A8gent

Generator

Process Documenter and SOP Generator

Add your process steps, assign each to a human or an AI agent, and generate a formatted SOP with automation recommendations you can hand to a team.

Interactive documenter

Document a business process for automation

Map your process steps and see how much can be automated with AI agents.

Free

Automation potential

54

Fix first

Client Onboarding is a moderate automation candidate.

Agent steps

2 of 5

Steps that can be handled by AI agents.

Time savings

30-60 min

Estimated savings per execution of this process.

Guided action plan

1Document each step with trigger, action, and output
2Tag steps as human or agent
3Identify data sources for agent steps
4Build and test the first automated step

Recommended course

Auto guided

AI Agent Course for Business Owners

Start by identifying which process steps are truly repeatable before investing in automation tooling.

Open course path
Contents
  1. Fast answer
  2. Quick facts
  3. What this covers
  4. Why it matters
  5. Where to go next
  6. Outcomes
  7. Step by step
  8. The measured cost of keeping processes in people's heads
  9. How much of a documented process is typically automatable
  10. What changed in 2025-2026: SOPs became agent specifications
  11. The methodology: what the generator does with your steps
  12. What a generated SOP is missing before it can specify an agent
  13. A worked example: invoice chasing at a five-person accounting firm
  14. The limits, and where to go next
  15. At a glance
  16. Mistakes to avoid
  17. FAQ
TL;DRAnswer-first

Fast answer

The process documenter turns your process steps into a clean, formatted Standard Operating Procedure with automation recommendations after you add each step and assign it to a human or an AI agent. You break the workflow into discrete steps, each with a clear trigger, a single action, and a verifiable output, because automation requires precision and vague steps like handle the request cannot be automated while specific ones like extract the invoice number can. Each step is tagged human or agent based on rule complexity, variability, and stakes, so rule-based and consistent but low-stakes steps become prime agent candidates while judgment-heavy steps stay with people. The tagged steps compile into a structured SOP with ownership columns, tool recommendations, and exception handling. Each agent step includes a recommended tool, a fallback if it fails, and an escalation path to the nearest human step. Five to ten steps per SOP is ideal, and the document doubles as a specification for setting up the actual automations.

Facts last verified 2026-08

Quick facts

Time lost to missing documentation
About 5 hours/week per knowledge worker
Operational cost of poor documentation
20 to 30% higher (IDC, cited)
Managers reporting knowledge-loss pain
56% say it hurts onboarding
Work activities technically automatable
About 50% (McKinsey)
Businesses with fully automated workflows
Only about 4%
Reclaimable time via automation
About 240 hours/year per employee

On this page

What this page covers

A tool visitor should leave with a decision, not just a number: build now, prepare first, choose another workflow, or follow a course path.

  1. 01Steps
  2. 02Assignments
  3. 03SOP
  4. 04Recommendations

Why does this matter now?

You break your workflow into discrete steps. Each step must have a clear trigger (what starts it), a single action (what happens), and a verifiable output (what proves it is done). This granularity matters because automation requires precision. Vague steps like 'handle the request' cannot be automated. Specific steps like 'extract invoice number from email subject line' can.

Internal path

Where to go next from this page

These links are part of the A8gent learning and conversion path. Use them to move from concept, to diagnosis, to workflow build, to course.

Start with readiness

What you should be able to do after this

  • Step decomposition
  • Assignment logic
  • SOP generation
  • Automation recommendations

How do you do it, step by step?

1. Document the process as it is, not as it should be

Including the workaround everyone uses. An SOP describing the official version specifies an automation nobody will use.

2. Mark owners honestly

Agent means the step could run unattended given the right inputs. If it needs judgement about a customer, a price or a person, it is human for now.

3. Add the four missing fields by hand

Trigger, named inputs, review rule, failure path. Twenty minutes of writing turns a description into a specification.

4. Replace the time estimate with a measurement

Time one real execution with a stopwatch. The built-in 15 to 30 minutes per agent step is a constant, not an observation.

5. Copy it out and store it where the team looks

Nothing persists on refresh. Export the text, put it where procedures live, and date it.

