BonoficiGrade your business
Menu
Lesson 5 of 5 · 6 min read

AI workflow automation: from process map to first step

How to score each step of a process map for AI workflow automation, where AI beats fixed rules, and how to start with one step and a human approval.

AI workflow automation works best one step at a time, starting from a process map that includes the exceptions. Score each step on volume, how clear the rules are, whether the data is available, the cost of an error and how often exceptions happen. Then pick one high-volume, low-risk step, use AI for the reading and classifying, keep a person approving, and measure before and after. Your existing systems stay the source of truth.

Start from the map, not the tool

Most automation projects start with a tool and look for something to use it on. It works better the other way round. If you have followed the earlier lessons, you have a process map with owners, systems, hand-offs and waits, an SOP and an exception log. That is what you need to decide which steps to automate, and how.

Without the map, you automate the documented process, and the first exception, such as the order without a PO or the invoice that does not match, sends the work back to a person with no context.

Score each step

CriterionQuestionScores well when
VolumeHow often does this step run?Daily or more; dozens or hundreds of times a month
Rule clarityCould you write down how to decide?The SOP gives clear rules, and two people would make the same call
Data availableIs the input somewhere you can reach?An email, a PDF, a spreadsheet, or a system with an export or API
Cost of an errorWhat happens if it is wrong?Easy to catch and cheap to fix before it reaches a customer or a payment
Exception rateHow often does it leave the happy path?Most cases are routine, and the exceptions are known and logged

Score each criterion from 1 to 3 and add them up. The scoring is rough on purpose: the point is to compare steps within one process, not to produce a number for a board. Steps with high volume, clear rules and a low cost of error go first. Steps where the rules live in one person’s head go last, or go back to the documentation stage.

What AI is good at, and what it is not

Not every automated step needs AI. Split each step into three kinds of work.

Kind of workUseExamples
Reading and interpreting messy inputAIReading supplier invoices and delivery notes; classifying incoming emails as order, query or complaint; extracting fields from a PDF order; drafting a reply for a person to check; matching records that do not share an exact reference
Applying fixed rules to structured dataPlain rules, or your existing systemIs the price within tolerance? Is the credit limit exceeded? Route to the right queue; create the record in NetSuite or SAP
Decisions with real consequencesA person, with the AI’s work in front of themApproving a discount, accepting a substitute, releasing a payment, answering an unhappy key account

A good automated step usually combines all three. Take order entry from email: AI reads the email and the attached PDF and extracts the customer, lines, quantities and requested date. Rules check that the customer exists, the prices match the price list and the credit limit is fine. A person reviews the draft order and approves it. Cases from the exception log, such as a missing PO or an unknown product code, are flagged with the reason rather than guessed.

Do not use AI where a rule will do. If the decision is “price differs by more than 2%: hold” (an example tolerance), write it as a rule. It is cheaper, faster and easier to audit.

Start with one step and keep a human approval

  1. Pick one step from the top of your scoring. Not a whole process.
  2. Define “done correctly” for that step, using the SOP. That definition is what you test against.
  3. Run it in parallel first. The automation prepares the output, the person does the step as usual, and you compare the two. Look at every disagreement.
  4. Then let the person approve instead of doing. The automation prepares, the person checks and approves. Stay here for as long as errors are still turning up.
  5. Give every exception a route. Each logged exception either gets a rule, goes to a named person with the reason, or stays manual. “The AI will work it out” is not one of the options.
  6. Only then consider removing the approval, and only for cases where approvals have been routine for a long stretch and the cost of an error is low. Many steps should keep a person approving for good.

Measure before and after

You cannot tell whether automation helped unless you measured first. The counts from mapping the process are your baseline. For the step you automate, record these before you start and again after a few weeks:

  • Time from trigger to done, including waits, not just work time
  • Number of hand-offs and systems touched
  • How often a person had to correct the output
  • Share of cases that went through without an exception
  • The revenue-side effect, where there is one: quotes answered the same day, orders confirmed sooner

If the numbers do not move, do not add a second step. Find out why first.

Your systems stay the source of truth

Automation should read from and write back to the systems you already use (your ERP, accounting package, CRM and inbox) rather than become another place where data lives. The sales order still lives in NetSuite, the invoice in Xero or QuickBooks, the customer in HubSpot. If the automation were switched off tomorrow, the records would be where they always were and the team could carry on by hand.

This also keeps the documentation useful. When a field is renamed or a price list changes, you update the SOP and the automation together, because they describe the same steps.

Checklist before you automate a step

  • The process is mapped end to end, with owners, systems, hand-offs and waits.
  • The step has a written SOP and an exception log.
  • You have before figures for the step.
  • A named person approves the output.
  • Every logged exception has a route.
  • Results are written back to your existing system.
  • Staff know which AI tools they may use and with what data (see the AI acceptable use policy template).

To check the wider picture before you start, the AI readiness assessment scores the work, the data, the people and the risk in twelve questions.

And for an outside-in view, the free Bonofici grader reads your company’s public pages and shows which processes are worth automating first.

Start with what your team already knows.

Book a 30‑minute call.We come back with the first three tools worth building, in writing.

Not ready for a call?

See where your company stands and which jobs are worth the most, from your public pages.