How to choose your company's first task to try with AI

Start with a result you can check

A task is better defined when it has an input, an expected output and someone who can review it. “Using AI in sales” leaves too many decisions open. “Preparing a draft follow-up to a quotation, using the approved terms and without sending it” makes it possible to check something specific.

My recommendation is to choose a task whose output you can check against known information and correct before using it. Access to the data, a clear quality standard and a designated reviewer are prerequisites for the trial. Frequency helps make an improvement useful, but repetition alone does not make a task suitable.

Also compare AI assistance with a template, a rule or a formula. For a standard email, a template with fields may be enough. To recalculate a total, a spreadsheet with checked formulas offers another option. The trial should help you choose the most useful approach for that task.

The first decision may also be to organise the inputs. If the catalogue contains conflicting prices or nobody knows which terms still apply, resolving that discrepancy is part of the work. An assistant can write using either version; the company must determine which one to use.

Three business tasks that call for different tests

These illustrative situations help clarify what work the assistant can prepare, what information it needs and who reviews the output.

Comparing suppliers:

The task is to extract the price, currency, stated taxes, freight, lead time and exclusions from each offer, with a reference to the document. Missing fields are marked as pending. A reviewer checks the extracted information and recalculates the amounts that can be compared. Selecting a supplier also requires assessing quality, capacity and terms that may not appear in those offers.

Making a pricing decision:

A recommendation requires information on costs, margin, demand, commercial terms and the company's objective. Without that information, requesting a final price does not provide a defensible test. A more limited first use would be to gather the necessary data or review a cost sheet using rules already defined. The person responsible for pricing retains approval.

The comparison changes the starting point: the email can be tested if the terms are clear; the table depends on comparable documents; the pricing decision requires broader analysis. There is no universal classification by department. The same task may be ready for a trial in one company and require preparatory work in another.

An observed improvement needs a specific measure

A study published in The Quarterly Journal of Economics in 2025 examined 5,172 support agents, some of whom gradually received access to an assistant. The researchers estimated an average increase of 15% in issues resolved per hour. Agents could edit or ignore its suggestions. The main analysis examines a staggered rollout that took place mostly in 2020–2021.Source: Brynjolfsson, Li and Raymond (2025) · Generative AI at Work

The measure and the conditions matter: the study measured issues resolved per hour in that operation. The result does not mean a 15% improvement in sales, nor can that benefit be assigned to a current assistant used in a different way. The study offers a reference for how to observe a task, without determining the outcome of our trial.Source: Brynjolfsson, Li and Raymond (2025) · Generative AI at Work

For sales follow-up, you could observe how much work it takes to obtain a correct message that is ready for approval. For quotations, you could track how many fields are extracted correctly and which clarifications remain pending. The measure should describe what we want to improve and retain a quality requirement that prevents an incorrect output from being accepted.

Measure until the work is ready to use

Generation time covers only part of the process. Also record the work of preparing inputs, reviewing the output, requesting clarification and making corrections. Define the same starting and finishing points for the usual and AI-assisted methods. If several people are involved, distinguish each person's working time from the waiting time between steps.

Separate practice from evaluation and set aside cases to check the procedure learned. Define in advance what makes those cases comparable. Avoid having the same person immediately work through the same document twice: they may learn from the first attempt. Alternating the order helps balance this, but does not eliminate that learning. Where possible, randomly assign comparable cases to the two conditions and review the outputs without knowing which method produced them.

Before starting, agree on what counts as a failure: inventing a term, omitting a tax, confusing currencies or attributing a statement to a source that does not contain it. Distinguish minor errors from those that prevent the output from being used. Also record when the assistant could not complete the task and someone had to finish it. Include every attempt in the assessment, including unsuccessful ones: measuring only accepted outputs would hide the effort spent correcting or redoing the others.

The first trial does not need to become an academic study. It does need to acknowledge its limits: a few cases can reveal problems and help refine the procedure, but do not establish a stable improvement across the company. Record the tool and its version when available, the date, the instructions and the corrections.

Record the initial training and setup effort separately. Minutes freed up after a task do not automatically become cash savings: the company must decide whether to use them to handle other cases, improve quality or reduce another workload.

If you get a good draft but the review consumes the time saved, change the scope or instructions and check again. If the source information remains ambiguous, work on that information. If errors arise that the reviewer cannot detect reliably enough, the task needs another form of control before its use is expanded.

Work sequence

Prepare, generate, review and correct

1. Prepare

   Agree on the task and gather current, authorised inputs.

2. Generate

   Produce a draft; it must still be checked before use.

3. Review

   Check data, terms and calculations against the sources and quality standard.

4. Correct

   Correct errors and recheck what prevented the output from being used.

The full sequence remains visible.

Waiting and initial training are recorded separately.

Training needs protected time and a real task

The D4SME survey published by the OECD in April 2026 offers a practical warning: 39% identified maintenance costs and 38% a lack of time for training as digitalisation challenges. These responses concern digital tools in general, rather than AI alone.Source: OECD (2026) · Empowering SMEs in the age of AI

The report collected 2,018 responses across twelve countries between the fourth quarter of 2025 and the first quarter of 2026. The sample does not represent SMEs as a whole; the figures are presented as averages across countries. Nor is it a comparable continuation of the earlier survey.Source: OECD (2026) · Empowering SMEs in the age of AI

Applied training should work through the chosen task using materials the team can use, review criteria and an output that can be assessed. Someone responding to quotation requests would learn to retain authorised terms. Someone comparing offers would learn to identify missing information and check calculations. Someone approving prices would learn to distinguish available information, assumptions and pending decisions.

The person responsible at the company needs to set aside time for practice and review, as well as arrange access and information permissions. By the end, the question is whether each participant can perform and check the task. Attending a session, remembering instructions and applying the procedure at work are different outcomes.

What to decide after the first trial

Expand use within a limited scope when the reviewed cases meet the quality standard, review is feasible and the improvement justifies the effort and adoption costs. Keep a record of failures: a tool that works with uniform offers may behave differently with incomplete documents or new formats.

If results vary widely, identify the types of cases where this happens and retain the usual procedure for them. If there is no useful improvement, you can abandon that use or redefine the task. Discovering this before extending licences also informs a management decision.

The worksheet accompanying this article helps define the task, record its sources, agree on the review and document what happened. It is a preparation tool; it does not produce an automatic diagnosis or certify that a process is ready for automation.

Conclusions

Start with a task you can check before using its output. For example, preparing a follow-up email to a customer who has already received a quotation requires authorised terms and someone to check them. If a clear source or a designated reviewer is missing, address that gap first or choose a more limited task.

Evaluate how much work it takes to obtain a correct output that is ready to use. Include preparing documents, generating, reviewing and correcting; retain unsuccessful attempts too. Compare that effort and the resulting quality with the usual procedure and, where appropriate, with a template or formula. This will help you decide whether the assistance offers a useful improvement, rather than judging only the speed of the first draft.

Organise training around the tasks the team needs to perform and review. At the end of the trial, each participant should be able to explain what they checked, what errors they found and when they need help. Expand use within a limited scope if the quality and effort justify it; otherwise, correct the documents, adjust the task, strengthen the review or retain the usual procedure while resolving the problem.