Barry Phillips argues that HR has an important role in helping to reduce token usage in organisations

Transcript:

Hello humans!  

And welcome to the weekly podcast that aims to address an important AI issue relevant to the world of HR in around 5 minutes. My name is Barry Phillips

Today I want to talk about something that sounds technical, but really is not: token usage.

Now, before anyone panics, this is not going to be a geeky AI episode. A token is basically a small bit of text that an AI tool has to read or write. So the more you give it, and the more it gives back, the more tokens you use.

In plain English, saving tokens usually means being clearer, shorter and less repetitive.  

Why aim to save on token usage? Simple. They cost money and the current shortage of chips suggests they’ll get more expensive not less at least in the foreseeable future.

For HR teams, this matters because employees are now using AI for all sorts of everyday tasks.

So here are four simple rules HR can encourage employees to follow to reduce token usage and with it the AI bill at the end of the month.

Rule one: ask for the shape of the answer upfront.

Do not let the AI ramble.

If you simply type, “Tell me about hybrid working policies”, you might get a long essay. Some of it may be useful. Some of it may be waffle. And all of it uses tokens.

A better prompt would be:

“Give me five bullet points on what an HR manager should include in a hybrid working policy.”

That one sentence does a lot of work. It tells the AI the topic, the audience, the format and the length.

You can also say things like:

“Keep it under 200 words.”

“Give me a table.”

“Write this as a short staff announcement.”

“Give me three options.”

The principle is simple: control the output before the AI starts producing it.

Less waffle. Fewer tokens. Better answer.

Rule two: give only the relevant background.

This is a big one.

Do not paste the whole kitchen sink into the prompt.

If you want help drafting an email to staff, the AI probably does not need your full employee handbook, last year’s engagement survey,  and the MD’s life story.

It needs the key facts.

Who is the audience?

What is the purpose?

What tone do you want?

What must be included?

Is there a deadline?

Is there anything it must avoid saying?

Think of it like briefing a human colleague. Too little context gives poor results. Too much context wastes tokens and can confuse the answer.

Rule three: reuse and refine sensibly.

Once the AI has produced something useful, do not automatically start from scratch.

Build on it.

You can say:

“Make that shorter.”

“Turn it into a staff email.”

“Make it warmer but still professional.”

“Rewrite it for line managers.”

“Give me a version in six bullets.”

That is often more efficient than writing a brand-new prompt every time.

But there is a warning here. If the chat has gone off in six different directions, start a new conversation. Old context can become clutter. And clutter can lead to confused answers.

So the rule is: reuse the conversation while it is still focused. Start fresh when it becomes messy.

That is not just good token hygiene. It is good common sense.

Rule four: turn off memory when it does not help.

Memory can be useful. It allows the AI to remember details from previous conversations, such as your preferred writing style, your role, or your organisation’s tone.

That can be helpful when you want personalisation. For example, “write this in my usual style”, or “use our standard policy tone”.

But memory is not always needed.

For routine HR work, such as summarising a policy, drafting a neutral job advert, producing interview questions, or creating a simple checklist, memory may add very little.

In some cases, it may bring in unnecessary background from previous chats. That means the AI may be processing more context than it needs.

So a good rule is:

Turn off memory when you do not need personalisation.

So, to recap, here are the four rules.

First, ask for the shape of the answer upfront.

Second, give only the relevant background.

Third, reuse and refine sensibly.

Fourth, turn off memory when it does not help.

And here is the bigger message for HR.

Saving tokens is not really about being technical. It is about being a better briefer.

Clear task. Clear audience. Clear length. Clear output.

That saves tokens and, frankly, saves everyone from AI-generated sludge.

As always thanks for listening. 

Until next week.

Bye for now!