The wrong order: tool before process
When a company starts thinking about AI, automation quickly becomes part of the conversation. That is natural. The problem begins when the discussion moves too quickly to tools and not enough time is spent on the process itself.
A better question is: which process currently costs us the most time, money, or errors?
AI should not be the starting point. It should be one of the possible answers after the process is understood.
If a company does not know how a process really works, how long it takes, how many people are involved, where delays happen, and how often work needs to be corrected, then AI may simply add another layer of tooling. Sometimes that layer is more expensive, harder to maintain, and less predictable than a simpler solution.
The problem is not AI itself. The problem is using AI before the company has understood the process, the data, and the actual operational friction.
What a baseline means and why it matters
A baseline is your current state before change.
It does not have to be a complex consulting exercise. In practice, a process baseline answers a few simple questions.
How long does the process take? How much does one case cost to handle? How many errors happen along the way? How many cases go through the process each month? How many people and systems are involved? Where is data copied manually? Where do exceptions appear?
Without these numbers, it is difficult to judge whether AI has improved anything.
You may have a new tool, an AI agent, a chatbot, or an automation. You may feel that the company is becoming more modern. But if you do not know that one case previously took 18 minutes and generated corrections in 12% of cases, you cannot prove that the new solution is better.
Maybe the process is now faster. Maybe the preparation step is shorter, but the review step is longer. Maybe AI helped one part of the process, but increased the need for human verification. Without a baseline, this stays invisible.
What to measure before implementing AI
You do not need to measure everything. A good starting point is three areas:
- time
- cost
- errors
The first area is time. How long does the process take from start to finish? How much of that is actual work, and how much is waiting? In many companies, the work itself may take 20 minutes, while the case waits for two days because someone needs to approve, provide data, or respond.
The second area is cost. This does not need to be a perfect financial model. Start with a rough calculation: how many people are involved, how much time they spend, and how often the process happens. If one person spends one hour every day copying data manually, that becomes a real monthly cost.
The third area is errors. How many cases require correction? How often does someone need to go back to a customer, employee, or supplier for missing information? How often are decisions reversed or data corrected manually?
After that, you can add a few additional elements: case volume, manual data entry points, the number of systems involved, and the most common exceptions.
These elements help you decide whether the right answer is AI, traditional automation, system integration, or simply a better process.
How to collect data without a large consulting project
Measuring a process does not need to become a large transformation program.
In a small or mid-sized company, start with one process that actually creates friction. It could be handling sales inquiries, preparing reports, processing invoices, updating the CRM, managing complaints, writing customer emails, or collecting documents.
Then walk through several real cases. Do not describe the process from memory only. Look at how it works in practice. Ideally, ask the person doing the work to show the full flow from start to finish.
Measure several iterations, not just one. One case may be too simple or too unusual. A few repetitions will show what is standard and where exceptions start to appear.
A simple spreadsheet is enough. The columns can be basic: case number, start time, end time, actual work time, number of errors, people involved, systems used, manual data entry points, and notes.
After one or two weeks, you will often see more than after a long internal discussion about where AI could be used.
You do not always need a full month of data. If the process has high volume, a shorter measurement period may be enough. If the process happens less often, you need more cases. The goal is not perfect measurement. The goal is data good enough to make a better decision.
AI, automation, or a simple script?
Once you have a baseline, you can ask the right question: what kind of solution makes sense?
AI makes sense where the work requires context, language, incomplete information, or many possible variants. It works well for research, analyzing inquiries, suggesting solutions, noticing details that are easy to miss, and creating customized messages for specific recipients.
A simple script makes sense where the task is repeatable, deterministic, and clearly defined. Examples include changing statuses, generating files, updating a spreadsheet, sending data to an API, creating records, or checking simple conditions.
An integration makes sense where the real problem is copying data between systems. If someone copies data from a form to the CRM, from the CRM to a spreadsheet, and then from the spreadsheet to a mailing tool, AI does not have to be the first answer. Connecting the systems may solve most of the problem.
A process change makes sense where the issue is unnecessary steps, unclear ownership, or decisions being made too late.
This matters because AI is attractive. It creates a sense of progress. But from a business perspective, the outcome is what matters: less time, lower cost, fewer errors, or lower risk.
A practical example: an agent versus a simple script
I tested this myself while working on a process for preparing customized emails.
The first approach was agent-based: the agent was supposed to handle the entire process end to end. It analyzed a spreadsheet with contacts, prepared messages, changed statuses, communicated with an API, and created drafts. It sounded good, but in practice the agent was doing too much.
One part worked correctly, another failed occasionally. The API returned errors. The agent interpreted data well in one run and worse in another. The whole process became non-deterministic.
The result was that preparing 10 emails took around 10 minutes. On top of that, there were API-related errors and inconsistent behavior: sometimes it worked better, sometimes worse.
The conclusion was simple: AI was needed only for one part — customizing the message. Everything else, including spreadsheet handling, statuses, API calls, and draft creation, should have been handled by a regular script.
After splitting the responsibilities, the script could prepare 50 emails in a few seconds, while AI did only what it was actually good at: personalizing the content.
This is an important lesson. Not every process needs to be agentic. Sometimes the best result comes from combining simple code, clear rules, and AI used only in the part where it makes sense.
A simple framework for measuring a process before AI
Before deciding on AI, go through a simple checklist.
- Measure process time. How long does the full process take? How much of that is actual work? How much is waiting?
- Estimate cost. How many people are involved? How much time do they spend? How often does the process happen?
- Check errors. How many cases require correction? How often is data incomplete? How often does work return to a previous step?
- Look at volume. How many cases happen daily, weekly, or monthly?
- Count the systems. How many tools are involved? Does the process move through email, CRM, ERP, spreadsheets, PDFs, or messaging tools?
- Find manual data entry points. These are often strong candidates for automation or integration.
- Look for exceptions. How often does the process deviate from the standard path?
- Assess data quality. Is the input data complete, consistent, and available? Is the output usable without manual correction?
- Finally, assess risk. What happens if the automation makes a mistake? Can the mistake be detected easily? Does a human need to approve the result?
Only after this review should you decide whether AI is needed.
The practical conclusion for business owners and managers
The best AI implementations do not start with a model, an agent, or a tool.
They start with understanding the work.
If you know how long the process takes, how much it costs, and where errors appear, you can make a sensible decision. Maybe AI is the right answer. Maybe a simple script is enough. Maybe you need an integration. Maybe the first step is to remove an unnecessary step or clean up the data.
It is also worth remembering that simple custom automation has become much more accessible than it was a few years ago. Writing a script that handles a specific part of a process is often not a project measured in weeks. Sometimes it is a matter of minutes or a few hours.
This changes how small and mid-sized companies should think about automation. You do not always need to build a large system or implement a complex AI tool. In many cases, a well-measured process, a simple script, AI used in the right place, and clear control over the outcome are enough.
That is less exciting than presenting a new AI tool.
But it is much more useful for the business.
Because the goal is not to implement AI.
The goal is to build a better process.
