This is a retelling of Arkady Moreinis’s material about implementing AI in companies. I shortened it and put it in my own words.
Many companies start implementing AI in the same way: they buy employees subscriptions to chatbots or build an interface on top of ChatGPT and insert it into an existing process. After a few months, it turns out that there is no noticeable financial return. Arkady explains this by saying that the process itself has remained unchanged. AI starts making money when the logic of the work changes around it.
Four mistakes
The human becomes the bottleneck
If AI works as an employee’s assistant, the speed of the process is limited by the speed of the person. A model can process thousands of tasks per minute, while an employee checks and passes on one task per hour. Most of the system’s capacity remains unused.
Automating isolated parts
When AI helps only a copywriter or only a designer, each part gets faster on its own, while the connections between departments remain unchanged. The company’s overall efficiency barely changes.
Working without guardrails
A system without defined boundaries starts inventing facts and taking random actions.
Competing with general-purpose models
A company will not beat ChatGPT at the level of basic algorithms. The value of a proprietary solution comes from data: private, cleaned and properly structured.
Five steps
1. AI runs the process, and the human joins at specific points
Arkady suggests turning the model upside down. AI does the main work, and a human is involved only in two cases:
- a legal signature is needed or someone must take responsibility;
- there is not enough data, and the decision has to be based on experience and intuition.
When handing a task to a human, AI must provide the full context: what is happening, what has already been done, and what the question is. The employee should not have to understand the situation from scratch.
2. The agent owns an end-to-end metric
An AI agent should be assigned to a final business metric that runs across several departments. The department’s job title or function is secondary.
The example from the material: an agent is responsible for filling all available appointments with specialists. It handles incoming requests. If there are not enough bookings, it calls previous clients. If that is still not enough, it launches advertising or cold outreach itself.
3. Guardrails and a metric
To make the system’s behaviour predictable, two things must be fixed in advance:
- boundaries that the system cannot cross without human approval;
- a measurable success metric.
If the metric is met and the system stays within its boundaries, small deviations in individual actions are considered normal operation.
4. Data and expert experience
A general-purpose model answers in an average way, based on the public internet. To give a company an advantage, Arkady recommends:
- collecting and structuring private data that general models cannot access;
- digitising the experience of the company’s best specialists: their notes, decisions and analyses of difficult cases. This creates digital twins of experts for unusual tasks.
5. Pay for results
It is not necessary to build everything in-house. AI agencies are emerging that use their own infrastructure to work quickly and cheaply and charge for the result (a success fee) instead of an hourly rate.
Checklist before a pilot
- A target metric is defined: it is clear which financial or operational metric AI should improve.
- Guardrails are set: it is described what the system does on its own and what requires human approval.
- Context transfer is configured: when a human is called in, AI passes on the history and essence of the problem.
- AI runs the process: a human checks the result and makes decisions at specific points.
- The system relies on company data: proprietary, cleaned and prepared.