One of the companies we work with had dozens of recurring procedures written down, spread across several areas, with their steps, their owners and even their accounting codes. Another one had nothing written at all. Both needed the same thing.
The first company's procedures lived in Word, on a file server, and nobody opened them at the moment of doing the work. None of them triggered a task, so the tax filing, the fortnightly reconciliations and the fiscal close depended on somebody remembering the date. The second company sat at the same point, without the documentation work behind it.
We lead with this because it runs against what is usually assumed, that having your processes documented brings a company closer to automating them. Both were missing the same thing: that what they do triggers a task on its date and leaves a record that it happened.
The same holds for artificial intelligence, which almost every mid-sized company has already paid for in some form. Licences for the team, an assistant wired to the documents, a pilot in one area. Each one helps the person using it, and a year later few can point to the line of the income statement where it shows.
A current model answers well when it has the context of the problem, and what it almost never has is how this company buys, who approves a credit note, or why the inventory in a certain warehouse gets counted late. That information sits spread across the ERP, spreadsheets and the head of whoever has been there for years, so without it the assistant works like someone who started on Monday.
Every new pilot rebuilds that context from zero, so the second one costs nearly what the first one did and the third one too. A year in, the company holds three tools with no connection between them and no way to tell what each one contributed.
What we find when we go in to map the operation
The two cases above are not unusual, and alongside them the same three things show up almost every time.
Supervision runs entirely through one manager who, to find out what is behind schedule, has to ask process by process. One or two people appear in most of the critical procedures, so their absence leaves the company with no guide to continue them. And where documentation does exist, the written procedure and the one people actually run drift apart without anyone noticing, which we confirm in every interview because we ask how they do it today and compare that against the document.
The reports management decides with are usually in the same state. Inventory turns, weekly sell-in, accounts receivable ageing, days payable and days receivable get built by hand in a spreadsheet every time they are needed, and the analyst's time goes into assembling them instead of reading them.
A procedure that runs on its own needs four things
Out of those engagements came the short list of what each process was missing in order to hold up without somebody pushing it, and the four turned out to have nothing technological about them.
Before writing a procedure, we take inventory
Mapping without an order produces forty procedures and no priority, so we start somewhere else. Each team lists from memory the activities it performs, without going into how it performs them. For each one they note the average time it takes, the blockers they run into and the problems that repeat. The area coordinates that list and rates every activity in two columns, criticality and complexity.
With that inventory in hand, procedures get mapped by what they move, starting with the ones that affect the income statement and the quality of the work leaving the area. What is frequent but cheap stays in the queue, and that keeps the first weeks off processes nobody misses.
Then comes the interview, one per procedure, with a master prompt of our own so the output comes out consistent across areas and across interviewers. The procedure is born as a draft, with an owner assigned, and gets corrected with whoever runs it before it stands.
Each step of the procedure moves up on its own
The question every manager asks at this point is how much of this AI can do, and the answer changes depending on which step you look at. A monthly close has steps a system solves today and steps that will stay with a person for a long time.
We classify each step into three types. Deterministic ones extract, load, calculate or reconcile, and have a correct answer. Judgment ones analyse, prioritise or decide, and depend on something that is not in a table. Rendering ones produce a document, an email or a report out of information that already exists.
That type sets how far each step can climb, and the move up happens when the step demonstrates something at the level below rather than when somebody decides it is time.
Every change of level is justified by what the one below showed, and management authorises it. What makes measuring it this way useful is that progress shows in indicators of the procedure itself rather than in a general impression: how much time the person spends on it, how long the full cycle takes, how many errors have to be corrected and how much volume the same team handles.
The monthly report says where the money came in
In the diagnostic we agree in writing what the process we are about to touch costs today, and everything that follows is compared against that number. An hour freed up becomes savings when you can point to where the money comes in, and in practice there are three ways:
- A hire that no longer happens.
- More volume handled without hiring anyone else.
- A close that drops in days.
The monthly report says which of the three applies that month, and when none of them applies yet it says that too.
KOJ holds the function for a defined period and hands it back
What a mid-sized company needs here is somebody to run the artificial intelligence function while there is nobody inside who can, and for that somebody to leave once there is. It is the same arrangement used to bring in an interim finance director.
While it lasts, KOJ answers if a procedure did not run. The handover happens during implementation, with the client's team taking procedures one by one, and the contract closes when that team holds them without us.
The three phases are contracted separately and each one produces something that stands on its own, so a company can keep the diagnostic and take it to another firm, or take the diagnostic and the plan and execute with its own team. That way the decision to continue gets made on evidence instead of committed up front.
Where it starts
Before building anything there are two things to measure: where the opportunity is and how ready the company is to take it. The opportunity sits in the decisions and the processes that move the most money, and the readiness depends on the state of the data, on how well written the processes are, and on who inside can take ownership of each one.
That is what the AIM diagnostic measures, and it runs three or four weeks and is contracted separately. It ends with the company's processes ordered by what they move, the matrix of what can run with AI and what cannot, the inventory of available data and the baseline agreed in writing, so a company that decides to stop there keeps a map it can use with whoever it wants.