Pick a process that repeats, write down how it really runs and cut the steps nobody needs. Then choose the lightest tool that will do the job, decide where a person signs off and test it on old cases.

To automate business processes well, work in this order: pick one process that repeats, write down how it actually runs today, remove the steps nobody needs, and only then choose a tool. The tool is usually the easy part. Automation built on a process nobody has looked at closely just does the wrong things faster.

Which processes are worth automating first

The best first candidate is a dull one, such as logging a new enquiry and passing it to the right person, chasing a form that hasn’t come back, or copying an approved order from one system into another.

Four questions sort the candidates quickly:

  • How often does it run? A task done twice a year isn’t worth automating, however tedious it is.
  • Are the steps the same each time? If every case is different, you need a better checklist before you need software.
  • Where does the information arrive? An email, a web form or a record in a system can be picked up automatically. A phone call or a paper form needs a person, or an extra step, first.
  • What does a mistake cost? Start where an error is caught and put right easily, not where it reaches a customer or moves money.

Leave out anything the team is still arguing about. If two people do the same job two different ways, an automation will follow one of them, and the other person will work around it.

Map the process as it really runs

Sit with the person who does the work and watch them do it once, from start to finish. The written procedure, if there is one, may be out of date. Note down:

  • The trigger: what starts it, such as an email arriving, a form being submitted or a date passing.
  • Each step, in order, and the system, spreadsheet or document it happens in.
  • Each decision, and the rule behind it. “The manager approves any order over a set value” is a rule. “It depends” is a conversation to have before you build anything.
  • The exceptions: what happens when information is missing, late or wrong. This is usually where the time goes, and where an automation is most likely to break.
  • The finish line: what a correctly finished case looks like, so you can check the automation against it later.

A page of notes is enough. The test is whether someone else could follow it without asking you a question.

Cut steps before you automate any

With the process on paper, look for steps that exist only because of how the work used to be done: re-typing details that are already in another system, saving the same file in two folders, or a manager approving something they have never once turned down. Steps like these can often simply stop. A step you delete costs nothing to run, and nothing to fix when it breaks.

It is the same order I work in during my monthly AI partner sessions: watch the task done as it is today, take out what isn’t needed, and only then automate what is left.

Choose the lightest tool that will do the job

Once the process is tidy, pick the simplest tool that can run it. From lightest to heaviest, the options look like this.

Rules inside software you already use. Many accounting packages, CRMs and email services have automations of their own, such as filing messages, sending reminders or creating a task when a record changes. If the whole process sits inside one system, start there.

A workflow tool that connects your apps. When the process crosses two or three online services, a workflow tool passes information between them. Zapier calls each workflow a Zap: a trigger, the event that starts it, followed by one or more actions. Microsoft’s cloud flows in Power Automate can start when an event happens, such as an email arriving, at a set time of day, or when someone selects a button. Check what your Microsoft licence includes before you build on it.

Desktop automation for software that can’t be connected. Older desktop systems often have no way in for a workflow tool. Desktop flows are the robotic process automation side of Power Automate, and Microsoft says they can automate legacy applications as well as modern web and desktop ones, working through the application’s on-screen elements. Anything that works through screens is worth testing again whenever the software behind it is updated.

An AI step for anything that needs reading. A fixed rule can’t read an email and decide what it is about, or pull the details out of a PDF that arrives in a different layout each time. An AI step can, with a person checking what it produces until you trust it. The guides to what an AI agent is and how AI reads and checks business documents go into this in more detail.

A bespoke system. Sometimes the process has no proper home at all, and the other tools have nothing to connect to. An automation needs one place to read from and write to, and one clear record of the stage each case has reached. When the work is spread across several spreadsheets, that place doesn’t exist yet, and the first job is to build it. The guide to moving a spreadsheet into a web app covers how I go about that.

An example from my own work

The case management console I designed and built for GK Finance, a mortgage and protection adviser in Northampton, replaced many spreadsheets. It manages customers, enquiries, case stages, documents and notes. GK Finance’s public website, which I also designed and built, is a separate project.

Decide where a person signs off

Every automated process needs a line drawn through it: the steps that run on their own, and the steps a person approves. Put that line before anything that is hard to undo, such as paying money out, sending a customer a decision or deleting records. At Alter Domus, where I was Head of AI & Automation, multi-agent trade settlement ran with a person approving every irreversible step and a full audit trail. A small firm needs the same habit at a smaller scale: a named approver for those steps, and a log of what ran and when. The rest of that record is on the page about AI in regulated financial firms.

Take extra care when the automation decides something about a person. Under the UK GDPR, a significant decision is one that has a legal effect on someone, or a similarly significant effect. If it is taken with no meaningful human involvement, the organisation must have safeguards in place: telling the person about the decision, and letting them make representations, obtain human intervention and contest it. Decisions that rely on health or other special category data are restricted further. The text is in Article 22A, Article 22B and Article 22C of the UK GDPR. This is practical guidance, not legal advice: if your process makes decisions like these, take advice on it before it goes live.

Test it on old cases before it goes live

Before anyone stops doing the work by hand, prove the automation on real examples, including the awkward ones you found when you mapped the process.

  1. Replay past cases. Run a batch of finished cases through it and compare what it does with what actually happened.
  2. Run it alongside the manual process for a short period, with a person checking every output.
  3. Measure with your own numbers. Time the task before and after, and count the cases that still needed a person. That tells you whether to extend it, change it or stop.
  4. Plan for failure. Decide what happens when a connected system is down or an input is missing, and who gets told.
  5. Name an owner. Someone in the business should know what the automation does, where its log is and how to switch it off.

When a workflow needs building and testing properly, that is the work I do in agents and business process automation, and the guide to budgeting for an automation project covers the build and running costs.

You don’t need a finished brief to start. Book a free 30-minute AI consultation, with no obligation, and I will look with you at how your business runs and where AI could save time. If a project makes sense, I will set out the scope and fee in a written proposal.

More insights

All insights