PRIMER

How AI actually works in a business.

For everyday work, follow five steps: read, plan, act, check, then retry or ask a person for help.

GET IN TOUCHSend an enquiry

Tell us which idea you want to try and the task it could help with.

Our team will reply within one business day.

Back to writing
← REBOTIFY WRITINGBY · · 6 MIN READ

Every “how does AI work” explainer online starts the same way. Neurons. Tokens. Embeddings. Transformers. The math of prediction. And then, somewhere around the third paragraph, it tells you about something called the attention mechanism and you realise the explanation is written for an engineer, not someone trying to decide whether to bring AI into their business.

Here is what every textbook explainer misses: none of that math tells you what AI does on a Monday morning when your inbox has two hundred emails and you have to figure out which one matters.

The useful answer is about how work is organised around the model: what starts a task, what the AI may do, and who checks the result.

Start with practical tasks: sorting email, tracking follow-ups, preparing for meetings, assembling reports, preparing document reviews, drafting customer replies, and researching prospects. AI is poor at replacing judgement about a relationship. It can help prepare the work around that judgement.

Start with two layers.

The first layer is the model. A large language model such as Claude or GPT uses your instructions and background information to generate an answer. It produces text a piece at a time; those pieces are called tokens. Patterns learned during training help it write, analyse, plan, and explain. Useful answers still need checking.

Knowing a model can produce a useful answer does not tell you how to make it part of a reliable business process.

The second layer is how the work runs.

When AI carries out a task in connected business tools, the work needs a repeatable process. These five steps make that process easier to understand and show where it can go wrong.

  • Read. The AI gathers the information for the task: an email, a support ticket, a contract, or new figures in a spreadsheet. A demo can use perfect inputs. Real work arrives incomplete or contradictory. The system needs rules for missing information, including when to stop and ask.
  • Plan. The AI works out what should happen next. What is the email about? Does it need a reply or a supervisor? Which tools are needed, and in what order? A wrong plan can waste time or send the work in the wrong direction.
  • Act. The AI uses an allowed tool: looks up a customer record, drafts an email, or prepares a meeting booking. Sending a message, moving money, or making a commitment must wait for the agreed human approval. The more systems involved, the more connections there are to check.
  • Check. The system checks whether each action worked. Did the draft appear? Did the other system accept the request? Did the customer record save? Planning an action is not proof that it happened.
  • Retry or ask for help. If the action worked, move to the next item. If it failed, try again where it is safe to do so, or pass it to a person with the details. Set a limit on retries and a clear rule for when a person takes over.

These steps are where a demo meets everyday work.

A supplier’s demo may run those steps once with clear data, responsive tools, and working connections. The test is what happens when those conditions change.

Daily use is different. The queue is long. Information is incomplete. Connected tools limit how many requests they accept, and access credentials expire. Models misunderstand unfamiliar cases. A small number of unchecked errors can still cause real trouble for customers.

A working AI employee needs four things around those steps.

First: saved guidance and history. Keep a record, owned by the customer, of what the team approved, corrected, or rejected. Include writing style and customer-specific rules. That record should remain available when the model changes, so each task does not start from scratch.

Second: human approval. Anything affecting a customer, a contract, or a commitment pauses for review before it goes ahead. The reviewer needs the relevant information and sources, as well as the draft. A confident answer can still be wrong.

Third: connections to the tools your team uses. The AI needs authorised access to the customer records, inboxes, and documents required for its task. A demonstration in the supplier’s own system does not establish that it can do the same work in yours.

Fourth: a work log and alerts. What last ran? What is waiting? What failed? What was passed to a person? If someone reports a missed task three days later, the team needs a record it can inspect.

Skipping any of these makes failures harder to prevent, notice, or recover from.

Consider how those steps apply to sorting an inbox overnight.

One hundred emails arrive overnight. The AI reads the subject, sender, and relevant history. It works out which need a reply, a file lookup, or a supervisor. It drafts replies, retrieves relevant documents, and sorts lower-priority mail. Then it checks the drafts and waits. A person reviews them, corrects a reply that misses the firm’s tone, and approves what can be sent. The correction is saved as guidance for future drafts.

The next night brings another queue. With the corrections saved, the team should see fewer repeated mistakes. In this example, by week three, email that used to wait for the team at 9am is already sorted and drafted for review. The work gets faster because the process improves, not simply because it runs more often.

That is the difference between a useful AI answer and AI doing useful work in a business: a repeatable process, saved guidance, human approval, and a record of what happened.

A founder or manager cares whether Monday morning is calmer, customers get timely replies, and failures are spotted before they cause more trouble. The useful question is whether the system completes the task safely in the tools the team actually uses.

New model releases can bring better answers. The business still needs a process that keeps what the team has learned: which sources to use, when to pause, and what people approved. A better engine does not replace those working rules.

AI works in a business by reading, planning, acting, checking, and retrying or asking for help. Its value depends on whether those steps complete the task safely, whether corrections carry forward, and whether people catch errors before they cause harm.

Start with a clear job. Use the model to prepare the work and the checks around it to keep the work reliable.

Related See how we put AI to work →

START WITH ONE TASK

Have a recurring task? Tell us what needs doing.

Send an enquiry

Tell us which idea you want to try and the task it could help with.

Our team will reply within one business day.