Service: AI Department Audit →
Free to shareFive stages: from a task map to agent chains. For a manager who hasn't started yet, and for one whose rollout is already running chaotic.
Right now, money gets wasted on AI two ways. First: buy subscriptions and wait for a miracle. Second: cut people before it's clear who was covering the fire drills. This guide is about the third, boring way: count first, decide after.
It's for two kinds of managers. The first hasn't started and doesn't know where to begin. The second is already rolling AI out: tools are bought, mandated from above, but every employee is figuring it out by trial and error. Both need the same sequence of steps. Only the scale differs, covered in Stage 5. There's no list of magic tools here: tools change every six months, the sequence doesn't.
Inside a company, AI works as an amplifier for an employee. It rarely becomes a replacement, and when it does, the result is usually worse than promised. An employee who understands their repeating tasks and hands them to a machine becomes worth more to the company than before. A company that just cuts headcount after buying licenses usually finds out the cost at the first real crisis.
The trend toward mass layoffs under the AI banner is real. Following it isn't mandatory. Below I'll show why the choice should be made on numbers, and how to run them.
The methodology draws on 11+ years of enterprise process analysis in the energy sector (SAP, SCADA, BPMN, UAT) and hands-on AI automation projects since 2023. Author of the ANSS specification standard.
The five stages here are covered at the level of principles. The full version adds instructions for each stage, table templates, and a case study.
Full version →A familiar story. The director watches a webinar, buys a company-wide subscription, sends out an email: "we're using AI now." Three months later, a handful of people actually use it, the skeptics feel vindicated, and the budget gets written off.
The reason is boring. Nobody looked at what people's workday actually consists of. AI is good at taking over specific, repeating tasks. Until those tasks are named, there's nothing for it to take. A tool with no point of application stays a toy.
I've seen cases where even serious dev shops mandated that engineers write code through AI, and nobody explained how to actually work with it. People learn alone, by trial and error. If that's what happens to professional programmers, imagine the sales department.
The conclusion is simple: rollout starts with observation, and the purchase comes after.
For one week, every employee logs their tasks: what they did, how often, how many minutes. Five minutes at the end of the day, a simple table. A week later you have a map of what the department actually does.
| Task | Times/week | Min per time | Total h/week |
|---|---|---|---|
| Management report from CRM | 1 | 180 | 3.0 |
| Replies to routine client emails | 25 | 8 | 3.3 |
| Negotiations with key clients | 4 | 60 | 4.0 |
| Moving requests from email into a table | 15 | 6 | 1.5 |
| Handling non-standard complaints | 2 | 45 | 1.5 |
The three highlighted rows in this example add up to 7.8 hours a week. At 80 zł/hour, that's roughly 2,700 zł a month for one employee, before you even talk about quality and speed.
Spoiler: the map won't match the job descriptions. Tasks you didn't know about will surface. A report someone spends three hours assembling by hand. Emails eating half a day even though it's the same template every time. Duplication, where two people maintain the same spreadsheet.
Now ask three questions about each task on the map. Does it repeat regularly, the same way each time? What does an hour of the person doing it cost? What happens if the machine gets it wrong?
Partial mode is better to start with than full automation: the employee sees where the machine gets it wrong and learns to manage it.
The output of this stage: a short list of tasks with an estimate of freed-up hours. Instead of "we're adopting AI in general," you get three specific places where it pays off.
The most underrated part. AI does what's written. It doesn't do what you meant. A vague brief produces a vague result, and the brief is what's at fault: the machine did its job.
Every task you automate needs a short document: what's the input, what's the output, what are the rules, what must never happen. I call that last part invariants: boundaries the machine never crosses, no matter what. Never send an email without human review. Never change data in the accounting system. Never promise a client a deadline.
Here's what that looks like for "drafting replies to client emails." Input: the client's email and the thread history. Output: a draft reply in the "For Review" folder. Invariants: never send without human approval, never quote prices or deadlines, never answer complaints, which go straight to a human. Five lines, and the machine's behavior is already predictable.
For complex systems, I develop an open standard for these specifications, ANSS. For one office task, a page of text is enough. But that page has to exist.
Roll out one task at a time. The temptation to launch everything at once is strong, and few resist it. The result is always the same: chaos and rollback.
Acceptance is done by whoever used to do the task by hand: they're the best judge of where the machine is cutting corners. Agree on the criteria upfront: what counts as "working." Two weeks of running in parallel: human and machine do the same thing, and the results are compared. Then the human shifts into a supervisory role.
Something good comes out of this too: the employee who went through this cycle starts finding the next tasks to automate on their own. You end up with more than one agent: you get a person who knows how to commission them.
One employee went through the cycle and showed the numbers. Now their experience carries over to the department. This works better than any directive: a colleague with results is more convincing than a consultant.
At the department level, things appear that didn't exist at the individual level: shared specifications, an agent registry, rules for who's allowed to automate what. Sounds bureaucratic. Saves you from a mess.
Only the unit of automation grows over time; the stages stay the same. Level one: a single task, a human checks every result. Level two: a bundle of one employee's tasks, the machine runs the whole process, a human reviews the outcome. Level three: chains of agents, where constant supervision is replaced by checkpoints and the invariants from Stage 3. Advanced teams live at level three, but they get there through the first two. Trying to start at level three gives you chaos that costs more to clean up than climbing the ladder would have.
The audit will show that a ten-person department frees up, say, thirty percent of its capacity. From there a manager has three paths, and the choice should be made on numbers.
The hidden risk has a name: reserve criticality. People in a company do more than what shows up in a spreadsheet. They hold knowledge that's nowhere in the documentation. They cover the fire drills. Cut support down to the bone because "AI handles routine tickets now," and the first major outage will show you the price of that decision: there are no routine tickets during an outage, and the people who knew how to handle the non-routine ones are already gone.
That's why an honest audit gives every role two estimates: how many hours it frees up in normal mode, and what happens in an abnormal one if those people aren't there. Almost nobody makes the second estimate, and that's a mistake: the losses from a single outage can wipe out a year of payroll savings.
The reverse is also true. Caution is itself a risk: a competitor who cuts payroll hard wins on price here and now, and in a mature, non-growing market, scenario A performs worse than it would in a growing one. This section isn't an argument for kindness. Every scenario has a gain and a price, and you need to count both before choosing. The ones who lose are usually the ones who only counted one.
There's one rule, and it's where many rollouts fall apart. If you told the team the goal of adopting AI is to strengthen people, you have to hold that line in your decisions too. One visible layoff "optimized" right after an audit, and there will never be another honest audit at that company. Employees read the gap between words and actions perfectly.
That's actually a purely pragmatic argument for the softer scenarios from the previous section. The team's trust is worth more than a couple of salaries.
You can go through the five stages on your own, and this guide is enough for that. If you'd rather go through the stages yourself but with support, there is AI training on your own tasks: you build the solution yourself, and a consultant guides the work. One warning: self-run audits often degrade into a formality, because employees fill out tables for the boss, and the boss has no time to check them. An outside person is there to make sure the data is real. Employees describe their week more honestly to someone who has no say over their bonus or position. And a fresh pair of eyes notices routine work that people inside stopped seeing long ago.
I run these audits: a work-week snapshot, a task map, an automation matrix, a capacity and risk calculation across three scenarios, a rollout plan, and a recommendation on which task to start with. From seven days for one role, three to four weeks for a department.
If you'd like to discuss an audit for your department, get in touch.
Full version → Send a message