How do I ask questions about my AI spend in plain English?

The AI Advisor answers any question about your AI spend using your actual call data — cost trends, waste patterns, feature performance, and more. Type a question in plain English and get a data-grounded answer in seconds, with full reasoning shown.

How it works — bring your own AI provider key

AI Advisor uses the AI provider key you already connected to Cognocient (Anthropic, OpenAI, or Azure OpenAI). Each question is a direct call from Cognocient's backend to your provider, using your key. The API cost appears on your own provider bill — Cognocient does not charge for Advisor queries separately.

Your browser
      │
      │  "Which feature has the worst efficiency this month?"
      │  Session token: Bearer eyJ...  (your Cognocient login)
      ▼
Cognocient platform
      │
      ├── 1. Authenticate your session
      ├── 2. Fetch your spend data (scoped to your account only)
      ├── 3. Look up your connected provider key (Anthropic / OpenAI / Azure)
      ├── 4. Call your AI provider using YOUR key
      │         ↳ billed to your Anthropic / OpenAI account
      │         ↳ ~$0.003–$0.01 per question at current pricing
      ├── 5. Return the answer, grounded in your data
      ▼
Your browser — answer appears in the Advisor panel

Prerequisites

What your key powers

What you can ask the AI Advisor

The Advisor can answer any question that can be answered from your API call metadata, spend history, and waste analysis. It cannot read your actual prompt or response content.

Cost investigation

• "Why did our AI spend increase 40% last month?"
• "Which feature had the biggest cost increase in the past 2 weeks?"
• "What caused the spend spike on November 14th?"

Budget and forecasting

• "Are we on track to stay within our Q4 AI budget?"
• "If usage grows 20% next quarter, what will our bill be?"
• "Which departments are at risk of exceeding their budget this cycle?"

Optimisation opportunities

• "What is the single biggest cost-saving action we could take right now?"
• "Which features should we prioritise for model optimisation?"
• "How much could we save if we implemented prompt caching on our top 3 features?"

ROI and board prep

• "What is our AI ROI by department?"
• "How does our cost per support ticket compare to industry benchmarks?"
• "Which AI features can I confidently call 'investment' in the board deck?"

Example: investigating a cost spike

Here is a real example of an AI Advisor session investigating a 40% cost increase:

You: Why did our AI spend increase 40% last month?

AI Advisor: Your November spend was $8,420 vs $6,015 in October — a 40% increase ($2,405). The increase is almost entirely attributable to one feature: document-extractor, which grew 182% in spend (from $890 to $2,512). Two factors drove this:

1. Call volume increased 3× after your v2.1 deployment on Nov 3rd — this is legitimate growth, your document processing pipeline is being used more.

2. Average token count per call increased 47% — your v2.1 deployment appears to include more context in the prompt. This is likely contributing to slower latency too.

The other 17 features had normal month-over-month variance (±12%).

You: What can we do about the token count increase?

AI Advisor: Three options, ordered by impact:

1. Enable prompt caching — document-extractor sends repetitive system prompts. Cognocient has detected 68% prefix similarity across calls. Enabling Anthropic prompt caching would reduce cached-prefix costs by 90% — estimated saving: $840/mo.

2. Trim context before sending — the v2.1 prompt appears to include full document content even when only a section is needed. A preprocessing step that extracts only the relevant section before the API call could reduce average input tokens by 30–40%.

3. Route to a smaller model for simple extractions — 61% of document-extractor calls have low task complexity scores (< 2.5). Routing these to claude-haiku-4-5 would save approximately $620/mo with no quality impact on those calls.

Multi-scenario simulation

Beyond answering questions, the Advisor can run forward-looking simulations: "What would our bill look like if we applied all current routing recommendations?" or "Model three scenarios: 10%, 25%, and 50% usage growth next quarter."

Each simulation shows a before/after cost comparison with the assumptions made explicit. Simulation results can be exported as a PDF for budget planning meetings.

What data the Advisor can access

The AI Advisor has read-only access to:

• All call metadata: timestamp, model, token counts, cost, status, latency
• Attribution headers: feature name, department, session ID, user ID
• Waste analysis results and classification data
• Budget configuration and consumption data
• Recommendations history and applied/dismissed status
• Anomaly history and resolution notes

Next steps: Forecast · Feature Intelligence · Executive View