The right first project fits in one sentence: when a specific event occurs, a defined task is performed and someone knows how to check the result. Start with that circuit. You can expand it after testing it.
Imagine a quote request arriving through your website. You copy the name into a spreadsheet, find the message, prepare a record, then work out who should handle the next step. Automation becomes interesting as soon as you can describe that transition precisely. What needs to come in? Where should the information go? Who needs to act? What happens if something is missing?
Our advice: first observe a real task with the person who handles it. The CNIL also recommends starting with an identified need before deploying generative AI.1 Here is a method for moving from that observation to a first measurable trial.
Start with a task, not a tool.
“Automating administration” is still too broad to decide what to build. “Creating a tracking record from a complete form” already defines a scope of work. You can identify a beginning, an end and an expected result.
We suggest looking for three characteristics: a recurring task, a rule you can explain and an error you can recover from. If every case requires negotiation, a fresh interpretation or a sensitive decision, start by preparing information for the person responsible.
Five ideas from everyday SME work
Illustrative scenariosIncoming enquiries
Create a record, check the fields and notify the relevant person.
Quote preparation
Gather case data and pre-fill a working document.
Reporting
Bring defined figures together in a monitoring view at a chosen frequency.
Document filing
Suggest a category and save a copy in the authorised folder.
Appointment follow-up
Prepare a reminder based on a confirmed appointment and the agreed rules.
These ideas are not a universal ranking. In your business, a simple duplicate-entry task may deserve a trial before a document assistant. Conversely, if input data is often incomplete, improving the form may be the first useful action.
For sales sequences, the guide automating prospecting without losing the human touch covers segmentation, triggers and follow-ups in more detail.
Rules execute. AI suggests. You approve.
Automation can simply connect tools under defined conditions: if the form contains the required fields, create a record and send an internal notification. Adding a language model is not essential in this scenario.
Generative AI can be considered for summarising a free-text message or preparing a rewording. It can also produce a plausible but inaccurate response, a limitation highlighted by the CNIL.1 Therefore provide a source that can be consulted and approval appropriate to the action.
- 1InputThe enquiry arrives
A form sends the intended information.
- 2RulesThe file is prepared
Fields checked, duplicates searched for, record created.
- 3AI, if usefulA summary is suggested
The original message remains available for checking.
- 4Person responsibleYou decide what happens next
You confirm the need and approve the response.
A document is missing, a duplicate appears or the summary looks doubtful? The case joins a queue for manual handling, with its reason. It does not disappear in the process.
The essential distinction lies between “prepare” and “act”. Pre-filling a quote and sending it to the client are two separate actions. For a first pilot, we recommend having the document reviewed and keeping sending under human control. Also define who takes over if the service becomes unavailable.
A successful first automation is a transition your team understands, can check and knows how to take over.
The working principle proposed in this guide
Count the net time, then the total cost.
Before discussing savings, establish a basis for comparison. Over a period representative of your activity, record the number of relevant cases and the time actually spent on each. Separate routine cases from those you know must be handled manually.
Then estimate the work that will remain: reviewing, correcting, handling exceptions and maintaining the connection between tools. To avoid double-counting, include routine corrections in the time per case and additional interventions in monthly monitoring.
How much time really remains?
Editable fictional example. Enter only the cases the process could handle.
60 × (8 − 3) − 60 = 240 minutes per month.
Initial setup time and expenses are not included. The pilot must confirm the assumptions.
The calculation covers only the cases you selected. Work on other cases remains outside the comparison. If the result is negative, your assumptions indicate extra workload: the scope, method or level of oversight should be reviewed.
Four cost areas to include in your decision
- Setup
- Scoping, configuration, data preparation, testing and training.
- Operation
- Subscriptions, service calls and any hosting.
- Monitoring
- Reviewing results, maintenance and incident handling.
- Takeover
- Documentation, data export and the ability to continue manually.
Available time does not automatically become money saved. Decide what it will be used for: responding more attentively, clearing a backlog or developing an activity. If you assign a value to an hour in your financial calculation, use an assumption specific to your business and state it explicitly.
