Most businesses do not lack ideas for AI automation. They have too many, and no good way to tell which one is worth doing first. This guide is a method for that choice: list the candidates, score each on five plain questions, pick one, and measure it before and after so you know whether it worked.
Why use a method instead of a list?
A list of ideas, such as 25 business processes you can automate with AI, shows what is possible. It cannot tell you which of your own processes is a good fit, because that depends on how your team actually works. Two companies can both "answer customer emails" and have very different tasks underneath: one has a handful of repeat questions, the other has every message different.
A method also protects you from the loudest voice. Without one, the first project is often whatever the most confident person in the room finds exciting, or whatever the demo made look easy. Scoring forces the question onto evidence you can check.
How do you build a list of candidate processes?
Start by collecting, not judging. Ask the people who do the work, because they know where the time goes and managers usually do not. Useful prompts:
- What do you copy from one place and paste into another?
- What do you read every day just to decide where it goes?
- Which task do you put off because it is dull but must be done?
- Where does work wait for someone to type something in?
- Which report do you rebuild by hand every week or month?
Walk through a normal week of the inbox, the shared drive and the spreadsheets people keep on the side. Anything that moves information from one form to another is a candidate. Aim for a rough list of ten to twenty items; you will cut it down in a moment, so do not polish it.
What should you write down about each candidate?
One short card per process is enough. If you cannot fill in a line, that gap is itself useful information, because it is the part nobody has examined yet.
- What starts it (an email arrives, a form is submitted, a date passes)
- The steps, in order, including the exceptions people handle without thinking
- The inputs it reads and where they live
- The output it produces and where that goes next
- Who checks the result today, and how
- Roughly how often it happens and how long one case takes
How do you score each process?
Give each candidate a score from one (poor fit) to five (strong fit) on each of five questions. Keep the scale coarse. The aim is to separate the clear winners from the clear non-starters, not to produce a decimal.
- Frequency. How often does it happen? Daily work gives a system many chances to prove itself and many chances to repay the effort. Yearly work does neither.
- Volume. How much total effort does it take across the team? A small task done by many people can add up to more than a large task done by one.
- Rules clarity. Could you explain what a correct result looks like, including the usual exceptions? If two experienced people would disagree, the system will be unsure too.
- Data quality. Are the inputs available, readable and reasonably consistent? Scanned pages, missing fields and information kept only in someone's head all lower the score.
- Risk. How costly is a mistake, and how easily is it caught and undone? Score this one the other way round: five means a mistake is cheap and easy to catch, one means it is costly or permanent.
Score as a small group, with at least one person who does the work. Where you disagree by more than a point, talk about why, because the disagreement usually points to an exception someone had forgotten.
| Candidate | Frequency | Volume | Rules clarity | Data quality | Risk (5 = low) | Total |
|---|---|---|---|---|---|---|
| Sorting the shared inbox | 5 | 4 | 4 | 4 | 5 | 22 |
| Pulling fields from supplier invoices | 4 | 3 | 4 | 3 | 3 | 17 |
| Drafting replies to unusual complaints | 3 | 2 | 1 | 3 | 2 | 11 |
| Building the yearly compliance summary | 1 | 3 | 3 | 2 | 2 | 11 |
The total is a guide, not a verdict. Read the individual scores as well. A one on rules clarity or data quality is a warning that the process needs fixing before any software can help. A low score on risk (a costly or permanent mistake) does not rule a process out, but it tells you the result needs a person to approve it before it goes anywhere.
Does the process need AI at all?
Sometimes the honest answer is no. If the steps are fixed and the inputs are always in the same shape, ordinary automation, such as a scheduled job or a rule in your existing software, is simpler, cheaper to run and easier to predict. AI earns its place where the input varies: free-text emails, documents with different layouts, requests written in a customer's own words.
A useful test: if you could write every rule on one page and no case would surprise you, you probably do not need a language model. If the work is mostly reading something messy and deciding what it means, you might. Many good designs use both, with fixed steps for the parts that must behave the same way every time. What an AI agent is explains how those pieces fit together.
How do you pick the first one?
Take the highest total, then apply four tie-breakers. They matter more than most people expect.
- A clear owner. One person who understands the process and will tell you when the output is wrong.
- One main system. Work that touches one mailbox or one application is simpler to connect than work spread across six.
- A small first slice. Automate the common case and send everything else to a person.
- Something you can measure. If you cannot say how you will tell it worked, choose another candidate.
Resist picking the most impressive candidate. A modest first project that works and can be measured builds the trust and the know-how for harder ones. It also gives you real figures for what AI automation costs in your own business rather than someone else's.
How do you measure before and after?
Measure before you change anything. Once the automation exists you cannot recover the original figures, and memory is a poor substitute. Take a sample of real cases, large enough that one odd week does not distort it, and record the same things you will record later:
- Time per case, from the moment it arrives to the moment it is done
- How many cases arrive in a typical period
- How often a result has to be corrected or redone
- How long cases wait before anyone starts them
After launch, run the new process beside the old one for a while on real cases and compare like with like. Add one more figure: the share of automated results a person had to change. A high correction rate eats into the saving. Before reaching for a bigger model, try tighter instructions or a smaller scope.
Only report what you actually measured. If the sample was small, say so. A modest, honest figure is more useful when you decide whether to expand than a confident guess.
What if nothing scores well?
That is a finding, not a failure. Low scores on rules clarity and data quality usually mean the process itself needs attention first: write down the steps, agree what a correct result is, and move the inputs somewhere consistent. This work pays off even if you never automate anything. Alternatively, start with a task that only reads and retrieves, such as searching your own documents, where a wrong answer is easy to spot and nothing is changed.
What is the next step?
Pick your top candidate, fill in its card and write down the figures you will measure. If you would like a second opinion on the scores, the workflow, the technical dependencies and whether AI is the right tool, you can request an assessment and we will go through it with you. When you have a candidate, see how our AI automation services start with a short assessment.