How to optimise your agency's budget with the BYOK model

Introduction: a budget you do not measure is not a budget, it is a surprise

With a personal API key, AI spending stops being a flat fee and becomes a variable you steer. That is the model's main advantage, and also its main trap: what can be steered can drift.

The good news is that this spending follows a simple formula, and none of its terms is mysterious:

cost = number of calls × length sent and received × price of the model chosen

Three terms, three levers. This article covers all three in order of actual return — and the highest-return one is not the one that comes to mind first. If the model itself is still abstract to you, start with what BYOK changes for an agency.

1. Before reducing anything: establish the baseline

No optimisation can be judged without a starting point. Take one full month of real consumption, per site and per task type, before touching a single setting. Without that figure, any saving announced afterwards is an impression.

What the baseline must contain, at minimum: the month's total cost, its breakdown by site, the task that consumes most, and the number of calls. Four numbers, recorded once. They will serve you for a year.

⚠ And a classic trap: the most visible task is almost never the most expensive one. Content generation is visible; automated background processing running across an entire fleet is not, and it is often what weighs.

2. The three components of the budget — and the one everyone forgets

The number of calls. How often the AI is invoked. It is the most intuitive component and the least compressible without losing function: cutting calls usually means removing a feature.

The model's price. Price gaps between models from the same provider are significant, and they do not track quality linearly. A more expensive model is not better at every task — it is better at difficult ones.

The length sent. This is by far the most underestimated component, because it is invisible. Every call carries a context: instructions, examples, page content, history. That context is billed every single time, even when it is identical from one call to the next. An unnecessarily long instruction, repeated a thousand times, costs a thousand times.

3. Lever 1 — the right model for the right task

This is the highest-return lever, and it requires no development: only one arbitration, made once.

The principle is to stop handling every task with the same model. Sort your uses into three families:

• mechanical — short rewrites, extraction, classification, tagging, single-paragraph summaries. A fast, inexpensive model is enough, and the result is indistinguishable;
• editorial — text meant to be read by a human and published. This is where quality shows, and where a premium model earns its price;
• analytical — reasoning over data, arbitration, diagnosis. Quality dominates, volume is low, so total cost stays modest even at a high rate.

The most common mistake is paying the editorial rate on mechanical volume. It is also the easiest to fix.

To open accounts with the main providers and compare their pricing yourself: OpenAI, Anthropic, Google, Mistral.

4. Lever 2 — never pay twice for the same thing

Three mechanisms, in order of simplicity.

Do not reprocess what has not changed. A product description whose source text has not moved has no reason to be regenerated. A simple version marker on the source is enough to avoid thousands of calls across a fleet.

Shorten the context, not the output. Sending a whole page when three paragraphs would do multiplies the cost without improving the result. Look at what you send before you look at what you receive: that is where the invisible savings hide.

Make batch jobs resumable. A fleet job that fails three quarters of the way through and restarts from zero pays three times for nothing. A job that resumes where it stopped pays once. It is an engineering detail that shows up directly on the invoice.

These three are process, not settings — see automating agency processes.

5. Lever 3 — protect the budget by construction, not by vigilance

A budget resting on someone's attention eventually slips, because attention is not a reliable resource. The protections that hold are the ones that act without being thought about:

• a monthly spending cap at the provider, set for each client key;
• an alert at an intermediate threshold, sent early enough to react before the cap;
• a rate limit on automated jobs, so that a badly closed loop costs an incident rather than a month of budget;
• one key per client, never a shared key. A shared key makes attribution impossible, and any overrun becomes anonymous.

That last line is the most important of the four. One key per client is what makes the budget readable — and it is free.

6. What it changes in what you bill

Once the spending leaves your contract, your offer reformulates itself. Three lines become billable that were not clearly so before:

• initial setup — provider selection, key creation, cap configuration, model-per-task arbitration;
• supervision — monthly reporting, drift detection, threshold adjustment;
• optimisation — the work described in this article, carried out on the client's fleet, with a measured before and after.

That third line has a rare property: it sells on a measurable result. You show last month's invoice and the next one. Few agency services can say the same.

On reasoning over a long engagement, see the ROI of a 6-month SEO strategy.

7. The minimum dashboard — four figures, once a month

• Total cost for the month, all clients combined
• Cost per site, sorted descending — the top three explain almost everything
• Cost per task — the one that reveals the background job you had forgotten
• Change versus the previous month, as a percentage, with a written explanation whenever it exceeds what you consider normal

The fourth is the only one that truly matters over time. A stable budget you cannot explain is not under control, it is merely quiet.

Conclusion

Optimising an AI budget requires neither an extra tool nor a rare skill. It requires three arbitrations made once — the right model per task, no reprocessing of the unchanged, caps by construction — and four figures recorded once a month.

The BYOK model does not make these levers possible: it makes them visible. Under a flat fee these optimisations exist too, but the vendor pockets them. Here, your client does — and you are the one who made them.

One closing remark, because it decides everything that follows: the first month of measurement is worth more than the next six of optimisation. Start by recording, not by cutting.

Further reading: why AIFORYA chose BYOK and what BYOK changes for an agency. Our catalogue is here: AIFORYA plugins — premium versions come with a full refund within 14 days.