What is AI automation?
AI automation combines defined automated steps with AI that can process emails or documents, for example, or draft text. Rules, checks, and human decisions remain in place where mistakes could cause harm.
Which workflows can be automated?
A good candidate recurs often, and its correct outcome can be described. Inputs may vary in format, but it must be clear when a person needs to decide.
| Current work | Possible automation | Human review |
|---|---|---|
| Sales enquiries arrive through different channels | Collect, classify, create a CRM record, and assign a task to the owner | Approve unclear or high-value enquiries |
| Fields are copied from invoices or documents | Identify fields, check format, and pass them to the next system | Check missing or contradictory information |
| A weekly report is compiled from several spreadsheets | Collect data, calculate, summarise, and send | Explain exceptional changes |
| Customer emails are routed manually | Identify topic and urgency, draft a reply, and select the right team | Approve the final reply or a sensitive case |
If a decision is based on a specific field or rule, traditional automation is cheaper and easier to predict. AI is suitable when the input is free-form text or a document.
Automating Microsoft 365 workflows
In Outlook, Teams, and SharePoint workflows, incoming mail can be classified automatically, tasks created, files collected in the right library, and approval notifications sent. Power Automate works well for defined steps; AI is worth adding only where free-form content needs to be understood. Critical output—a quote, customer reply, or document awaiting approval—remains subject to human review.
Manufacturing automation
In manufacturing, benefits usually start with manually moving order, work order, and quality-control data. Automation can, for example, pass orders to an ERP, notify about deviations, and prepare shift summaries. Before adding AI, agree on what event triggers the work, what output counts as correct, and who is responsible for exceptions.
Which process is suitable for a first pilot?
Good candidate
The task recurs at least weekly, inputs are available digitally, each step has an owner, and the result can be measured by time, volume, errors, or waiting time.
Fix the process first
Rules keep changing, no one is responsible for the outcome, access to the data is not permitted, or a single error could cause uncontrolled harm.
Four questions before development
- What event starts the work, and where does the input come from?
- What result shows that the work has been completed correctly?
- Which exceptions need a person, and how quickly?
- How much time, waiting, or errors occur now?
How much manual time might a workflow consume?
Enter the approximate volume of one recurring task. The calculator estimates possible time savings. It does not account for development, licensing, review, or maintenance costs.
Formula: number of tasks × minutes × workdays × share that can be automated. The actual result depends on exceptions, review needs, and workflow reliability.
What affects the cost of AI automation?
The price of the work cannot be set based only on the words “AI automation.” It depends on the systems to connect, inputs, checks, and maintenance.
Scope and integrations
One form and spreadsheet are simpler than several CRMs, inboxes, databases, and users with different permissions.
Testing the AI step
Free-form text or documents require a real test sample, quality criteria, and a route for uncertain results.
Reliability
Validation, duplicate protection, logs, alerts, and a recovery plan are part of the solution, not later add-ons.
Running costs and maintenance
AI model usage, infrastructure, licences, monitoring, and maintenance of changing interfaces affect the total cost.
After mapping the process, I document the pilot scope: inputs, outputs, connections, checks, metrics, and expected running costs. That lets you compare the price with the cost of the current work.
Four steps to a verifiable pilot
Process mapping and baseline
I map the current workflow and measure its volume, time, waiting time, and typical errors.
Technical plan
I choose connections, access, rules, AI steps, checkpoints, and error handling.
Limited pilot
I test with a small volume of data or one use case, log the results, and refine exceptions.
Measurement and handover
I compare with the baseline, document administration, and agree on monitoring and maintenance.
Automation must fail visibly and safely
Failures need to be noticed. Before launch, we agree which results the system checks, which it sends to a person, and how it reports a problem.
Minimum necessary access
Each connection receives only the permissions needed for its task. Access and owners are documented.
Data minimisation
The workflow does not send fields that are not needed for the result. Sensitive content needs a separate assessment.
Logs and notifications
Important steps leave an auditable trail, and interruptions reach the designated owner.
A person at decision points
Uncertain, rare, or high-impact cases do not proceed automatically to a final outcome.
What can change after launch?
Automation sometimes needs changes because tools, access, and work rules change. Maintenance needs depend on the number of connections and how important the process is.
An interface or data field changes
An API version, spreadsheet column, or form name can change. Checks must distinguish an empty result from a successful result whose value is zero.
Access expires
A token, password, or user permission needs renewal. An owner and notification route should be set before the connection breaks.
AI step quality shifts
A change to the model, instructions, or input can affect classification and drafts. A standing test sample helps detect changes before expansion.
The process itself changes
If a team adds a manual detour but the automation is not updated, two different versions of the truth emerge. Change ownership must include the workflow too.
The maintenance plan records the owner, failure notifications, log checks, and change testing.
What to compare after the pilot
- Time spent: manual minutes per case and per month.
- Lead time: time from input arriving to the result being ready.
- Quality: share of errors, corrections, and exceptions routed to a person.
- Reliability: successful runs, interruptions, and recovery time.
- Total cost: development, usage, licences, monitoring, and maintenance.
The number of automated tasks alone does not show value. If automation creates more checking or corrections, the workflow needs to change or the trial should stop.
Compare periods of the same kind
The comparison is more useful when workload and case complexity are similar. Also account for time spent on human review, corrections, and resolving interruptions.
The trial summary shows what worked, which cases I sent to a person, and what failed. Based on that, you decide whether to expand, change, or stop the solution.
Questions about AI automation
What is the difference between traditional automation and AI automation?
Traditional automation follows set rules. AI can process emails and documents, for example, or draft a response. Important decisions remain with a person when needed.
How does an AI automation project begin?
I start by mapping one process: its inputs, steps, exceptions, owner, current time spent, and desired result. This helps assess whether automation is worthwhile and what to measure.
Do existing programs need to be replaced?
In most cases, existing programs do not need to be replaced. I connect them through a suitable interface. The options depend on access to the tools in use.
How do you measure the return on AI automation?
Before the trial, I record the workload, manual time, number of errors, and waiting time. Afterwards, I compare the same measures and account for development and usage costs. The calculator provides an estimate, not a promise.
Can automation work entirely without a person?
Not everything should be automated. I route high-impact decisions, uncertain answers, and exceptions to a person. The goal is to reduce unnecessary manual work, not remove necessary oversight.