Business process automation: where to start
The safest first project does not cover the entire company. It chooses one frequent process, a measurable loss, a small user group and a manual completion route when automation fails.
Choose a problem, not a technology
“We want to implement AI” is not yet a business problem. A better starting point is concrete: sales staff copy forms into a CRM, proposals wait three days for approval, only one person knows a case status or the team searches dozens of documents for an answer.
A process suits a first pilot when it is frequent, has an owner, can be followed end to end and produces a measurable change. It should not initially cover the entire company or a critical operation without an emergency procedure.
- enough frequency to collect evidence
- a clear beginning, end and process owner
- a current cost in time, errors or delay that can be compared
Follow one real case step by step
A workshop should not rely only on how the process is supposed to work. Observe a real form, email, spreadsheet, approval and exception. Workarounds and manual corrections usually determine the actual scope.
Record inputs, owner, decision, output and system at every step. Mark copying, waiting, missing status and repeated checks. Only then is it clear what should be simplified before automation.
Use rules and integrations before AI
When a decision is deterministic, a standard rule is cheaper and more predictable than a generative model. A webhook can send a form to a CRM, validation can check completeness and a schedule can create a reminder.
AI becomes useful for text, documents, classification and ambiguous data. In that case define sources, allowed actions, answer quality, permissions, logging and human approvals. NIST frames AI risk management across the system lifecycle rather than around model choice alone.
Design a pilot with a fallback
A pilot should support a small group, a limited case type and a clear measure. Recommendation mode is often useful: the system proposes an action while a person approves it.
Every failure needs a visible status, log and manual route. If an integration times out or the model cannot find a source, the case must not disappear. It should enter a queue with an explanation.
- one user group and one case type
- human approval at the beginning
- an alert, log and manual completion path
Measure before scaling
Compare process time, manual steps, corrections, ownerless cases and exceptions requiring people. It is not enough to say the team uses AI. A tool can be opened frequently without shortening the process.
Only after several cycles should the company add case types, integrations or permissions. Stopping or changing a pilot that creates no value is a valid result, not a failure.
Frequently asked questions
What process is best for a first pilot?
Does automation always need custom software?
When should AI be used?
What is a useful success measure?
Sources and further reading
Show us one process held together by memory, spreadsheets and copying. We will separate the problems, quick integrations, pilot scope and decisions that should remain human.
Process card: steps, roles, data, exceptions and success measure
Continue reading
If "Business process automation: where to start" is relevant to your project, these guides may help with the next decision.
How much does an online restaurant ordering system cost?
Cost and scope of online restaurant ordering: MVP, menus, carts, payments, delivery, operations panel, integrations, maintenance and risk.
Education-platform MVP: what the first version should include
Education-platform MVP scope: users, roles, content, quizzes, results, progress, administration, payments, testing and growth plan.
Construction company websites: 12 elements that help win better enquiries
A practical website checklist for contractors and B2B manufacturers: offer structure, projects, catalogues, local SEO and quote-request journeys.