Run Smarter · AI Token Optimisation

AI token optimisation that cuts your bill, not its quality

If you pay for AI by usage, every word in and out costs money. We find where your tokens go, trim prompts, cache what repeats and route each job to the right model, so you pay less for the same results.

Take the AI Readiness Check See examples

Example

A month of usage pulled from your AI accounts

Spend broken down by task, team and tool

Prompts trimmed and repeated context cached

Each job routed to the right model

AI + Experts + Your Data

Any one of these on its own falls short. Together they give you the speed of AI with answers you can trust.

Does the heavy lifting

Finds the waste in prompts, context and model choice.

EXPERT

Checks it's right

Checks quality holds after every change, so cheaper never means worse.

YOUR DATA

Keeps it grounded

Your real usage logs, so savings are measured rather than guessed.

HOW IT WORKS

How it works, step by step

Here is a typical job, from the moment it arrives to the moment it is done. Watch it move through.

01A month of usage pulled from your AI accounts
02Spend broken down by task, team and tool
03Prompts trimmed and repeated context cached
04Each job routed to the right model

Where it pays off

Real examples of where AI Token Optimisation pays off. Filter by the part of the business you care about.

CONTROL

Spend broken down by task

See which jobs, teams and tools use the most tokens, in pounds rather than jargon.

All sectors

COSTS

Shorter, sharper prompts

Instructions rewritten so the model gets what it needs and nothing it doesn't.

All sectors

COSTS

Prompt caching set up

Long instructions and documents you reuse are cached, so you stop paying full price every time.

All sectors

COSTS

The right model for each job

Sorting and tagging on a small, cheap model. The big model saved for work that needs it.

All sectors

COSTS

Batch jobs that can wait

Overnight reports and bulk tasks sent through batch services at a lower rate.

EcommerceProfessional services

CONTROL

Budgets, alerts and key controls

Caps and alerts so a runaway script or leaked key can't run up a surprise bill.

All sectors

Start small. Prove it works. Then scale.

No big transformation projects. We find the one thing worth doing first, prove it with a pilot, then keep building month by month.

01 · 3 MINUTES

AI Readiness Check

Seven questions and a straight answer on whether AI fits your business yet.

02 · 1 DAY

1 Day Business Analysis

A day with your team to find the quickest wins. Free for a great fit.

03 · 2 TO 4 WEEKS

Pilot Build

One workflow built, tested and handed over, so you see it working first.

04 · MONTHLY

Monthly Support

Monitoring, training and the next build. Leave at the end of any month.

IN DEPTH

The detail, in 9 short parts

01 / 09

What Is AI Token Optimisation?

AI models charge by the token. A token is a small chunk of text, often a short word or part of a longer one. Most large language models split text up using a method called byte pair encoding, so common words are one token and rarer words become two or more. As a rough guide, 1,000 tokens is about 750 English words.

You pay for input tokens, which are everything you send in: system instructions, documents, chat history and the user's question. You also pay for output tokens, which are everything the model generates in reply. Output tokens usually cost several times more than input tokens, so a chatty answer can cost more than a long question.

AI token optimisation is the work of paying less for the same results. In practice that means sending fewer tokens, caching what repeats, asking for only the output you need and choosing the right AI models for each job. Then we check that output quality holds, because a saving that makes the work worse isn't a saving.

02 / 09

Why AI Costs Grow Faster Than You Expect

Most surprise bills don't come from one big job. They come from small habits repeated thousands of times a month, and from AI workloads that quietly grow as more of your team starts using them.

These are the patterns we see most often when we look at token usage for a small or medium business.

Long system prompts sent with every request, full of old examples nobody reads
Whole documents pasted in when one section would answer the question
Entire conversation histories re-sent with every new message, so each reply costs more than the last
The most powerful model used for simple sorting, tagging and summarising
Verbose responses where a short structured answer would do
Agents that make several tool calls per user task, each one sending the full context again
Retry loops and overnight scripts that run unnoticed until the invoice arrives
API keys shared between tools, so nobody can tell which one is spending

03 / 09

How Our Token Optimisation Works

We start with your real token usage and finish with a lower bill you can see in pounds. Every step is tested before it goes live, and nothing changes in your tools without you seeing the before and after.

Measure your token usage

We pull a month of usage from your AI accounts and break it down by task, team, tool and API key. You see the token count, the number of tokens per request and the total cost for each job. In most businesses a handful of jobs make up most of the spend, so that's where we start.

Context management and retrieval

Sending more context than the model needs is the most common source of wasted tokens. We use retrieval augmented generation, which means document retrieval that finds the relevant paragraphs instead of sending the whole file. We also summarise or trim long chat histories so you stay well inside context limits and the context window doesn't fill with old messages.

Leaner prompts and system instructions

We rewrite system prompts so they're shorter and clearer. Repeated rules, outdated examples and polite padding come out. Clear instructions often give more relevant responses as well as costing less, because the model has less to wade through.

Prompt caching and semantic caching

Prompt caching lets the provider remember the fixed start of a request, such as your instructions or a product catalogue, and charge much less for it on later requests. Semantic caching goes a step further: when a new question has the same intent as one you've already answered, the saved answer is reused and the model isn't called at all. Both save tokens and cut response generation time.

Model routing

Not every job needs the biggest model. Model routing sends each request to the smallest model that does it well, based on query complexity. Sorting emails or tagging invoices goes to a small, cheap model. Drafting a tricky reply or analysing a contract goes to a more capable one. We test model performance on your real examples before switching anything.

