What to automate first in a service business

Inventory the work, not the tools

For a week, record the tasks that repeat, the interruptions, the copying between systems and the waiting. Note who starts each task, what they receive, what they decide and what they deliver. Ask what happens if it is late and which mistake would be hard to repair. That inventory reveals more useful opportunities than a list of fashionable apps.

Group tasks by outcome: answering an enquiry, preparing a quote, confirming a service, coordinating the team or closing the day. Automation can remove steps inside that flow, but it must keep an owner. If nobody answers for the outcome today, the first job is to clarify operations.

A four-criteria matrix

Score each task on frequency, time and friction, availability of information and risk. The best candidates are frequent, visibly painful, backed by reasonably clean data and reversible if wrong. A rare but annoying task may not pay for the build. A frequent, high-risk task may need an assisted mode.

Include variation. If every case is completely different, AI may help prepare but not complete. If eighty percent share a pattern and the rest need judgment, design an automatic route for the common case and a handoff for the rest. Do not force exceptions down a path that does not suit them.

• Frequency
• Minutes and people involved
• Data quality
• Variation and exceptions
• Cost and reversibility of an error

Three common first automations

In customer service, an agent can answer approved questions, collect details and summarise intent. In administration, an automation can turn forms or emails into structured tasks. In knowledge, an assistant can search policies and cite the source. All three start with information that already exists and can run in proposal mode before executing actions.

Scheduling, payments and changes to records come later, once the base interpretation works. Each integration adds surface for failure and for security. Add it because it removes a measured step, not because it appears in a demo. Keep human confirmation when an action commits money, capacity or the relationship with a customer.

Define success and an exit date

Pick twenty to fifty representative cases and a baseline. Define criteria per case: correct data, source, time, handoff and review. Test in parallel, classify failures and set a date to continue, adjust or withdraw. A pilot with no exit condition turns into permanent software by inertia.

Then ask what the business learned. Perhaps the biggest gain was documenting policies or discovering that enquiries arrive incomplete. That learning can justify a second phase different from the one you imagined. Good automation does not just move work; it makes the process more visible and governable.