When to recognise your business needs AI, or not

When to recognise your business needs AI

There is probably a job in your business that everyone wishes took less time. Preparing customer updates. Pulling figures into a weekly report. Finding the information needed to answer a straightforward question. You may already have wondered whether AI could help, then found yourself looking at tools without knowing quite what to ask them to do.

That is a good reason to pause and look at the work itself. Before choosing software, establish which problem is worth solving, who would benefit and how you would recognise an improvement. You can then give AI a specific job and see whether it earns its place.

Find the job behind the frustration

“We should use AI in sales” leaves you with almost every decision still to make. “Our account manager spends Friday putting customer updates together, and customers still have to chase us” is a much better starting point.

Talk to the person preparing those updates and someone who receives them. Where does the information come from? What takes the time? Is the writing difficult, or is most of the effort spent chasing missing details?

The answer matters. If nobody records when a job will finish, a faster way to write emails will leave customers just as poorly informed.

Walk through the task from beginning to end. Look for the particular step causing the trouble, and describe the improvement you want in everyday terms: fewer customer chasers, less time spent assembling updates, or more accurate information reaching the right person.

Choose an approach that suits the work

Once you understand the task, it becomes easier to judge what help it needs. A shared template might resolve inconsistent updates. Ordinary automation might move figures between systems. AI could help turn a set of job notes into a readable message.

An agent becomes an option when the system needs to choose its own steps and use tools to complete the job. That adds decisions about access, authority and supervision. It is worth taking on that work when the task calls for it.

Anthropic makes a similar recommendation in its engineering guidance on effective agents: start with the simplest approach that works, and add complexity where needed.

For a first workflow, drafting something that a colleague checks may be plenty. It gives you a useful place to learn without handing over the customer relationship.

Compare candidates with this scorecard

If you have several ideas, this handy scorecard gives you a consistent way to compare them. Complete it with someone who does the work; they will usually spot practical difficulties that a planning meeting misses.

Give each item 0 if you cannot answer it yet, 1 if you have part of the answer, or 2 if you can answer it confidently. Add a short note explaining your score.

• Is the problem worth solving? Who struggles with it, and what would improve for them?
• How often does it happen? How many times a week or month? How long does it take, including corrections?
• Can we use the information needed? Is it available, reliable enough and approved for use in the chosen tool?
• Do we understand the process? Are the steps reasonably consistent? What happens in unusual cases?
• Who will take responsibility? Who knows the work and can make decisions about the pilot?
• What happens if it gets something wrong? Can someone spot the mistake and correct it before it causes harm?
• How will we judge improvement? What will we measure, and what quality must the work meet?
• When would we stop? What would make us pause the test, and how would the job get done instead?

The totals help you choose where to look first. This is a discussion tool, rather than a tested formula for predicting success.

Use judgement alongside the numbers. An attractive score cannot make up for information you have no permission to use or mistakes nobody can check. Work involving payments, employment decisions, safety or binding commitments needs more specialist consideration than a first experiment should carry.

A task that barely happens, changes every week or takes longer to check than to complete may also be a poor starting point. Pick something with enough repetition to learn from and enough value to justify the effort.

Try it on a small piece of real work

You do not need to redesign a department to find out whether an idea is useful. Choose one task, agree who will run the trial and give it a sensible time and spending limit.

Write down what the AI will do and what the person will still do. Agree which information it can use, whether it can change anything and where approval is needed. Keep the current process available while you test.

Then try ordinary cases and the awkward ones. Count the time spent checking and correcting, as well as the time spent producing the first answer. Before starting, agree what result would justify carrying on and what would make you stop. This gives you something firmer to discuss than whether people liked the demonstration.

Worked example: customer update drafts

Take a business where an operations coordinator sends 30 customer updates each week. Finding the details, writing the message and checking it takes about 12 minutes per customer—six hours in total.

The team decides to try AI drafting for two weeks. It will use job records the business has approved for that purpose. The coordinator will check and send every message; the AI will have no access to send emails or alter the records.

Across the 60 updates, the team wants to see whether it can halve the time per message without introducing important mistakes. Dates, job status and promises to customers all need checking. If a detail is missing, the draft should say so.

The coordinator records corrections and time spent reviewing. An invented promise, an unauthorised disclosure or a message sent without approval would stop the trial immediately. If the records are too patchy or checking takes too long, the team will need to rethink the approach.

Suppose the total time does fall to six minutes per update. That saves three hours a week. The next question is what the coordinator can do with those hours. If other commitments mean only two hours can be reassigned, the business has saved three hours but gained two hours of useful working capacity.

At an illustrative staff cost of £30 an hour, the time saved is worth £90 a week on paper. Whether the business is financially better off depends on what changes: perhaps it avoids paid overtime, or uses the extra capacity for work that earns money. Software, setup, training and running costs also need paying for. Review time is already included in the six-minute figure, so it should not be deducted twice.

That is a more useful account of the result than claiming £90 in weekly savings simply because a task became quicker.

Decide whether it deserves a bigger role

At the end of the trial, ask the people using it: is the work accurate, is the checking manageable, and has it made a worthwhile difference?

If it has, you have a reason to continue. If one particular problem is getting in the way, decide whether another short trial would resolve it. If the effort outweighs the improvement, stop and use what you have learned to choose a better candidate.

Your first AI workflow should give you a useful improvement and a clearer understanding of where AI fits in your business. That is a sound basis for deciding what to do next.

Want help choosing and testing your first workflow?

Cariboo can help you work through the options, plan the pilot and build it with your team. Explore how we can help.

‍

‍