Auditing your workflow for agentic AI
The Audit phase of the A.G.E.N.T. playbook, in practice — build an honest picture of how work is done today, define the outcome that actually matters, and remove the unknowns before you automate anything.
Audit before you automate
Every agent we ship starts with an honest audit of the workflow it will run. Not a full process map with every branch documented — most of those steps are about to change — but a clear picture of how work actually moves today: the objective, the data in play, the systems that touch it, and the people who execute or review it. Skip this and you get well-built agents pointed at the wrong problem: automation that accelerates activity nobody needed, adds complexity, and opens governance gaps you only discover in production.
Focus — the outcome, not the activity
The audit is outcome-driven. The question is never just "what tasks happen here" but "what result actually matters." A workflow is worth automating when improving it moves something real — quality, velocity, cost, governance — not when it simply produces more output faster. So we start with why the workflow matters strategically, what success would look like beyond finishing tasks quicker, and how you'd know the improvement created genuine value. Get Focus right and every later decision — architecture, orchestration, guardrails — has a target to optimize against.
Action — map, interview, document
Three concrete moves. Map the goals, data flows, systems, and human roles the workflow depends on, across both internal delivery and client-facing environments. Interview the operators, reviewers, and stakeholders who live in the process — that's where you find the friction, rework, and policy risk that no diagram surfaces on its own. Then document the desired business outcome, not just the outputs, so the agents you design are optimizing for value rather than volume. Pay particular attention to compilation and "checking" work — humans verifying other humans' output — which is almost always ideal agent territory.
Goal — an as-is and a to-be with no unknowns
The Audit ends with two things: an unvarnished picture of the current state and a precise target outcome. Automating without the first is the fastest way to inherit hidden risk — undocumented exceptions, silent dependencies, decisions that only live in someone's head. Defining the second is what lets you design agents, governance, and orchestration with intent instead of guesswork. Before moving on, we name the unknowns that still remain and close them. Only then is a workflow ready to leave audit and enter design.
Where Audit fits in the two-month sprint
Audit is weeks one and two of a focused eight-week sprint — paired with Gauge, where opportunities get prioritized. The rule for a first sprint: pick one high-value workflow with clear inputs and outputs, meaningful manual effort, and real business impact when accelerated — but not a mission-critical process, because you need room to learn safely. Map the current state honestly, find the 20% of activities creating 80% of the value, and treat everything else as a candidate for agent automation. The discipline is momentum over perfection: prove it on one workflow, then scale the pattern.
The A.G.E.N.T. playbook — Audit, Gauge, Engineer, Navigate, Track — was introduced by Dirk Hofmann and Ulla Kruhse-Lehtonen in the Harvard Data Science Review. This piece reflects how AR Data applies the Audit phase in delivery.
Read the full A.G.E.N.T. field guide →Ready to audit a workflow?
We run the Audit phase on one high-value workflow — as-is reality, target outcome, and the unknowns to close before build.
