- List repeated work
- Choose a narrow task
- Keep a baseline
- Run alongside
- Review the result
Start with the week you just had
I wouldn't start with a tool catalogue. I'd look at the work that kept coming back last week. Information copied between systems. A report assembled from messages. The same question answered again. A proposal waiting because somebody had to find the previous version. Those moments give me something specific to investigate before deciding where AI belongs.
Then ask what made the work difficult. Sometimes the problem is that nobody owns it, the source information is wrong, or a decision has not been made. AI can produce another draft, but it cannot settle a pricing policy the business has never agreed on.
Choose a task where a mistake is recoverable
Internal summaries, first drafts, extracting fields from material you are allowed to use, and finding information in an approved knowledge base are sensible candidates to examine. They have outputs a person can inspect before anything reaches a customer.
Avoid making your first experiment a system that moves money, changes a contract, gives professional advice, or makes a consequential decision about someone. The right starting point depends on your work. The useful question is whether you can see an error and put it right before it causes harm.
- What input does the task need?
- Can you explain the acceptable result?
- Can the existing process continue if the tool stops?
Use what you already have
Your current spreadsheet, inbox, CRM, or booking tool may already support a rule or template that solves the problem. Try that before adding a model. An automation that copies a confirmed booking into the correct place does not become more useful because it is described as an agent.
Where language is the difficult part, add AI to that step. Let it turn a long message into a structured draft, while ordinary software checks the fields and a person makes the decision. You do not need to replace the whole business system to test that boundary.
Keep a baseline and a stop condition
Record how the work happens now: completion, errors, waiting, and checking time. Run the proposed workflow alongside it. Look at what people actually use, not how impressive the generated output looks.
Decide when you would stop the experiment. If the output needs so much repair that nobody uses it, return to the task definition. If the source material is unreliable, fix the source. If the narrow task works, expand deliberately rather than connecting every available tool.
Work through an example before buying anything
Suppose a business receives enquiries through a website form and an inbox. The owner wants an AI assistant because requests are difficult to keep track of. Before buying it, trace a request through the current process. The form reaches an inbox, someone reads it, a question is forwarded to a colleague, and the answer comes back in a different thread. Nobody can see whether the customer received a reply.
The first change in this example could be a shared record with an owner and a visible status. That does not require a model. Once that works, AI could help summarize a long request and draft a response from approved information. It should leave an incomplete request visibly incomplete rather than guessing the missing detail.
Compare the proposed draft with what the responsible person would send. Record which corrections come from missing information, wrong interpretation, or tone. If most of the work is still finding the correct source material, improve that source before spending more time on the prompt. The result you are looking for is a request that reaches the right person with enough information to act.
Write the decision down
Finish the experiment with a short decision record: continue, change the task, or stop. Include the examples tested, the errors that mattered, what people had to repair, and what the system was allowed to do. Keep unresolved questions visible rather than averaging them into a confident summary.
If you continue, name who owns the source information and who notices a failed run. If you stop, keep what you learned. Discovering that the task needs a clearer process is useful work; it prevents you from turning an organizational problem into another piece of software nobody trusts.
Sources you can check
The links support the platform details and technical distinctions. Recommendations and examples are editorial judgement.