In brief

  • Start not with the most visible process, but with an operation that has a repeatable input, clear rules, a measurable outcome, and manageable exceptions.
  • Assess candidates by frequency, effort, the cost of delays and errors, data maturity, the number of exceptions, and integration complexity.
  • The first pilot should cover one complete workflow: from the incoming event to the outcome, status, and a clear human action when an error occurs.
  • Record the baseline and control rules before development. Otherwise, it will be impossible to demonstrate the impact after launch or develop the system safely.

Start with the Bottleneck, Not a Fashionable Tool

Automation often begins with a technology choice: companies want a bot, a CRM system, an AI assistant, or an integration. This turns the problem upside down. A tool does not produce results by itself if the operational constraint it is supposed to remove has not been defined. First, identify an area where work is regularly delayed, repeated, lost between employees, or dependent on manually transferring the same data. Only then should you select the implementation method.

A good first candidate usually has a clear input and output. The input might be a new inquiry, a document, a payment status, or a webhook event. The output might be a created record, a verified dataset, an assigned owner, a generated QR code, or an exception notification. The more precisely these boundaries are defined, the easier it is to estimate the scope, verify quality, and prevent the pilot from turning into an endless restructuring of the entire company.

Create a Map of Manual Operations

Before prioritizing, compile a short process register. Do not begin with a large policy document: it is enough to trace the path of a customer, document, or task and speak to the people who do the work every day. A manager sees the formal workflow, while the person performing it knows the real workarounds, repeated checks, and places where data has to be reconstructed from messages. These details determine the complexity of future automation.

  1. 01

    Record the Triggering Event

    Describe exactly what starts the process: a message, form, status change, schedule, file, or employee action. If the trigger cannot be identified unambiguously, the automated workflow will run late, twice, or at the wrong moment.

  2. 02

    Document the Actual Actions

    List the systems, channels, data copying, approvals, and waiting periods in their real order. Separately mark actions that exist only because the current tools are not connected.

  3. 03

    Separate a Rule from a Decision

    A rule can be expressed as a condition and checked consistently. A decision requires context and human accountability. These parts can be combined in one process, but a machine must not be given the authority to make an ambiguous choice without due consideration.

  4. 04

    Collect the Exceptions

    Specify what happens when data is incomplete, a service is unavailable, a duplicate inquiry appears, statuses conflict, or an operation is canceled. Rare branches often make an automation that looks simple on a diagram expensive to operate.

  5. 05

    Assign an Outcome Owner

    The process needs a person who defines the correct outcome, approves changes to the rules, and resolves disputed cases. Without an owner, errors will be passed from one department to another.

Evaluate Candidates with a Consistent Matrix

The most frequent process should not always come first. High volume creates potential impact, but unstable rules, poor source data, and numerous exceptions can make a pilot unjustifiably complex. Each candidate therefore needs a qualitative assessment based on the same criteria. It is not necessary to convert it into a financial model immediately: for the first pass, a low, medium, or high rating with a short rationale for each score is sufficient.

CriterionWhat to DetermineSign of a Good First Pilot
Frequency and effortHow often the operation repeats and how much attention it consumesA regular flow of similar actions
Cost of delayWhat the business loses while the task waits for someone to handle itThe delay affects the customer, revenue, or the next stage
Cost of errorThe consequences of an incorrect result or omissionThe risk is visible, measurable, and can be checked before action
Rule maturityWhether employees make the decision consistentlyThe primary path can be expressed as conditions
Data qualityWhether stable fields, identifiers, and status history existInput data is available and verifiable
ExceptionsHow many non-standard branches occur in practiceAn exception can be handed to a person without losing context
IntegrationsWhich systems must read and modify dataA documented API or reliable exchange method is available
MeasurabilityWhether the before and after states can be comparedCycle time, errors, queue size, or the share of manual steps is known

Priority goes not to the process with the greatest number of clicks, but to the process with the best balance of beneficial impact, manageability, and the cost of exceptions.

