Start automation with a frequent, understandable process that has a clear owner and a manageable failure cost. Measure the current workload, automate a bounded part and compare the result. Do not begin by giving an AI system broad authority over an unclear process.
Rules-based automation and AI solve different problems. Moving a confirmed record between systems may need ordinary integration. Interpreting variable documents may benefit from AI, provided uncertain results can be reviewed and mistakes can be contained.
Find the right pilot
Ask staff where they repeatedly copy information, chase updates or categorise predictable requests. Select a process whose inputs and outcomes can be described clearly.
A suitable pilot might prepare a draft summary from an enquiry or move approved information into another system. A poor first pilot might make high-impact decisions with no review, unclear policy or no way to reverse errors.
Identify the owner before implementation. Someone must decide what counts as correct, review failures and authorise changes. Without that ownership, an automation can quietly become another unreliable system.
Map the existing process
Record the trigger, inputs, actions, handoffs and exceptions. Include the work people do mentally but never write down, such as recognising a familiar customer’s unusual request or checking an ambiguous reference.
Measure a representative period. A quiet afternoon may not reflect month-end volume or seasonal peaks. Capture both processing time and rework, and distinguish elapsed waiting time from staff effort.
The mapping exercise may reveal a simpler improvement: remove a duplicate approval, standardise a form or stop collecting an unnecessary field. Automation is not always the first or best change.
Choose rules, AI or a combination
| Task | Starting approach | Reason |
|---|---|---|
| Route a request using a selected department | Rules | The input already states the decision |
| Transfer an approved order | API integration | The action is structured and repeatable |
| Suggest a category from free text | AI with review where needed | The input is variable and ambiguous |
| Approve a consequential action | Defined policy and appropriate human control | Mistakes may have material consequences |
A combined workflow can use AI to propose a structured result and ordinary rules to validate required fields. Human review then handles ambiguity or actions beyond a defined threshold.
Do not treat a confidence score as proof of correctness. Evaluate the actual task with representative examples, including unusual and adversarial input where relevant.
A worked enquiry example
Imagine a fictional business receiving project enquiries by email. The pilot extracts a proposed service category, a short summary and any stated deadline. It creates a draft for a colleague to review before information enters the CRM.
The system does not send a quotation or promise a delivery date. It flags missing information and preserves a link to the original message. If extraction fails, the message remains in the ordinary review queue.
This scope is intentionally limited. It tests whether preparation helps staff without transferring final responsibility for customer commitments to the automation.
Calculate value with honest assumptions
Suppose the process handles 250 items a month. Current preparation takes six minutes per item. The pilot leaves two minutes of review per item, while 20 exceptions need an additional five minutes each.
The arithmetic is:
- Current effort: 250 × 6 = 1,500 minutes, or 25 hours.
- Routine review: 250 × 2 = 500 minutes.
- Additional exception work: 20 × 5 = 100 minutes.
- Remaining effort: 600 minutes, or 10 hours.
- Illustrative capacity released: 15 hours a month.
At an assumed internal loaded cost of £22 an hour, that represents £330 of staff capacity. If tools and monitoring cost £90 a month, the illustrative net capacity value is £240 before implementation and change-management costs.
These are invented planning inputs, not measured Flowzelo results. Released time does not automatically become cash savings. If implementation cost £2,400, dividing by £240 suggests ten months of simple payback only if the assumptions hold; it excludes financing, tax and changing volumes.
Protect data and limit authority
Send only the information needed for the task. Decide whether personal data or confidential documents are suitable for the selected provider and configuration. Review contractual terms, retention and international processing arrangements before using real records.
The ICO’s AI data-minimisation guidance is relevant when designing what information an AI system can access. Its AI governance guidance provides further context for responsibilities and risk assessment.
Restrict connected accounts to necessary actions. A summarisation workflow does not need permission to delete customer records or change payment information. Treat incoming documents as untrusted information, not instructions granting the automation new authority.
Design exception handling and monitoring
Every item needs a visible state: pending, completed, awaiting review or failed. Staff should know what happened and what action is safe next.
Use bounded retries and duplicate prevention for connected systems. Keep a clear route back to manual processing. A pilot should be easy to pause without losing the original work.
Monitor error categories, review effort and unresolved items. Rising exception rates may indicate a changed document format, a provider update or a poor fit between the task and the automation.
Evaluate before expanding
Run the pilot against a baseline and inspect errors, not only successful examples. Separate quality failures from delays and configuration problems. Ask staff whether the new process is genuinely easier to operate.
Expand only when benefits survive realistic review and exception costs. Add one meaningful capability at a time and repeat the assessment after major changes.
For a bounded pilot with a clear owner and measurable outcomes, explore Flowzelo’s AI and workflow automation service.
Last reviewed: . Guidance is general; project requirements and platform behaviour should be checked for your circumstances.
Explore the related Flowzelo service →