Where AI fits in your business, and where it does not
AI fits where work is frequent, text-heavy and tolerant of a checked draft. It does not fit where a mistake is expensive and rare. A map of your business.
The short answer
AI fits a task when three things are true: the task happens often, most of it is reading or writing, and a mistake is cheap to notice and correct. It does not fit when a mistake is expensive, rare, or something a customer expects a person to have judged. Apply those three questions to the tasks in your business and you have a map that is more useful than any list of use cases.
The map, by kind of task
| Fits well | Fits with a person deciding | Does not fit |
|---|---|---|
| Sorting and routing incoming email | Drafting replies to customers | Handling a complaint end to end |
| Summarising meetings and documents | Preparing quotes and proposals | Setting prices |
| Extracting data from invoices and forms | Matching invoices to orders and flagging mismatches | Approving payments |
| First-draft job descriptions and internal texts | Screening applications against stated criteria | Hiring decisions |
| Answering standard questions from your own material | Suggesting next steps in a sales conversation | Committing the company to terms |
| Tagging, classifying, translating for internal use | Drafting contracts from your templates | Legal advice |
| Internal search that understands questions | Weekly report commentary | Anything a regulator would expect a human to have decided |
The middle column is where most of the value sits: AI does the frequent, text-heavy preparation, and a person makes the decision that matters. Left column tasks can run largely unattended once monitored. Right column tasks stay human, and the strategy should say so explicitly.
How to draw your own map
- List the tasks, not the roles. Ask each team for the ten things they do most often. You will get a list of about fifty tasks for a small company.
- Score each task on the three questions. Frequency, text share, cost of a mistake. A simple high, medium, low is enough.
- Place each task in one of the three columns. Frequent and text-heavy with cheap mistakes: fits. Frequent and text-heavy with expensive mistakes: fits with a person deciding. Everything else: does not fit, or not yet.
- Look for the split. Many “does not fit” tasks have a preparation part that does. Move the preparation to the middle column and keep the decision on the right.
- Pick three from the left and middle columns. Those are your first projects, and the map is box one of your strategy.
What changes over time
The columns move slowly. As models improve and as your own experience grows, some tasks migrate from the middle column towards the left, because the logs show mistakes are rarer than feared. Almost nothing moves from the right column, because it was placed there for reasons that have nothing to do with capability: what customers expect, what regulators require, what the business is willing to have decided by a probability.
What this means for you
Draw the map with your team in an afternoon. It replaces the endless question “what could we do with AI” with a shorter one: which task in the left or middle column do we start with. It also gives you the right-hand column in writing, which is the part of an AI strategy that goes unstated and is later regretted.
Frequently asked questions
Is there any department where AI does not fit at all?
No, and none where it fits everywhere. Finance has invoice reading, which fits, and payment approval, which does not. HR has drafting job descriptions, which fits, and hiring decisions, which do not and are regulated more strictly. The map is drawn per task.
What about customer-facing uses?
They fit when the AI answers from your own material and a wrong answer is cheap to correct, such as opening hours or how to start. They do not fit when a customer reasonably expects a person to have judged the situation, such as a complaint or an exception. The line is what the customer expects, not what the technology can do.
How do we handle tasks that are frequent but high-stakes?
Split them. Let AI do the preparation, reading, extracting, drafting, and keep the decision with a person. Invoice processing is the classic case: the model reads and matches, a person approves anything that does not match. The frequent part is automated; the expensive part is not.