Agentix Labs methodology

Which Processes Are Usually Good Starting Points

Strong first use cases sit at the intersection of repeatability and transparent rules. People spend their time not on expert judgment, but on transferring, checking, routing, and monitoring information. Automation handles the technical work, while the responsible employee steps in where a choice or confirmation is required. This reduces resistance from the team: the system is not trying to replace expertise, but removes the mechanical delays surrounding it.

  • Receiving and routing incoming inquiries: normalizing contact details, checking required fields, finding duplicates, assigning an owner, and tracking the initial status.
  • Synchronizing data between the website, CRM, Telegram, email, and an internal system with a single operation identifier.
  • Preparing recurring reports from already structured sources when formulas and access rules have been agreed.
  • Notifications about events, deadlines, and deviations when the message includes context and leads to a specific action.
  • Processing documents and questionnaires with mandatory subsequent validation when the quality of source files varies.
  • Technical monitoring of queues, integrations, and background jobs, with exceptions handed to the responsible employee.
  • Preliminary qualification according to approved criteria without automatically making a sensitive commercial decision.

When choosing between several candidates, favor the one that can be closed into a small but complete workflow. Automatically recording data without a status, controls, or failure handling merely moves manual work elsewhere. Value appears when the system receives an event, validates it, performs the permitted action, saves the result, and informs a person if the primary path cannot be completed.

What You Should Not Automate First

A poor first candidate is a process that the team cannot yet perform consistently. If the rules change every week, roles are undefined, and the outcome depends on verbal agreements, a software implementation will cement the existing confusion. In that situation, first agree on the workflow and data, remove unnecessary approvals, and only then design an automated process.

  • Rare operations for which development and maintenance would be more complex than manual execution.
  • Processes without an owner, a single outcome, or an agreed source of data.
  • Scenarios with numerous exceptions that cannot be formalized or handed safely to a person.
  • Critical decisions where an error is irreversible and no independent review of the outcome has been provided.
  • Attempts to replace the CRM, accounting system, and all communications with one large project without a separate pilot.
  • Automation based on incomplete spreadsheets, different identifiers, and statuses that departments interpret differently.

Describe the Target Process Before Development

For the selected candidate, create two diagrams: how the work is performed now and how it should be performed after launch. The notation can be simple, but participants, events, data, and branches must be unambiguous. Formal BPMN is useful for complex workflows, yet even a table of steps is better than a verbal description. What matters is not the beauty of the diagram, but the ability to verify that every state has a next step and an accountable person.

  1. 01

    Define the Pilot Boundaries

    Specify one input type, the primary path, a particular outcome, and the list of systems. Leave new channels and additional scenarios for the next iteration.

  2. 02

    Agree on the Data

    Define required fields, the source of truth, a unique identifier, permitted statuses, and history retention rules. This prevents parallel versions of the same operation from appearing.

  3. 03

    Describe the Controls

    Decide who can see the queue, how a stuck task is detected, where the error reason is stored, and what action an operator can take.

  4. 04

    Record the Baseline

    Measure available indicators before launch: input volume, cycle time, the number of returns, omissions, and manual touches. Do not invent baseline values if they have not been collected before.

  5. 05

    Formulate the Acceptance Criterion

    Describe observable behavior for the normal case, a duplicate, incomplete data, a cancellation, and an unavailable external system.

Build in Reliability and Manual Control

What distinguishes working automation from a demonstration is its behavior during a failure. The network may be unavailable, a webhook may be delivered more than once, an external API may return an intermediate status, or an employee may modify data at the same time as a background job. The design therefore needs to cover more than the primary happy path. Reprocessing the same event must not create a second inquiry or perform an irreversible action twice. Stable identifiers, state checks, and idempotent operations are used for this purpose where applicable.

Observability is required from the first pilot. The minimum set includes a transition log, the current state, the time of the latest successful step, a clear reason for stopping, and a safe retry method. A chart showing the total number of requests does not explain why a particular inquiry failed to reach an outcome. The system should connect a technical event to a business object without exposing unnecessary personal data in logs.

