Answer
How do you turn a manual process into an agent workflow?
Watch it being done, write down the decisions rather than the steps, and find the ones nobody can state as a rule.
Watch somebody do it, record the decisions rather than the actions, and mark which decisions can be stated as rules. The ones that cannot are the agent's job; the rest are ordinary automation.
The first move is observation rather than description, because a described process and a performed process differ in a consistent way. Descriptions contain the intended path and omit the exceptions, the checks people make without noticing, the cases they handle differently and the judgements that have become automatic. Those omissions are precisely the content that determines whether an automated version works, so asking somebody to write down how they do it produces a document that is accurate and incomplete in exactly the wrong places.
What to record is the decisions, not the actions. Opening the record, checking the field, sending the message are steps; deciding this is urgent, deciding this customer gets the exception, deciding this needs a person are decisions. The steps are the easy part to automate and the decisions are where the process actually lives, and a specification built from steps produces something that performs the motions and gets the outcomes wrong.
Then classify each decision by whether it can be written as a rule. Most can, and this is usually a surprise: what felt like judgement turns out to be a condition nobody had articulated, and articulating it is itself valuable. The residue that genuinely resists — reading what a customer meant, weighing two considerations, recognising an unusual case — is the part that needs a model, and it is normally one or two decisions rather than the whole process.
The exceptions deserve a separate pass because they are where automated versions fail and where descriptions never go. Ask what happens when the usual thing does not apply, and keep asking until the answers stop. Every exception is either a rule to write, a case to route to a person, or evidence that the process is more varied than a single workflow can serve. Discovering the third at this stage is much cheaper than discovering it after building.
Build the smallest version that handles the common path and routes everything else, then run it alongside the manual process for one full cycle with the disagreements recorded. That comparison is the specification for the second version, and it is built from real cases rather than from anyone's account of them. Businesses that skip it build the version described to them and then discover what the process actually contains.
One thing to decide early: what the automated version does not do. A process that has been performed by a person contains judgements they made for reasons they may not be able to state, and inheriting all of them is not the objective. Being explicit that the workflow handles this and not that produces a narrower thing that works rather than a broad one that half works.
Ask how it is done and you get the steps; watch it being done and you find out what is actually being decided.
Siddharth Sharma, Context Theory
Related questions
What if different people do it differently?
That is a finding worth having before automating rather than a problem to resolve during it. Where the variation is preference, pick one and the automation standardises it, which is a benefit. Where the variation reflects something real about the cases each person handles, the workflow has to accommodate it or serve only a subset, and knowing which you have determines what to build.
How long should the observation take?
Long enough to see an awkward case, which is usually a few instances rather than a formal study. The observation that matters is the one where the person does something the description did not mention, and that happens quickly. If several instances pass with nothing unexpected, the process is more uniform than most, which is itself useful to know.
METHOD
Every figure below carries its source and the date it was verified. Nothing on this page is asserted.
The numbers on this page.
| What | Value | Specific to |
|---|---|---|
| Sub-15-minute compliance — automated routing vs manual only | 62.5% vs 39.1% | Category-wide |
| Odds of qualifying a lead — replying within the first hour vs after it | 7× | Category-wide |
2026 speed-to-lead benchmark · verified
Oldroyd, McElheran & Elkington, "The Short Life of Online Sales Leads", Harvard Business Review (March 2011) · 1.25M inbound leads across 2,241 US firms · verified
What is specific to this page.
| Kind | Claim | Check it against |
|---|---|---|
| Workflow | Described processes contain the intended path and omit exceptions, unnoticed checks, differently handled cases and automatic judgements, which are precisely the content determining whether an automated version works. | Comparing a written description of a process against an observation of it being performed. |
| Response | Recording decisions rather than actions locates where the process lives, because a specification built from steps produces a system that performs the motions and reaches different outcomes. | Listing the decisions in a process separately from its actions and checking which determine the result. |
| Software | Most decisions that feel like judgement turn out to be unarticulated conditions, leaving one or two that genuinely resist statement as rules, which are the only parts requiring a model. | Attempting to write each identified decision as a rule and counting those that resist. |
| Constraint | Variation between people performing the same process is either preference, which automation can standardise, or a response to genuinely different cases, which the workflow must accommodate or exclude, and knowing which determines what to build. | Comparing how two people handle the same case and asking each why they did it that way. |
Each row would be wrong on another industry's page. Where a sourced figure exists it is in the table above instead; these are the constraints that shape the work and do not happen to be numbers.
Start with the measurement.
Reading about a benchmark is not the same as knowing your own number. The audit produces yours, measured rather than estimated.
$497 · delivered in 5 business days · credited against month one