Output length control

Output tokens cost the most, so we control output length. That means asking for structured output formats such as a short list or a few fixed fields, setting sensible maximum lengths and removing instructions that invite long explanations. Shorter answers also arrive faster, which improves perceived latency for your team and customers.

Batch work that can wait

Overnight reports, bulk tagging and other jobs that don't need an instant answer can usually go through a provider's batch service at a lower rate. We move the right jobs across and keep anything customer facing on the normal service.

Cost controls and usage limits

Finally we set budgets, alerts, quotas, usage limits and key restrictions on each account. A runaway script or a leaked key then hits a cap instead of running for days, and the right person gets an email when spending trends change.

04 / 09

Input Tokens, Output Tokens And Your Cost Structure

Providers price input and output tokens separately and usually quote a rate per million input tokens and per million output tokens. Each model has its own rate, and the rates change often, so we always work from [current rate per million tokens] for the models you use.

Understanding your own cost structure tells you where to look first. If most of your spend is input token costs, the savings come from context management, caching and leaner prompts. If output dominates, output length control and structured answers matter more. If a few expensive models carry most of the total cost, model routing usually gives the biggest win.

We give you a simple view of average tokens per request, input versus output share and cost per task. It turns a confusing invoice into something you can act on.

05 / 09

Preserving Output Quality While Costs Fall

Cheaper is only better if the results hold up. Before we change anything, we save a set of real examples from your business with answers you're happy with. After every change we run the same examples again and compare them side by side.

This is where AI, Experts and Your Data work together. AI can suggest a shorter prompt or a cheaper model. A person who knows your work judges whether the invoice coding or customer reply is still right. Your own examples make the test fair, so savings are measured rather than guessed. If quality drops, that change is rolled back.

06 / 09

LLM Cost Optimisation For Claude, ChatGPT And Gemini

The same principles apply across providers, but the controls differ. Anthropic, OpenAI and Google each price their AI models differently, handle prompt caching in their own way and offer different spend limits and quotas. We work across all three and set up the controls each one offers.

We also look at tools built on top of these models, such as automations in Make, Zapier or n8n, coding assistants and customer service bots. They often hide token consumption inside a bigger monthly fee or a separate API account, and that's frequently where the waste is.

If your team uses per-person subscriptions such as ChatGPT Business, Claude Team, Microsoft Copilot or Gemini in Google Workspace, those aren't billed per token. There the saving comes from reviewing licences, and we focus token optimisation on the automations and tools that use the APIs.

07 / 09

AI Cost Management That Ties Spend To Business Value

Lower AI costs are good. Knowing what you get for them is better. Good AI cost management links spend to a business outcome, such as cost per ticket answered, cost per invoice processed or cost per lead qualified. Those unit economics show whether a tool earns its keep.

We tag usage by tool and business unit, so each team can see its own costs. That brings simple financial accountability without slowing anyone down, and it helps you decide where more AI spending would pay off and where it wouldn't.

For most SMEs this sits alongside wider IT and infrastructure spending rather than replacing it. The aim is cost efficiency and better resource utilisation, not cutting AI use for the sake of it.

08 / 09

Is Token Optimisation Right For Your Business?

Token optimisation pays off when you pay for AI by usage and the bill is big enough to matter. It's a good fit if any of these sound familiar.

You've built a chatbot, assistant or automation on the Claude, OpenAI or Gemini APIs
Your AI bill has grown month on month and nobody is quite sure why
A developer or agency built your AI tools and has since moved on
You want to roll an AI tool out to more staff or customers but worry about the cost
You've had one surprise bill already and don't want another

What it costs and how long it takes

Most projects start with the 1 Day Business Analysis, then a pilot that fixes the biggest jobs first. Pilots start from £1,500 and typically takes 2 to 4 weeks. We report the saving in pounds against your starting baseline, so you can judge it against what you paid.

09 / 09

Your Data Stays Safe While We Look

To find the waste we look at usage reports and, with your permission, a sample of prompts and responses. We agree up front what we can see, work in your accounts with access you control and remove that access when the project ends.

We follow UK GDPR principles: only the data we need, kept for as short a time as possible. Where prompts include customer or staff details, we can work from anonymised samples instead. [Confirm data processing agreement wording before publishing.]

FAQ

Common questions

Will cheaper mean worse results?

No. We save real examples before changing anything and test them again after every change. If quality drops, we put that part back.

Which AI providers do you work with?

Anthropic (Claude), OpenAI (ChatGPT) and Google (Gemini), through their APIs and the tools built on them.

What if we only pay for ChatGPT or Copilot seats?

Per-person subscriptions aren't billed by token, so the saving there comes from licence reviews rather than token optimisation. We'll tell you which applies on the first call.

How do I reduce AI costs?

Start by measuring where your tokens go. Then trim prompts and context, cache repeated instructions, route simple jobs to cheaper models, ask for shorter answers and set budgets and alerts. Together these can significantly reduce costs without hurting quality.

What is AI cost optimisation?

It's the practice of getting the same or better results from AI for less money. It covers token usage, model choice, caching, batch processing, licences and the controls that stop costs creeping up again.

How much is 1,000 tokens in AI?

About 750 English words as a rough guide. The cost depends on the model and on whether they're input or output tokens. Providers quote rates per million tokens, at [current rate per million tokens] for each model.

What is the best AI cost tracking tool?

Start with the usage dashboards each provider already gives you, broken down by API key and project. For several providers or tools, a simple monthly report that tags spend by task and team is often enough for an SME. We'll recommend what fits your setup.

What is semantic caching?

Semantic caching saves answers to questions you've already handled. When a new question has the same intent, the saved answer is reused and the model isn't called. It works well for repeated customer questions.