At Kanexio, the scope is defined after observing the process, and any potential gains are verified through a pilot. Licences and external services are among the elements to examine in the total cost.2
Compare your tasks before choosing.
Take five tasks that recur during your week. For each one, enter an approximate monthly volume and the time spent on one case. This initial record provides a basis for discussion; then correct it using measurements.
The framework below selects tasks for which you consider the rules clear and the access and error handling defined. It then ranks them by gross time currently spent. A task at the top deserves examination but is not necessarily the most profitable to automate.
Five tasks. One starting point.
Rename the tasks and estimate their volume and duration. Complete at least one row. The others can remain blank.
The framework compares the current time you enter. It predicts neither time saved nor project profitability.
Once two candidates have been identified, compare the practical difficulties: do the software tools have the right connections? Who can grant access? How should a failure be reported? Can the person handling the case easily correct the result? These answers may move a lower-volume task to the top.
If every row remains “to be specified”, the next step is already clear: document how the process works. Take a routine case, an incomplete case and an exceptional case, then write down the expected response for each.
A first pilot, with a beginning and an end.
We recommend starting with a single input, a defined result and one person responsible. For example: when a complete enquiry arrives, a working record is prepared for review. The final quote and its sending remain in the usual process.
The scope to document before building
- Trigger. The exact event that starts the process.
- Data. The necessary information, its sources and authorised access.
- Expected result. What must be created or prepared, and where.
- Review. Who checks, using which criteria and sources.
- Exception. Where an unprocessable case goes and who is notified.
- Measurement. Total time, corrections and incidents to compare with the initial situation.
Before connecting a tool to real files, determine which information may be sent to it and what the provider may do with it. The CNIL recommends defining authorised uses, checking data processing terms and training users.1 This scoping deserves particular care when files contain confidential or sensitive information.
Then test the planned situations, including errors. Should a retry create a second record? What happens when a document is missing? How can the source of a summary be traced? Have the person who will use the process walk through it, then adjust the points that make their work harder.
At the end of the pilot, decide from observed results: continue, adjust the scope or stop. Keep a clear takeover procedure and a named owner. This first experience will also provide a better basis for estimating the next project.
Let's identify your first useful automation.
Bring a task that takes up your time, the tools involved and, if you have completed it, your framework. We can then clarify the starting point.
Questions before getting started.
Which task should a small business automate first?
We recommend choosing a repetitive task with clear rules, authorised data and errors that are easy to recover from. Compare its volume and current time, then deduct oversight and maintenance to estimate net time. The framework in this article helps prepare that choice.
Is artificial intelligence always necessary?
No. A transfer between two tools, a notification or classification based on a specific field can follow rules. AI may be considered when text needs interpreting or a summary needs preparing. Its output must remain verifiable, and manual takeover must be planned.
Is the time estimated by the calculator guaranteed?
No. The calculator applies your assumptions: eligible cases multiplied by the difference between current time and remaining time, less monthly monitoring. It does not include initial setup or expenses. The pilot must be measured to establish what time is actually freed up.
Can we start without changing all our software?
That is the first scenario to examine: look at the features of your current tools and their connection options. Feasibility depends on access, supported actions and your process rules. A missing or overly limited connection may change the scope.
Sources and method.
This guide proposes a Kanexio decision-making method. The cases described and the calculator's initial values are illustrative. They do not represent a client's results. The references below support the passages associated with them.
- CNIL
Capabilities and limitations of generative AI, concrete needs, and governance of uses and data. Published on 18 July 2024. Source consulted on 7 September 2026.
How to deploy generative AI? - Kanexio
Internal commercial source: scope of the Automation & AI offer, scoping and monitoring method. Source consulted on 7 September 2026.
Our approach to automation
What we do not quantify
No average hours saved per SME, productivity percentage or payback period. Those figures require data specific to your business. The calculator presents an arithmetic scenario; the framework organises your inputs. Neither replaces measurement of a pilot.
The comments about data help prepare the scope. They do not constitute legal validation of your system. Applicable arrangements must be examined according to the data, tools and intended use.