Start with the process, not the technology
Before choosing an artificial intelligence model, understand the work it will support. Process mapping should describe what actually happens: activities, people, tools, decisions and rework loops. Bucap’s guide to process automation similarly places analysis of existing workflows before the evaluation of solutions.
The practical objective is a verifiable map and a process record. Only then should you decide whether to use AI, conventional automation or an organisational change.
Define boundaries, outcomes and ownership
Choose a bounded process rather than an entire department. For example, consider the journey from receiving a purchase request to approving or rejecting it. This is an illustrative example, not a client case study.
Record the following:
- Starting event: what triggers the work.
- Final outcome: which state confirms completion.
- Scope: which activities are included and excluded.
- Owner: who approves the map and authorises changes.
- Observed problem: where delays, errors or duplication are suspected.
Avoid objectives such as “make the department intelligent”. Instead, formulate a testable question: how long does a complete request take to reach a decision, and where does it wait?
Reconstruct actual work with the people doing it
Involve operators, the process owner and those responsible for the systems used. Ask them to reconstruct completed cases, including routine requests, incomplete submissions and unusual situations. Compare interviews with available documents, messages and records, using only authorised materials.
Draw separate lanes for the roles involved. Within each lane, describe activities using a verb and an object: “check availability”, “request missing information”, “approve purchase”. Connect steps with arrows and show decisions together with the conditions governing each path.
Include personal spreadsheets, email exchanges and informal approvals. ZeroDivision’s guide for SMEs recommends documenting actual times, people involved, problem areas and repetitive activities.
Create a record for each activity
Attach enough information to each step to support a discussion about automation:
- Input: the document, message or data required.
- Output: the result produced and its recipient.
- Operator and system: who performs the work and where.
- Rule: the criterion used to proceed.
- Exception: a condition that interrupts the normal path.
- Evidence: the record that allows execution to be checked.
In the purchase request example, the completeness check might receive a form and return either a “complete” status or a request for further information. Specify which fields are mandatory and who resolves missing information.
If a rule differs between departments, show the variants. Do not conceal them behind a generic activity labelled “check”.
Establish the baseline
Separate active handling time from waiting time. Also record cases received, completed, reopened and returned for additional information, specifying the observation period and definitions. When data is missing, mark it as unavailable: an estimate should not appear to be a measurement.
Where suitable logs exist, consider process mining. Appian describes this technique as using system logs to understand workflow performance and potential bottlenecks.
Check that events can be linked to the same case and that timestamps are consistent. Then compare the digital reconstruction with operators’ accounts, recording any unlogged steps separately.
Separate simplification, rules and AI
For each activity, first consider whether to remove or simplify it. A duplicate approval without a clear purpose requires an organisational decision, not necessarily a model.
Then distinguish three options:
- Rule-based automation: mandatory field checks, notifications and predefined assignments.
- AI assistance: classifying free text or extracting information from documents with varying formats.
- Human intervention: sensitive decisions, ambiguity and exceptions to policy.
AI document processing and automation of subsequent operations are distinct components, as Appian illustrates through IDP and RPA.
In our example, assess AI for reading an unstructured attachment, while keeping approval rules and responsibilities explicit.
Design controls and a pilot
Create a second map of the future process without overwriting the current one. Highlight where AI operates, which data it receives and who checks its output. For personal or confidential data, involve privacy and security specialists before selecting tools and providers; document intended access, retention and transfers.
Define what happens when a document is unreadable, a system is unavailable or an output is wrong. Provide a review queue and an alternative manual procedure.
For the pilot, establish test cases, acceptance criteria, an evaluation owner and stopping conditions. Compare quality, total processing time and review effort against the baseline. Bucap recommends testing and monitoring after implementation.
Connect the map to decisions
Business process mapping represents workflows, roles and friction points. Automation opportunity analysis identifies repetitive activities that can be delegated to automated workflows. An AI adoption strategy connects opportunities with usage rules and risk oversight.
The mapping exercise should also support a useful conclusion: this particular step should not be automated yet.
Next step
Bring this process into your company
Tell us where the process slows down. We will identify the most useful starting point.