  • Irreversible and financially sensitive actions require separate confirmation, limits, or the four-eyes principle.
  • Secrets and credentials are stored outside the code, and service account permissions are restricted to the required operations.
  • An error moves the task into a clear status and preserves context instead of disappearing after a generic exception handler.
  • Manual intervention must not destroy history: the process owner should be able to see who changed the state, when, and why.
  • A rollback plan defines how to return temporarily to the manual workflow without losing new data.

How This Approach Works in Real Systems

The QRush case did not automate the entire P2C business, but a specific operational flow: receiving inquiries in real time, performing an initial assessment using trader-defined rules, and coordinating isolated trading accounts. The fast event-driven workflow is separated from the administrative backend, which handles users, profiles, limits, statuses, audit records, and commands. This example demonstrates the importance of clear boundaries: the system makes repeatable decisions, while rules and access remain manageable.

TransferBot solves a different, limited task: it guides a transfer request through a Telegram conversation, an orchestrator, and a controlled browser workflow, then returns a QR code and a payment link. Its architecture separately accounts for the queue, request state, cancellation, waiting, errors, and prevention of restarting an active transfer. The user receives a simple interface, while a controlled multi-step process is preserved internally.

How to Launch the First Pilot

Once you have selected a process, do not attempt to cover every channel and exception immediately. Divide the work into verifiable stages: start with the data and primary path, then add controls and exceptions, and expand only after stable operation. This order helps uncover incorrect assumptions sooner and avoids connecting too many systems before the process model itself has been validated.

  1. 01

    Preparation

    Assign an owner, describe the current and target workflows, confirm access, and collect baseline observations without artificially improving the metrics.

  2. 02

    Technical Prototype

    Verify receipt of the event, data quality, and the ability to perform the key action safely in a test environment. At this stage, proving the integration matters more than building a complete interface.

  3. 03

    Limited Pilot

    Connect the agreed flow, leave disputed decisions under human control, and review exceptions every day. Record new requirements separately instead of changing the rules without visibility.

  4. 04

    Acceptance

    Compare actual behavior with the criteria, and test duplicates, cancellations, errors, and recovery. Evaluate the impact only from data that was actually collected both before and after.

  5. 05

    Development

    Expand channels, roles, and scenarios after stabilizing the first workflow. The next candidate may be an adjacent process that uses already validated data and integrations.

If mapping the process reveals that the main problem lies not in manual actions but in unclear accountability or poor data, that is also a useful result. Sometimes the correct first step is not development, but simplifying the procedure and establishing consistent statuses. Agentix Labs designs automation after this analysis so that technology strengthens a functioning process instead of concealing its weaknesses.

Frequently asked questions

Which process is best for the first automation project?

A recurring process with a clear input and outcome, stable rules, available data, and a limited number of exceptions. The impact should be observable, and an error should be transferable safely to an accountable person.

Do we need to implement a CRM system first?

Not always. A CRM system is useful when the process involves customers and it will serve as the single source of state. However, the first pilot can also work with another system of record. The choice depends on where reliable data already resides and who is accountable for its quality.

Should automation begin with AI?

AI is appropriate for unstructured text, classification, or preparing a proposed decision, but it does not replace process definition. If a task can be solved with transparent rules, a conventional integration is often easier to control. Sensitive decisions should retain human review.

How can we demonstrate the impact of automation?

Before launch, record the available baseline indicators and a method for measuring them consistently after launch. Compare full cycle time, queues, manual touches, returns, and errors, but do not substitute the number of completed technical requests for the business outcome.

What should we do if a process has too many exceptions?

Separate the primary path from the exceptions. Automate the stable part and hand disputed cases to a person with full context. If the primary path cannot be separated, simplify the process and agree on the rules first.

Sources

  1. Object Management Group: Business Process Model and Notation 2.0.2Checked 30 July 2026
  2. RFC 9110: HTTP Semantics — Idempotent MethodsChecked 30 July 2026
  3. OpenTelemetry Documentation: Observability PrimerChecked 30 July 2026
This article was prepared by the Agentix Labs editorial team with AI used for research, structure and drafting. Vladislav reviews the final text, facts and recommendations.