How do I cut my AI bill with one-click recommendations?

Use Cognocient's Recommendations page to find your biggest cost savings and apply them without touching your code.

Goal: Identify and apply the top cost-saving opportunities in your AI spend — without any code changes.

Time: 30 minutes to apply. Savings appear in the next billing period.

Prerequisite: At least 7 days of API calls through the Cognocient proxy so there's enough data to analyse.

Step 1 — Open Recommendations

Click Recommendations in the sidebar (this is a different page from AI Advisor — Advisor is a natural-language Q&A chat over your spend data; Recommendations is the one-click-apply savings list this guide walks through).

Cognocient analyses your call patterns and ranks recommendations by estimated monthly saving. Each card shows:

What the issue is — e.g. "chatbot uses GPT-4o for 3-token classification outputs"
Estimated saving — e.g. "$420/month"
Confidence — based on how many calls it's seen
One-click apply — creates a routing rule automatically

Start with the highest-saving recommendation at the top of the list.

Step 2 — Apply the top recommendation

Click Apply Recommendation on the top card. For a model-switch recommendation, Cognocient immediately creates a routing rule that reroutes matching calls to the cheaper model — there's no separate preview or confirmation step first.

How it works

Cognocient creates a routing rule that intercepts calls matching the scope and rewrites the model in the request before forwarding to the provider. Your code keeps sending gpt-4o. The proxy silently sends gpt-4o-mini. Your code sees the same response format.

Step 3 — Check the result after a few days

The rule you just created now lives on the Routing Rules page, alongside its estimated monthly savings figure. There's no built-in "actual savings realized" tracker or automatic quality signal on the recommendation itself — check your feature's real cost trend in Feature Intelligence and your own quality signals (e.g. error rate, user complaints) directly.

If something degrades, go to Routing Rules, find the rule, and either Pause it (keeps the rule but stops applying it) or delete it — there's no one-click "revert" action, but pausing has the same effect.

Step 4 — Work through the list

Repeat for each remaining recommendation. Common findings:

Finding	Typical saving
GPT-4o used for short classification outputs	60–85% cost reduction
Identical prompts sent repeatedly (no caching)	75–90% via prompt cache
Batch-eligible workload using real-time endpoint	50% with Batch API
Context window bloating after 5 turns	20–40% with context trimming

If you don't see recommendations yet

The Advisor needs enough data to be confident. If your call volume is low:

Make sure calls are going through the proxy (check Live Calls tab)
Make sure you've added X-Cost-Feature headers — without them, all calls look like a single unattributed pool
Wait 7+ days for call patterns to stabilise

See Tag Your First AI Call if attribution isn't set up yet.