AI Workflow Automation Examples, Specified Step by Step

What a useful example has to specify

The parent guide catalogues six process shapes that recur across small operating teams. This page takes three of them and builds them out, because the shape is the easy part. Everything expensive lives in the steps.

Each step below carries one of three labels. Fixed code means a rule somebody wrote once: a lookup, a calculation, a write to a system. It costs almost nothing per run and behaves identically every time. Model judgment means the step genuinely cannot be enumerated in advance, so a model reads the input and decides. It costs money and varies on every run, forever. A human approval means somebody has to be answerable for what happens next, and no amount of accumulated confidence changes that.

All three examples pass the test Google's engineers use for which work deserves automating. They call it toil: manual, repetitive, automatable work that produces no enduring value and grows at least as fast as its source. Two of the three grow directly with sales. The third does not, and it is included precisely because that makes it the marginal case.

Example one: routing an inbound lead

A form submission arrives and somebody has to work out whether it is worth a salesperson's time and whose it is. Left manual it is the sort of job that gets done in batches on Friday, which is also why the good leads go cold.

1. 01A form is submitted, or a record appears in the CRMTrigger
   Two sources, one process. The second one matters: leads entered by hand at a conference should not take a different path just because they arrived differently.
2. 02Look up the company and fill in the firmographicsCode
   Domain to company record, headcount, country, existing customer or not. Entirely mechanical, and the largest single block of manual minutes in the process.
3. 03Decide whether this record is the same company as one you already haveAI judgment
   “Acme GmbH”, “ACME Holdings” and a personal email address from a contractor working for them. No rule settles this reliably, which is what makes it a genuine judgment step.
4. 04Map a free-text job title to a segmentAI judgment
   “Head of Ops”, “Operations Lead”, “COO” and “I do a bit of everything” all need to land somewhere. The last one is the interesting case, and the reason a lookup table stops working.
5. 05Apply the routing rules and assign an ownerCode
   Territory, segment, round-robin within a team. Once the two judgment steps above have produced clean values, this is arithmetic.
6. 06Write the record and notify the ownerCode
   No approval here, and that is the point: this process commits the company to nothing. Nobody outside is told anything, and a wrong assignment is fixed by dragging a record.

The exception that breaks it. A large customer submits the form again through a different subsidiary, and the deduplication step confidently merges it into the existing account, hiding a new opportunity inside a closed one. It will happen, it is knowable now, and the fix is a rule that any merge touching an account above a certain value stops and asks. That rule costs nothing to add on day one and is painful to retrofit after somebody has lost a deal to it.

Example two: the recurring client report

Every Monday, figures from three systems become a document that goes to a client. The work is dull, the deadline is immovable, and the person who does it is usually senior enough that it is an expensive way to spend a morning.

1. 01Nine o'clock on MondayTrigger
   A schedule, which is the simplest trigger there is — and the one most likely to fire on a bank holiday when the source data has not landed yet.
2. 02Pull the figures from each systemCode
   Fixed queries against fixed sources. If this step is hard, the problem is your reporting stack, and no automation layer will disguise it.
3. 03Reconcile the numbers against each otherCode
   Totals from two systems should agree within a known tolerance. A mismatch is a rule, not a judgment: if the gap exceeds the tolerance, the run stops rather than reporting a number somebody will later have to retract.
4. 04Write the commentary on what moved and whyAI judgment
   The only genuinely open-ended step in the process. Which movements are worth explaining, which are seasonal noise, and how to say so without overclaiming a cause the data does not support.
5. 05A person reads the commentary and releases itHuman approval
   Permanent, because the document goes to a client with the company's name on it. Approving an assembled report takes two minutes; producing one took a morning. That asymmetry is the entire return on this process.

The exception that breaks it. A source system is late, so the reconciliation passes on partial data and the commentary explains a decline that never happened. The approver, reading a fluent and internally consistent document, agrees with it. This is the failure mode the EU AI Act names in its human oversight requirements, which expect an overseer to stay aware of the tendency to over-rely on a system's output. The practical defence is to make the gate show its working: the approver needs the freshness of each source in front of them, not just the finished prose.

Example three: the supplier renewal review

A contract comes up for renewal and somebody should check whether it still makes sense before it rolls over silently. This is the marginal example of the three, because it happens perhaps forty times a year and does not grow much with the business.

1. 01Ninety days before a renewal dateTrigger
   Calculated from the contract record. If your contract records do not carry reliable renewal dates, that is the actual project, and it is a data problem rather than an automation one.
2. 02Assemble the current positionCode
   What is being paid, against what was budgeted, alongside whatever usage figures exist. Mechanical, and tedious enough that it is often skipped entirely when done by hand.
3. 03Read the terms and summarise what changed since last timeAI judgment
   Contracts have no fixed layout, which is exactly the condition that makes a model worth paying for. The output is a summary with the clauses quoted, not a verdict.
4. 04Renew, renegotiate or cancelHuman approval
   Permanent, and not a close call. The decision commits money and is awkward to reverse, so it stays with a person who is answerable for it — regardless of how good the summary in front of them has become.

The exception that breaks it. Auto-renewal notice periods are quietly different per contract, so a ninety-day trigger is too late for the one supplier that wanted one hundred and twenty. The process runs perfectly and the decision arrives after the window closed. Notice periods belong in the trigger calculation, and finding that out from the contracts rather than from a missed deadline is the whole value of writing the process down first.

What the three have in common

Laid side by side, the pattern is more useful than any individual example.

Two things stand out. The model does one or two jobs in each process, and everything around it is ordinary lookups and writes — which is why a system that calls a model for every step is charging you judgment prices for arithmetic. And the human approvals are not there because the software is immature. They are there because those two steps commit money or reach a client, and that property does not change however many clean runs accumulate behind it.

Each example also needs an owner: one person who would notice next March that the commentary has quietly stopped being right. NIST puts govern at the centre of its four AI risk functions, ahead of measuring and managing, and a process with no named owner has skipped the step that makes the rest meaningful.

Limitations and when not to use this

• All three examples are composites built to show the shape of the work. The step counts are real properties of the processes as described here, not measurements from any organisation, and your version of each will have steps these do not.
• Step counts are a poor proxy for effort. One fixed-code step that reconciles three systems can be more work to build than four straightforward ones, and the table above deliberately does not pretend otherwise.
• Nothing here describes what any particular product does. The split between code, judgment and approval is a property of the process; whether a given tool can express that split is the subject of the tools guide.
• Nothing here is legal, financial or procurement advice. The renewal example in particular touches contract terms, where the details are specific to your agreements and jurisdiction.

Sources

1. Eliminating Toil — Google, The Site Reliability Workbook Accessed 27 July 2026
2. Article 14: Human oversight — EU Artificial Intelligence Act Accessed 27 July 2026
3. AI Risk Management Framework — NIST Accessed 27 July 2026

Related reading

• AI Workflow AutomationThe catalogue these three are drawn from, plus the decision table for whether your own process belongs in the category.
• How To Automate WorkflowsWhat you personally do, week by week, to get one of these examples from manual to running.
• SOP AutomationHow to write your own process down to the level of detail these examples are specified at.
• Accounts Payable AutomationA fourth example taken further still, with a downloadable SOP and workflow blueprint attached.