The measured cost of keeping processes in people's heads

Undocumented process is not a tidiness problem; it is a recurring tax with published estimates. Knowledge management research finds knowledge workers lose about 5 hours per week waiting for information from colleagues or recreating work that already exists, which compounds to roughly 260 hours per person per year. IDC research cited across the documentation industry puts the operational cost of poor documentation at 20 to 30 percent higher costs from rework and error correction.

The cost concentrates at two moments: when people join and when people leave. In knowledge-loss surveys, 56 percent of managers report that departed employees' undocumented knowledge makes onboarding new people harder and less effective. Every SOP this tool generates is insurance against both moments, because the process stops being something a specific person knows and becomes something the business owns.

There is now a third moment: when you automate. An AI agent cannot infer a process that exists only as habit, and ambiguity that a human would resolve with common sense becomes an error an agent commits at machine speed. Gartner attributes the coming wave of agentic project cancellations, over 40 percent by end of 2027, primarily to management failures like undefined business value and weak controls rather than model capability. A written SOP with explicit steps, owners, and failure paths is the cheapest control that exists.

How much of a documented process is typically automatable

Once a process is written down step by step, the automatable share is usually larger than teams expect. McKinsey's long-running automation research estimates about half of current work activities could technically be automated, and workplace surveys find 94 percent of workers performing repetitive, time-consuming tasks in their roles. Yet workflow surveys find only around 4 percent of businesses have fully automated their workflows, a gap explained less by technology than by the absence of exactly the step-level documentation this tool produces.

The workers themselves can quantify the prize: survey respondents estimate they could reclaim around 240 hours per year if their repetitive tasks were automated, roughly six working weeks per person, and 51 percent of employees report spending at least two hours daily on repetitive work. When your generated SOP tags a majority of steps as agent-suitable, those benchmarks are why: documented business processes are disproportionately made of the rule-based, data-moving steps automation handles best.

The tagging discipline also protects you from over-automation. Steps involving judgment, negotiation, or high-stakes exceptions belong to humans, and a good SOP makes that boundary explicit with escalation paths rather than leaving it to the agent's confidence. In practice, most well-documented back-office processes land in a 60 to 80 percent agent, 20 to 40 percent human split, with the human steps concentrated at approval points and exception handling.

What changed in 2025-2026: SOPs became agent specifications

The audience for process documentation changed. An SOP used to be written for a new hire; it is now equally a specification an AI agent executes. That shift rewards a specific writing style the tool enforces: one action per step, explicit triggers, verifiable outputs, and named failure paths. Vague documentation was always annoying for humans; for agents it is disqualifying, because an agent cannot ask the person at the next desk what a step really means.

Tooling reflects the shift. Workflow platforms like n8n, Make.com, and Zapier effectively turn each documented step into an executable node, so a well-structured SOP can be translated into a running workflow almost mechanically, and screen-capture documenters can record current-state process for refinement. The teams moving fastest in 2026 maintain one artifact that serves both audiences: the SOP is the source of truth, and the automation is its implementation, updated together when the process changes.

Decision guidance: document the process that hurts most, but automate the process that is most stable. Those are often different processes. A chaotic process benefits enormously from documentation because writing it down exposes the chaos, but automating it freezes the chaos into software. Document first, run the documented version manually for a few cycles until the steps stop changing, and only then hand the stable steps to an agent. This sequence is slower by weeks and safer by an order of magnitude.

The methodology: what the generator does with your steps

You give the process a name and fill in steps. It opens with five rows, accepts up to ten, and will not go below three. Each row takes a description, an owner toggle set to human or agent, and a tool picked from a fixed list of nine: none, Autonoly, Zapier, Slack, email, CRM, spreadsheet, custom script, AI assistant. The generate button unlocks once you have a process name and at least three rows with text in them. Empty rows are dropped silently, so a five-row form with two blanks produces a three-step SOP and renumbers cleanly.

The only computed figure is the automation rate: the number of rows you marked agent divided by the number of filled rows, as a percentage. You set it yourself by toggling owners, which is worth stating plainly, because the output presents it as an assessment. Three bands then attach a sentence: 60 percent or above returns high automation potential, 30 to 59 returns moderate with advice to start with the agent-marked steps, and below 30 returns a note that the process needs mostly human judgement and suggests automating data entry and notifications with simple triggers. Those thresholds are fixed and the sentences are canned.

