Which business process should you automate first?

Start with a business problem, not an AI tool

Describe the problem without naming a product. For example: we spend three hours a week copying approved enquiry details into a second system, and some records are missed. That is specific enough to investigate. We need AI in our sales process is not.

A useful first automation has a clear beginning and end. It might start when a valid enquiry arrives and finish when a CRM record and a follow-up task exist. Automating the entire sales journey at once introduces more dependencies, more exceptions and a much harder test of whether anything has improved.

Also ask whether the task needs to exist. A report nobody uses should probably be stopped, not automated. A form that asks clients to repeat information you already hold may need simplifying before it needs connecting to anything.

1. Make a shortlist from a normal working week

Ask the people doing the work to record recurring tasks over a representative period. Note how often each task happens, active handling time, common mistakes and what causes delays. Include busy periods if your business is seasonal; one unusually quiet week can make an important task look insignificant.

Separate time spent working from time spent waiting. A client taking five days to approve a document is not five days of staff effort. Automation may improve visibility or prompt a response, but it cannot guarantee that an external decision happens sooner.

• Copying the same information between a form, spreadsheet and CRM
• Creating routine tasks after an enquiry, accepted quote or completed job
• Checking for missing documents and preparing a reminder for review
• Combining consistent data into a recurring internal report
• Sending an approved status update when a verified milestone is reached

2. Check whether the process is ready to automate

Before estimating savings, ask whether two trained team members would follow roughly the same steps for the same input. If not, clarify the process. Software cannot reliably apply a rule the business has not agreed.

Look at the information entering the workflow. Structured fields, stable identifiers and clear statuses are easier to work with than inconsistent notes spread across inboxes. Poor source data can create duplicate contacts, wrong reminders and reports that look precise but are not dependable.

You do not need perfect data everywhere. You do need a rule for missing, conflicting or ambiguous inputs. A sensible first workflow can put those records into a review queue instead of guessing.

• A clear trigger starts the work, such as an approved form submission
• The required inputs exist and their meaning is understood
• Routine steps can be written as a short checklist
• Someone owns exceptions and can pause the workflow
• The outcome can be checked independently of the automation saying it succeeded

3. Use a simple automation-priority score

Score each candidate from 1 to 3 against the five criteria below. Higher is better. This is a discussion aid, not a validated industry model or a substitute for financial analysis. Use observed workload wherever possible and write down why you gave each score.

Add the scores for an initial comparison, but do not let a high total override a serious risk. A process handling sensitive information, changing payment details or making contractual commitments needs stronger controls even if it looks attractive on time savings.

• Measured workload: 1 means occasional or low effort; 2 means a recurring burden; 3 means substantial, consistently observed effort
• Rule clarity: 1 means frequent judgement; 2 means a mix of rules and review; 3 means stable, agreed steps
• Data readiness: 1 means incomplete or inconsistent inputs; 2 means some cleanup is needed; 3 means reliable, accessible inputs
• Connection simplicity: 1 means several fragile workarounds; 2 means some integration work; 3 means supported connections or built-in features
• Safe recovery: 1 means errors are hard to reverse; 2 means mistakes need careful correction; 3 means errors are easy to detect and contain

A worked comparison: follow-up tasks or AI-written proposals?

Imagine a fictional service business comparing two options. Creating CRM follow-up tasks from a structured enquiry form scores 3 for workload, 3 for clear rules, 3 for data readiness, 2 for connection simplicity and 3 for safe recovery: 14 out of 15. The owner can check the queue before anyone contacts a customer.

Generating and sending bespoke proposals without review scores 2 for workload, 1 for clear rules, 2 for data readiness, 2 for connection simplicity and 1 for safe recovery: 8 out of 15. A wrong price or promise could create a much bigger problem than a missed internal task.

These scores are illustrative, not evidence about a particular business. The useful conclusion is to start with the predictable internal workflow. Proposal drafting might still become a later project, with approved source information and a person reviewing every proposal before it is sent.

4. Calculate net savings, not just minutes removed

Estimate gross handling time saved as the number of cases multiplied by the minutes removed per case. Then subtract the time people will spend reviewing outputs, resolving exceptions and maintaining the workflow. Include recurring software costs separately.

