The question we are asked most often is some version of “should we be using AI for this”. It is the wrong first question. It starts with the tool and goes looking for a problem, and it tends to end with a demo that looks impressive on its own and changes nothing about how the business runs.
The better first question is “what kind of work is each step”. Once a workflow is written down as steps, the right answer for each one is usually obvious, and AI turns out to be the answer for one or two of them. This is the method we use in every Learn & Recommend engagement, and it is simple enough to run on a whiteboard this week.
Start with the steps, not the tool
Pick one workflow that annoys people. Write it down as a numbered list of steps, from the moment something arrives to the moment the business has finished with it. For each step note three things: what comes in, whether anyone has to decide something, and what goes out.
Be honest about the exceptions. The step where “someone checks it” usually hides a decision. The step where “we update the sheet” usually hides a copy-and-paste between two tools. Most workflows have eight to fifteen steps once the hidden ones are written down, and the writing alone tends to settle half the argument about where AI belongs.
Six possible answers for any step
Every step gets exactly one of six answers. The test for each is short.
Connect the tools. The information already exists in another system and someone is retyping it. A booking that lives in email and gets copied into a job tracker. A lead that lives in a form and gets copied into a spreadsheet. Nothing needs to be decided. The two tools need to talk.
Automate it. The step runs the same way every time and a rule can make the decision. Send the follow-up three days after no reply. Create the folders when the job is created. Check for replies every two hours. If you can write the rule in one sentence, it is automation.
Use AI. The input is messy or written differently every time, the step needs judgment inside clear limits, or the right action depends on context a rule cannot hold. Reading an email and pulling out the client and the address. Writing a two-line note that fits this particular person. Scoring an idea against a brief.
Leave it human. The decision has consequences the business owns, depends on a relationship, or is rare enough that building anything for it would cost more than doing it. The conversation after a warm reply. The pricing call. The shoot and the edit.
Build focused software. The work does not fit any tool you have, and the workaround is a spreadsheet with forty tabs or a shared inbox with a colour code. A small internal tool built for exactly this job beats bending a product that was built for something else.
Simplify the process. The step exists because of something that happened in 2021. Remove it. This is the answer people forget, and it is often the most valuable one.
Notice that only one of the six involves AI.
A worked example
Here is a made-up request moving through a small services firm. The IDs are fictional and the firm does not exist; the shape is one we see constantly.
| Step | What happens today | Answer | Why |
|---|---|---|---|
| 01 | Request 0418 arrives by email and someone copies it into the tracker | Connect | The information already exists. Nothing to decide |
| 02 | Someone reads it to work out what is being asked and how urgent it is | Use AI | Every request is written differently. A rule cannot read it |
| 03 | It goes to the right person | Automate | Once the category is known, the routing is a rule |
| 04 | That person decides whether to quote, decline or ask for more | Leave it human | A judgment with consequences the firm owns |
| 05 | The work is tracked across three spreadsheets | Build focused software | The job does not fit the tools; the spreadsheets are the symptom |
| 06 | Five people get five separate status emails | Simplify | One update, once a day, to one channel |
One AI step out of six. That ratio is typical. In the systems we have built for clients the count is usually zero or one, and we have never built one where AI was the right answer for more than a third of the steps.
Two real systems that taught us the method
Gizoom Consulting: zero AI steps. Paul needed a steady stream of new conversations, and each campaign was the same slog: find the companies, find the right people and their emails, make sure the addresses work, send, follow up, keep track. Every one of those steps had a rule. The system we built pulls leads from a search, verifies every address, keeps one clean list so nobody is contacted twice, sends from warmed-up mailboxes and follows up until a reply comes in. There is no AI in it, on purpose. What the business needed was for it to run every day without anyone remembering to start it. Paul’s time goes into the conversations, which is the one step that is his.
Avision Digital: one AI step out of six. Bookings for real estate shoots arrived by email and no two looked the same. Someone read each one, made the folders, created a share link, opened the job in ClickUp with its tasks and told the team. The system uses AI for exactly one step: reading the booking and pulling out the client and the property. Everything after that is plain automation, because everything after that is the same set-up every time. The shoot and the edit stay with the team, because that is the work.
Automated Investments sits in between: AI writes one personal note per prospect, because a template reads like a template, and a rule stops all automation for anyone who replies. One AI step, one human step, the rest is plumbing.
Seven questions before you accept any AI step
A demo answers none of these. An operation depends on all of them.
- Where does it sit in the work? Which step, with what coming in and what going out. If the answer is “it is a chatbot”, it does not sit anywhere yet.
- What information can it reach? An AI step is only as good as the files and records it can read. If the brand document is in someone’s head, write it down first.
- How reliable is it on a bad day? Not the average day. The day the input is garbled, the API is slow and the person who normally checks is on leave.
- When do people stay involved? Name the step where a person reviews, and name the person.
- What happens when something goes wrong? Where does the failure show up, who sees it, and what do they do.
- Do people use it, day to day? A system the team routes around is a system that failed, however good the demo was.
- Does the operation run better because of it? Fewer hours, fewer errors, faster turnaround, measured against how it ran before.
If a proposed AI step cannot answer the first five before it is built, it is not ready to be built.
Signs you are about to over-build
The tool was chosen before the workflow was written down. More than a third of the steps are marked AI. No step is marked human. Nobody can say what happens when it fails. The pitch mentions “agents” and the workflow has three steps.
Any one of these is worth a pause. Two together usually mean the project should start again from the step list.
What to do this week
Write one workflow down as steps. Give each step one of the six answers. Count the AI steps. If the count is one or two, you have a project worth scoping. If it is zero, you may have a better project than you expected, as Gizoom did. If it is five, the list is probably hiding the real steps.
If you would rather have us do this across the whole operation, that is what Learn & Recommend is: one to two weeks, every important workflow mapped and marked, and a written recommendation you keep whether or not anything gets built.