in Q2."
Query error
Filter values don't exist. Misspelling, wrong ID format. Analyze which filter failed and suggest corrections.
Never let your synthesis layer say "no results" without explaining why.
08 / 15Personalization Is Not a Feature.
It's the Product.
A system that treats every user the same will plateau on adoption no matter how accurate it is.
Shortcuts
User-defined command aliases. /pipeline west expands to a full multi-line prompt.
5+ shortcuts = 2-3x engagement.
Scheduled delivery
"Every Monday, send me a summary of my open deals." Reactive to proactive. Insights without logging in.
What turns AI from a tool into a habit.
Long-term memory
"This client is price-sensitive." Remembered across sessions. The assistant adapts to you.
Memory feeds suggestion quality. Virtuous cycle.
Your AI Has Blindspots
If It Only Sees Chat.
Users operate in dashboards, spreadsheets, email threads. Their real context lives outside the chat window.
The broader principle
The AI should ingest the same stimulus the human sees. Context changes? AI context changes with it.
Embed AI where humans already work. Chat, email, team channels, dashboards. Don't force context-switching.
Let Users Build on Top
of Your Platform.
You can't anticipate every workflow. Give users composable building blocks. Let them assemble their own.
The primitives approach
Expose agents and tools as composable steps. Users chain them in natural language. You provide: gather, decide, act, synthesize. They describe the recipe.
Every step deterministic. No context loss. Full visibility. This takes you from "chatbot" to platform.
Distinct from personalization
Lesson #6 = adapting the system to users (shortcuts, memory). This = letting users create new capabilities that didn't exist before.
Best enterprise products become platforms. Users who compose their own workflows become your strongest advocates.
Meet People Where They Work,
Not Where You Built.
Your users live in email, team channels, and video calls. If AI only works inside your app, it will always have blindspots.
Email delegation
CC the assistant on any thread. It reads context, picks out the ask, runs work in the background. No context-switching.
Team channels
Tag it in a group conversation. Answers questions, pulls data, takes actions alongside the team. Not a 1:1 silo.
Proactive intelligence
Don't wait to be asked. Monitor for changes, detect anomalies, surface insights. Event-driven, not query-driven.
The measure of a great assistant isn't what it can answer. It's how little the user has to remember to ask.
12 / 15Build the AI System
with AI.
The meta-pattern. The value isn't in general intelligence. It's in encoded constraints that channel intelligence toward useful outputs.
What this looks like
- Machine-readable rules for code conventions, architecture, domain requirements
- Structured command files for complex multi-step operations
- New developers contribute effectively from day one
- Multi-week sprints collapse to days
The compounding advantage
A model with no context writes generic code. The same model with project conventions writes code that fits your codebase.
Teams that make their codebase legible to AI from the start compound that advantage every sprint. The gap is already a chasm, and it's widening.
The 10 Lessons
Your users speak a language the model doesn't
Assume sensible defaults, show your work
Guide discovery, don't hope for good questions
Security is not an LLM problem
Empty results are never "no data found"
Personalization is not a feature, it's the product
Your AI has blindspots if it only sees chat
Let users build on top of your platform
Meet people where they work, not where you built
Build the AI system with AI
Keep building.
More field notes at amansardana.work. Let's connect.
15 / 15