AI or automation: where should a business start?
Start with a specific task. If the input is structured and the decision follows stable rules, automation may be enough. If messages need interpreting, information summarising or variable language handling, AI may help. The two approaches can work together with validation and defined responsibilities.
By the end, you can:
- Distinguish rules from interpretation
- Choose a first application
- Measure review, errors and follow-up
Rules carry out defined decisions; AI interprets variable inputs
Automation can receive a form, check required fields, create a record and notify the responsible person. The journey depends on conditions we can write and test. If contact details are missing, the request goes for correction instead of proceeding as if it were complete.
Generative AI can propose an answer or interpret text written in different ways. The result needs evaluating for the specific task. Google Cloud documentation recommends considering quality, cost, response time and human review at critical points; a convincing answer in a demonstration is only the beginning of that assessment.
A decision tree for choosing the first path
Work through the questions in the order presented. The result is a possible solution to test, not a compulsory technology purchase. If source information is incomplete or contradictory, organise it before automating the journey.
| Question | If yes | If no |
|---|---|---|
| Does the task have a defined result and owner? | Map inputs and steps | Define the process before choosing a tool |
| Does the data arrive in predictable fields? | Test validation and routing rules | Study data preparation or AI support for interpretation |
| Can the decision rules be written down? | Start with automation based on those rules | Collect examples and assess whether AI interpretation is suitable |
| Could a failure cause an action that is difficult to reverse? | Include approval before that action | Define checks and recovery proportionate to the impact |
| Are there examples and an expected answer for evaluation? | Build a comparable pilot | Prepare the set of test cases first |
Three illustrative application examples
Sales enquiries: a form with service, location and contact details can populate a record and assign the team through rules. AI can be evaluated for summarising the free-text message while keeping original fields accessible. A person confirms ambiguous information before making a commitment to the customer.
Incoming documents: when formats vary, AI-assisted extraction and field validation can be studied before data is saved. The solution should identify incomplete documents, contradictory values and inputs it cannot handle. Capabilities depend on the chosen tools and testing with the actual formats.
Customer support: an assistant can help with questions about approved business information and refer situations outside its scope. Automation organises the record and handover to the team. These examples describe project possibilities; they are not measured customer results or features attributed to every Lyntev product.
A copyable worksheet for a pilot that supports a decision
Choose a task with a clear beginning and end. Keep the current method of carrying it out and define what you will compare. Prepare common cases, incomplete cases and relevant exceptions. If you assess different versions, use the same cases to understand the effect of the change.
Worksheet for your project
Task and person responsible for the pilot: Problem observed in the current process: Input and intended result: Approved information sources: Rules we can already write: Part requiring interpretation: Normal and incomplete examples, and exceptions: Expected result in each example: What the solution may do autonomously: What requires approval: Current time and pilot time including review: Errors, corrections and cases referred: Usage and follow-up costs: Criterion for proceeding, adjusting or stopping: Owner and how to resume the manual process:
Count the complete work, including review and exceptions
Measure through to the accepted result, including corrections and human intervention. For an assistant, prepare cases such as the following and keep the answer obtained, expected result and approval decision. These are test criteria, not results already measured.
In the proposal, separate implementation, service usage and follow-up. Define who updates information. When prices, rules or integrations change, repeat affected cases before expanding use.
| Test case | Result to check |
|---|---|
| Question about a service with a published price | Uses approved information without adding conditions |
| Question about an undocumented discount | Identifies missing information and refers it; does not create a price |
| Request missing required contact details | Asks for the missing data before completing the record |
| Request for an action requiring approval | Keeps the action pending until the process’s planned check |
Choose between an existing product and a bespoke project
For customer support, ContactoIA is an existing Lyntev product: provide the company’s services, prices, opening hours and rules. Test the answers in the laboratory. Only then activate the assistant on WhatsApp. It connects as a linked WhatsApp device, not through Meta’s official API; its use remains subject to WhatsApp’s rules.
Compare the task and its limits with each product’s published features. If you need to connect tools, process specific documents or implement your own rules, the first step is to study the process and available integrations. The solution may combine existing software, automated rules and an AI component evaluated for a bounded function.
Bring the completed worksheet to the initial conversation. An objective, a few examples prepared without unnecessary data and an acceptance criterion make it possible to discuss the project precisely. Choosing the technology follows understanding the work.
Sources and criteria
Content prepared with AI assistance and research using technical sources, with checks of the relevant examples and calculations. The references support the guide’s decisions; examples identified as fictitious let you try out the method.
