The gap between businesses getting real value from AI and businesses talking about it is usually not budget or technical skill. It is that the first group picked one unglamorous task and finished it.
Start with the task, not the tool
Do not begin by choosing a platform. Begin by finding work that is repetitive, text-based, low-risk and currently slow. The characteristics that make a task a good first candidate:
- It happens often — daily or weekly, not once a year.
- It is writing-heavy: drafting, summarising, reformatting, replying.
- A mistake would be annoying, not catastrophic.
- Someone can check the output quickly without special expertise.
Typical first wins: turning meeting notes into a summary, drafting the first version of a routine reply, tidying up a document, or turning a long thread into a short list of actions.
Write the data rule before you write the prompts
This is the step people skip and regret. Before anyone pastes anything anywhere, agree what must never leave your systems:
| Data type | Default position |
|---|---|
| Client or customer personal data | Not into public AI tools |
| Financial records and payroll | Not into public AI tools |
| Contracts and legal documents | Only in an approved, business-grade tool |
| Passwords, keys, access details | Never, in any tool |
| Public marketing text | Generally fine, with review |
The distinction that matters is between consumer tools, where your input may be used to improve the service, and business tools with a contractual agreement that it will not be.
Keep a human on the output — and mean it
AI produces fluent text, which is exactly why errors slip through. Fluent wrong answers do not look wrong. Set a rule that anything going to a client, a regulator or the public gets read by a person who is accountable for it.
You are not checking whether the writing is good. You are checking whether the claims are true.
Measure one thing
Pick a single number you already know, and compare after a month. Time spent on the task, or how many you get through in a week. Without that, you cannot tell genuine improvement from the feeling of being busy.
What this actually costs to start
A common reason small teams stall is the assumption that AI needs a budget line. The honest position:
| Approach | What it really involves |
|---|---|
| Free or built-in tools | Enough to prove whether the task is worth automating at all |
| A business-grade subscription | The point at which you involve client data or need confidentiality terms |
| Automation built for you | Worth it once you know the task and can describe the outcome you want |
| Governance and guardrails | Small effort, and it is what makes the rest safe to use |
The expensive mistake is not the licence cost. It is buying seats before you have a use case, then declaring AI does not work because nobody adopted it.
The five mistakes that waste the most money
- Buying licences for everyone before anyone has a working use case.
- Starting with the most sensitive data in the business.
- No written rule, so staff quietly use consumer tools on client data.
- Treating generated text as finished work and sending it out unread.
- Automating a process that should have been redesigned instead.
What to do next
If you want a structured view of where you are, an AI readiness assessment will tell you which processes are worth automating and which should stay manual. For turning that into working automation, see agentic AI and automation; for the rules and controls that keep it safe, AI governance and guardrails.
If you are deploying AI inside Microsoft 365, Copilot deployment covers the rollout and the permissions that make it safe. And if you simply want one task automated properly, start with a conversation.