For example, suppose 80 records a week currently take six minutes each to transfer and check: eight hours in total. If an automation removes that handling but needs two hours of weekly review and support, the net time released is six hours. At an illustrative staff-cost assumption of £25 an hour, that is £150 a week of capacity.

If the additional software costs £30 a week, the illustrative net benefit is £120 a week. A £1,800 setup cost would have a simple payback of 15 weeks at that steady workload. This excludes other costs unless you add them, including training, migration, taxes and any extra support. It is a worked calculation, not an AVS quote or a promised return.

Released capacity is not automatically cash in the bank. Payroll may stay the same. Explain whether the time will be used to serve more clients, reduce overtime, improve response times or avoid a future hire. Also test the calculation with lower volumes and higher review time before committing.

5. Decide whether you need automation, AI or a simpler process

Traditional automation follows explicit rules: when an approved event occurs, create a task, transfer fields or send a fixed template. AI is more relevant when an input needs interpretation, such as extracting a draft summary from a long email or suggesting a category for an unstructured request.

Do not add AI to a task that an ordinary field mapping or rule can handle reliably. AI-generated output can be incomplete or incorrect, so account for review time and a safe fallback. An AI Assistant that prepares a draft for approval is a different risk from one that sends it directly to a customer.

Sometimes the best answer is neither. Removing an unnecessary approval, agreeing a standard form or setting a clear owner may solve the problem without new software. Make that comparison before pricing a build.

6. Pilot one workflow with clear success criteria

Write down the current baseline, the proposed improvement and the conditions for stopping the pilot. Start with internal test records, then a small live group where a person checks the results. A fixed pilot period is useful only if it includes enough representative cases to expose the important exceptions.

For an enquiry-to-task workflow, success might mean every eligible enquiry produces exactly one task, all required fields are correct, the right owner is assigned and failures are visible. Compare active handling time as well. A workflow can be accurate yet still create more review work than it removes.

Keep the original records and a controlled manual route available during the pilot. Avoid running two unrestricted processes that both contact customers or create invoices. Define which system is authoritative and how the team should resume work if the automation is paused.

• Test a duplicate event and confirm it does not create duplicate work
• Test missing information and check that it reaches a review queue
• Disconnect an integration and confirm the owner sees the failure
• Retry a failed step without repeating actions that already succeeded
• Check access permissions and verify that unrelated customer information stays private
• Confirm that a changed or cancelled record stops any pending actions

What should you avoid automating first?

Be cautious with workflows where a mistake has serious financial, legal or personal consequences. Examples include changing bank details, approving refunds, accepting contract variations, making employment decisions or sharing sensitive client records. These are not good experiments simply because they take time.

Also avoid processes that are being redesigned every week, depend on unreliable data or have no clear owner. A low-volume task with complex exceptions may cost more to maintain than it saves. Stabilise the process or automate a smaller supporting step first.

Human approval is useful only if the reviewer has enough information and time to make a real decision. A button that staff routinely click without checking does not provide meaningful protection.

Do you need custom software for your first automation?

Often, no. Check your existing CRM, accounting and project tools for built-in rules or supported integrations. If one product can meet the requirement with appropriate permissions, logging and recovery, configuring it may be the simplest starting point.

Custom software becomes more relevant when important information spans several systems, your workflow has genuinely specialist rules or repeated workarounds are costly. Compare setup, ongoing subscriptions, support, data export and the ease of changing the process later. A cheaper initial build is not necessarily a cheaper system to run.

Keep the scope measurable whichever route you choose. Ask what will trigger the workflow, how success will be checked, what happens when a connection fails and who owns maintenance after launch.

Your first-automation decision checklist

Before committing, finish this sentence: we are automating this specific step because it currently creates this measurable burden, and we will know it works when these checks pass. If the answer remains vague, spend more time understanding the work before buying a tool.

The best first automation is not necessarily the largest opportunity. It is a useful improvement with clear rules, manageable risk and a team ready to own it. Once that works reliably, use what you learn to choose the next process.

• We have observed the workload rather than guessed it
• We have agreed the trigger, inputs, owner and expected outcome
• The expected benefit includes review, maintenance and software costs
• We can detect errors, pause the workflow and recover safely
• We have chosen a small pilot and defined what would justify expanding it