The estimated time saving is the crudest number in the tool and the most quotable, so treat it carefully. It is the count of agent-marked steps multiplied by 15 and by 30, presented as a range of minutes saved per execution. Four agent steps always returns 60 to 120 minutes, whether those steps are copying a reference number between two tabs or reconciling a month of transactions. There is no frequency input anywhere in the tool, so it cannot annualise anything, and you should not annualise it either without replacing that 15 to 30 minute constant with a timing you actually took. The copy-as-text button exports the whole thing as plain text with a header, one block per step and the summary, which is the format worth pasting into whatever your team already uses.

What a generated SOP is missing before it can specify an agent

The generated document records what happens, who does it and with what tool. An agent specification needs four things it does not capture, and adding them by hand takes about twenty minutes. The first is the trigger: what starts this process, and is it a time, an event, a message or a person noticing something. A process that starts because someone remembers to check a folder is not ready to be automated until the checking becomes an event. The second is the inputs and where they come from, named as fields in named systems rather than as 'the details'.

The third is the review rule, which is the difference between a workflow and an incident. For each agent-marked step, write down whether the output goes straight out or waits for a human approval step, and who that person is by role. Anything that emails a customer, moves money, changes a price, alters a record another team relies on, or makes a commitment on your behalf should start with a review and earn its way out of one. The fourth is the failure path: what the process does when the input is missing, the system is down, or the answer is ambiguous. Write the sentence 'if X is missing, do Y and tell Z', because an automation without that sentence silently does nothing, which is worse than failing loudly.

One more habit that pays: mark the steps you are unsure about rather than smoothing them over. Most processes contain one step that different people perform differently, and that step is nearly always where the automation breaks. A documented disagreement is a decision waiting to be made by a named person. An undocumented one becomes a bug report three weeks after go-live, and the fix is usually a rule someone could have written down on day one.

A worked example: invoice chasing at a five-person accounting firm

Take the practice manager at a small accounting firm running QuickBooks or Xero, documenting how overdue invoices get chased. Six steps: pull the aged debtors report (human, spreadsheet), identify invoices more than 14 days overdue (agent, spreadsheet), check whether the client has an open query or a payment plan (human, CRM), draft the reminder email (agent, AI assistant), send it (agent, email), and log the contact against the client record (agent, CRM). Four of six marked agent gives an automation rate of 67 percent, which lands in the high band, and an estimated saving of 60 to 120 minutes per execution.

Now do the work the tool leaves to you. The trigger is a schedule, every Tuesday at 9am, not a person remembering. The inputs are the invoice number, amount, due date, client contact and the query flag, each named in the system it comes from. The review rule is that step four drafts and a human approves before step five sends, at least for the first month and permanently for any client above a value threshold, because a reminder sent to someone who has already paid or who is mid-dispute costs more goodwill than the automation saves. The failure path is that if the query flag cannot be read, the invoice is skipped and listed in a summary to the practice manager rather than chased blind.

Then check the time estimate against reality. If the manager currently spends 40 minutes on this every Tuesday, the tool's 60 to 120 minute figure is wrong by a factor of two or three, and the real case is 35 minutes a week, or about 30 hours a year. That is still worth building, and it is a number you can defend. Our invoice chasing guide covers the build itself, and the risk scorer is worth a five-minute pass first because this workflow sends messages about money to customers.

The limits, and where to go next

The generator caps at ten steps, holds no branching, records no roles or names, has no frequency or duration input, and does not persist anything: refresh the page and the SOP is gone, so copy it out as soon as it is right. The nine tool options are broad categories rather than your actual stack. The automation rate is your own judgement reflected back at you, which makes it useful as a conversation with yourself and worthless as evidence to anyone else. And a linear list cannot express the two structures that break most real processes: a decision point with two different downstream paths, and a step that waits on somebody outside the business.

Use something else when the shape is wrong. For a process with branches and systems rather than a straight line, the workflow mapper and the workflow planner handle triggers, conditions and integrations properly. For turning a documented process into a scoped build brief, the workflow brief template gives you the document a developer or an agency can quote against. For deciding which of several documented processes to automate first, the task prioritizer ranks them, and the ROI calculator prices the winner.

