Savvy AI

The problem

SMBs pay for workflow tools that require them to configure everything. Setting up automation is its own project — schema mapping, trigger chains, if/then trees, error branches. Most teams either give up (leaving the tool paid for and unused) or over-configure for edge cases they will never hit (paying maintenance cost forever on hypotheticals). The tools are powerful, but the setup tax is what kills adoption. The agent-first thesis is that the setup tax is now avoidable. If the user can describe what they want in natural language and the agent can propose an execution plan the user approves, most of the configuration work becomes conversation. The hard part is making that reliable enough for production.

Key challenges

Natural-language authoring of business workflows is a genuinely hard product problem. Users describe outcomes vaguely ("send a follow-up to leads that go cold"), the mapping to concrete tool calls is often ambiguous, and mistakes are expensive (sending the wrong follow-up to the wrong list). The agent has to be confident enough to propose useful plans, cautious enough to check with the human before acting on ambiguity, and transparent enough for the user to trust what they're approving.

What we built

AR Data built Savvy AI as an agent-first workflow platform. The user describes the workflow in natural language; the agent proposes a step-by-step plan with the concrete tool calls it intends to make; the human reviews and approves; the workflow runs. Multi-step orchestration covers CRM updates, email drafting and send, calendar coordination, and third-party integrations via a common tool interface. The agent uses LangGraph for orchestration — deterministic state transitions inside the agent loop, human-in-the-loop approval gates at defined boundaries. Every action is logged, every decision is inspectable, and every workflow can be re-run in a dry-run mode that shows what would happen without side effects. Errors surface with context rather than as opaque failures.

Our approach

1. 1

   Plan-then-execute over agent-loop-and-hope

   Rather than a free-running agent that acts and reports, Savvy proposes a plan and asks for approval. This is the difference between production-viable automation and a demo that scares users.

2. 2

   LangGraph for deterministic orchestration

   Agent loops are non-deterministic by nature; LangGraph provides deterministic state transitions inside the loop, which makes production behavior reproducible.

3. 3

   Dry-run mode as a first-class feature

   Every workflow can be executed in a mode that logs what would happen without doing it. This is the safety net users need to trust the system with real actions.

4. 4

   Common tool interface for third-party integrations

   Instead of custom integrations per tool, a common interface (auth, invocation, error handling, retry) lets new integrations land in hours instead of weeks.

Key architectural decisions

Results

• Natural-language workflow authoring — describe, review, run
• Multi-step agent orchestration with human-in-the-loop approval gates
• CRM, calendar, email, and third-party integrations via common interface
• Dry-run mode for every workflow — preview before commit
• Deterministic state transitions inside the agent loop
• Structured logs of every decision and action for audit and iteration
• Rapid integration onboarding — new tools in hours, not weeks

Impact

Savvy AI represents a shift in how SMB workflow automation should feel — less configuration, more conversation. Getting the reliability right (plan-then-execute, LangGraph orchestration, dry-run, approval gates) is what made the shift viable in production rather than just a nice demo. The engagement also shaped how we approach agent-based product work for other clients: agent freedom is proportional to blast radius, and blast radius is controlled by structure, not by trust.

Tech stack