One of our agents was running about fifteen and a half cents a minute. That's fine, but we wanted to know where it was going, so we opened the per-minute breakdown. Four lines: the platform's voice infrastructure, the language model, the voice itself, and an attached knowledge base.
The model line was higher than it should have been.
Here's why. On this platform, once the total context for a turn crosses roughly four thousand tokens, the price steps up by about twenty percent. The agent's instructions were already around 2,900 tokens on their own. The attached knowledge base pushed every turn that touched it over the line.
Now the part that matters: that knowledge base was a placeholder. It largely repeated information already in the instructions. It had been attached early, with good intentions, and never revisited.
We detached it. Cost dropped to fourteen cents a minute. And the latency estimate improved as well — less context to chew through before the first word, and no lookup step in the middle of the turn.
Same answers. Less money. Faster.
More context isn't more helpful.
Context costs you in two currencies
This is the part worth internalizing. Every extra token in an agent's context is charged twice: once in money, and once in the delay before your agent starts talking.
The second one is what your customers actually experience. Nobody on a phone call notices your per-minute rate. Everybody notices the pause.
So the rule we work by now: only attach a knowledge base when its content is genuinely rich enough to earn the tokens, the per-minute fee, and the lookup delay. A knowledge base that restates the instructions is worse than no knowledge base at all — it's paying rent to slow you down.
There are other levers. A faster model tier buys you latency for a little more money. A cheaper voice or a lighter model buys you cost, but trades against how natural it sounds and how reliably it follows its guardrails — which on a sensitive line isn't a trade worth making. Cutting context is the one lever that improves both sides at once.
What this means for you
If your agent feels sluggish and expensive, the instinct is to add configuration. Usually the answer is subtraction.
Ask whoever runs it a plain question: what's in this agent's context on every single turn, and is all of it earning its place? Old instructions nobody pruned, a knowledge base someone attached in month one, examples that stopped being relevant.
The cheapest performance improvement available to you is almost always deletion.
Want us to look at where a phone agent could take work off your plate? Grab a free workflow audit.