The best reason to use this tool has nothing to do with automation. Written procedures are how a business stops depending on one person's memory, and most small teams have none. If you document five processes and automate none of them, you are still meaningfully better off: onboarding gets faster, holidays get easier, and the day someone leaves stops being a crisis. Automation is what you do after the writing, and the writing is the part that always pays.

At a glance

What the generated SOP contains, and what you have to add

ElementIn the generated SOPWhy it matters for an agent build
Step descriptionYes, up to 10 stepsThe sequence, in order, in plain language
Owner (human or agent)Yes, per step, set by youDrives the automation rate figure the tool reports
ToolYes, from nine fixed categoriesIndicates the system touched, not the account or connector
TriggerNoWithout it there is nothing to start the workflow
Named inputs and sourcesNoField names in named systems are what a build maps
Review or approval ruleNoDecides whether output goes out unattended
Failure pathNoWithout it, failures are silent instead of loud
Frequency and real durationNoThe tool assumes 15 to 30 minutes saved per agent step
Branching and exceptionsNoLinear list only; branch-heavy processes need the workflow mapper

What mistakes should you avoid?

  • Quoting the automation rate as an assessment when it is the ratio of steps you personally toggled to agent.
  • Annualising the 15 to 30 minute per step estimate, which is a fixed constant with no relation to your process.
  • Documenting the official process rather than the one people actually follow.
  • Leaving out the review rule on any step that contacts a customer or moves money.
  • Writing a linear SOP for a process that really has two branches, then discovering the second branch in production.
  • Losing the document by refreshing the page before copying it out.

FAQ

How does the tool calculate the automation rate?
It divides the number of steps you marked as agent by the number of steps you filled in, and shows that as a percentage. You set it yourself with the owner toggles, so it reflects your judgement rather than an analysis of your process. Three bands attach a canned sentence: 60 percent and above reads as high potential, 30 to 59 as moderate, and below 30 as a mostly human process. Useful for structuring your own thinking, not evidence to show a stakeholder.
Where does the time-savings estimate come from?
It is the number of agent-marked steps multiplied by 15 and by 30, expressed as minutes saved per execution. Four agent steps always shows 60 to 120 minutes regardless of what those steps are. There is no frequency or duration input anywhere in the tool, so it cannot know how long your steps take or how often the process runs, and it cannot annualise. Time one real execution and substitute your own figure before you quote a number to anybody.
How many steps should a process have?
The tool accepts three to ten and most useful processes land between five and eight. If yours needs more than ten, it is probably two processes with a handover in the middle, and documenting them separately makes both easier to automate. If it needs fewer than three, it is a task rather than a process and belongs in the task prioritizer instead. Ten is a cap in the code rather than a rule of practice, but it is a reasonable one.
Can the SOP be used as a specification for an AI agent?
Not on its own. It records what happens, who does it and with what tool. A build also needs the trigger that starts it, the inputs named as fields in named systems, the review rule for each automated step, and the failure path when an input is missing or a system is down. Those four are absent by design and take about twenty minutes to add. Written procedures became agent specifications the moment agents could follow them, and the missing four are what separate the two.
What should stay human even when it could be automated?
Anything where a wrong output is expensive to unwind: customer-facing messages about money or disputes, price changes, contract language, hiring decisions, anything clinical or legal, and any step where the input is a judgement about a person. The practical rule is that a step earns its way out of review rather than starting there. Run one month with a human approval step, count the corrections, and remove the review only if the correction rate is near zero.
Does the tool save my SOP?
No. Nothing persists and nothing is sent anywhere, so refreshing the page loses the document. Use the copy-as-text button as soon as the SOP is right and paste it wherever your team keeps procedures. The exported format is plain text with a header block, one block per step showing action, owner and tool, and a summary with the automation rate, which pastes cleanly into a document, a wiki or a ticket.
What if my process has branches?
This tool cannot express them. It produces a linear numbered list, so a process that splits on a condition (new client versus existing, under threshold versus over) has to be either simplified or documented as two SOPs. Two SOPs is usually the better answer, because the branch is where the exceptions live. If the branching is genuinely central, use the workflow mapper or the workflow planner, which handle conditions and multiple systems.

Sources & further reading

Was this page helpful?