AI agents for business: which tasks are closed and which are not

What does this solve?

After the demonstration, the impression remains that now everything can be automated, and then there is emptiness in place of the answer to where to put it on Monday. AI agents for business achieve success in a narrow class of operations and consistently fail in the neighboring one, and the signs by which the classes differ are known in advance. Below is a filter for the suitability of the task, a map by department with both columns, a calculation of the economics of one operation and a pilot’s route, based on the results of which the decision is made based on numbers, and not on feeling.

How it works

Task suitability filter

Five signs. All five are present - the task is suitable, one is missing - the cost of implementation increases, two are missing - the task is postponed.

1. Digital input. The data is located in the system to which there is access. A telephone conversation, a paper document and a verbal agreement are not an option.
2. Result to be verified. The readiness of the operation is determined mechanically: the letter has been sent, the fields have been filled in, the amounts have been agreed. “Customer satisfaction” is verified by a survey, not an agent.
3. Limited cost of error. The action is reversible or the error is caught before the consequences. Payment, dismissal, deletion of data are outside this class.
4. Access via API. The systems have an interface. Automation of clicks on someone else's site breaks down during the first redesign.
5. Repeatability. The operation is done regularly. A one-time task once a quarter is cheaper by hand.

Map by department

There is only one pattern in the table: the agent takes away the collection, reconciliation and preparation, and the decision with the consequences remains with the person.

Economics of one operation

It is not “savings per department” that is calculated, but the cost of a unit of work:

стоимость шага = токены запросов + вызовы платных сервисов
              + доля времени человека на проверку и правку
              + доля времени разработчика на поддержку инструментов

The cost of the same manual operation is placed against it - the employee’s time at his rate. The difference is calculated based on the actual volume for the period, and not on one illustrative example. The third and fourth terms usually decide the outcome: an agent that requires checking each result saves only typing.

Rule: consider the cost of one operation as a whole, including human verification and tool support, otherwise the savings exist only in the presentation.

How to connect

Step 1. Assign owner. One person within the company is responsible for the agent's results in the same way as for the employee's work. Without an owner, the agent's mistake belongs to no one, and the review becomes a dispute.

Step 2. Select one task by filter. Not a department, not a process - one operation with a digital input and a verifiable result. The list of candidates is collected from complaints about routine, not from a list of fashionable uses.

Step 3: Measure baseline before launch. How many operations per week, how many minutes for each, how many errors and reworks. The numbers are taken from the same system by which the result will then be calculated.

Step 4. Grant access to a minimum. Separate technical user, the right to read the necessary tables, his own token for each service. Pilot entry rights are not issued.

Step 5. Start with a prompt. The agent prepares the result, the person sends it. Each edit is recorded: what was corrected and why. After a few days, the list of edits stops being replenished with new types - this is a condition for moving on.

Step 6. Open reversible actions. Draft, comment, mark, service record. Irreversible actions remain subject to confirmation.

Step 7. Make an observation. A log of all launches, available to the owner, and a notification when the threshold is exceeded: the share of edits has increased, the number of failures has jumped, the consumption of tokens for the operation has doubled.

Step 8. Decide by the numbers. After a specified period, compare with the baseline on a comparable window. Expand - to a neighboring task of the same class, and not to the entire department at once.

Useful scripts

• First line of knowledge base support with transfer of complex requests to a person along with the collected context.
• Reconciliation of bank statements with orders and output of a list of discrepancies for analysis.
• Analysis of incoming applications: classification, data enrichment, distribution by managers.
• Daily summary of multiple systems, explicitly highlighting where data does not converge.
• Placing order in directories: duplicate counterparties, empty required fields, broken connections.
• Preparation of product cards and documents according to the supplier’s data, the final button is up to the person.

Limitations

Responsibility remains with the company. Legally significant decision, personnel assessment, refusal to the client, disputed amount - up to the person, regardless of the quality of the model.

Personal data. The request goes into someone else’s circuit, and this is a separate conversation with a lawyer and security before the pilot, and not after. The processing of personal data is subject to 152-FZ, and the choice of a model provider depends on it before the quality of answers.

Tail of rare cases. The agent closes mass requests, but not rare ones, and it’s the rare ones that eat up the employees’ time. The gain should be assessed by its share of the total volume, and not by the apparent relief.

Systems without API. An accounting program without an interface means either a manual bridge or a refusal to automate this step.

Fixed support cost. Tools break following other people's updates, prompts grow old along with the process. The agent requires maintenance just like any integration.

Silent degradation. The broken agent continues to produce plausible answers. Without a set of checks and a log, a breakdown is detected by a customer complaint.

Command resistance. An employee from whom an agent “takes away” part of the work will not correct his descriptions and send examples. The pilot comes in through the one to whom the agent removes the routine, rather than adding supervision.

How to check the result

Same metric from same source. The result is compared with the baseline for the same query to the same system. Changing the method of calculation along the way invalidates the comparison entirely.

Percentage of closed without a person. How many operations went further without corrections, how many required improvements, how many were redone from scratch.

Cycle time. How long it takes from the appearance of the input to the finished result. The metric catches the case when the agent is fast, but waits half a day for confirmation.

Unit cost. Tokens and services from the journal plus human minutes for verification - against the employee’s rate on the same operation.

Errors reaching the client. Counted separately from internal ones. One such error outweighs the speed statistics and is discussed by name during the analysis.

Complete viewing in a short window. Once a week the owner reads all the routes for the day in a row. The sample shows successful cases, the continuous review shows those where the agent confidently went the wrong way.

Checking the switch. Make sure in a live run that the agent stops at the flag without rolling out and without the participation of the developer.