Where the AI budget is going, what's anomalous, and the highest-leverage changes to make. You're trending $25,130 over budget this month — we found $12,790 in recommended savings below.
Slice MTD spend by department, team, individual, agent, or model. Use this to pick the right place to dig in.
| Department | Headcount | Avg/person | $/mo | Share | Trend |
|---|---|---|---|---|---|
| Engineering | 184 | $3,293 | $605,912 | 69.1% | -6.8% |
| Product | 62 | $1,596 | $98,952 | 11.3% | +28% |
| Legal | 16 | $1,119 | $17,904 | 2% | +5.6% |
| Marketing | 54 | $739 | $39,906 | 4.5% | +38% |
| Sales | 96 | $428 | $41,088 | 4.7% | +13% |
| Operations | 142 | $424 | $60,208 | 6.9% | +9.4% |
| Finance | 38 | $221 | $8,398 | 1% | -3.2% |
| People | 28 | $179 | $5,012 | 0.6% | +15% |
Track AI app usage, agents, models, and cost across the organization. Use the overview to find the department or owner that needs the next drill-in.
| Application | Category | Active employees | Engagement | Licenses | % License usage | Annual cost | Potential savings | App Owner | Renewal due | Labels |
|---|---|---|---|---|---|---|---|---|---|---|
 ChatGPT Enterprise3 agents | AI assistant | 700 | 47% | 1,000 | 70% | $144K | $43.3K | sean.west@initech.com | Dec 15, 2026 | Engineering |
 Claude Team3 agents | AI assistant | 130 | 7% | 245 | 53% | Est. $58.8K | $9.2K | eng.tools@initech.com | Feb 2, 2027 | |
 Microsoft Copilot2 agents | AI assistant | 860 | 62% | 1,200 | 72% | $360K | $41.8K | it.admin@initech.com | Mar 1, 2027 | Sales |
 Gemini for Workspace1 agent | AI assistant | 540 | 44% | 780 | 69% | $187K | $26.6K | it.admin@initech.com | Jan 20, 2027 | |
 Perplexity EnterpriseDiscovered | AI research | 260 | 39% | 420 | 62% | $50.4K | $8.8K | strategy@initech.com | Aug 12, 2026 | Shadow AI |
 Cursor Business2 agents | AI coding | 184 | 78% | 260 | 71% | $99.8K | $12.4K | eng.tools@initech.com | Nov 5, 2026 | Engineering |
 GitHub Copilot1 agent | AI coding | 210 | 83% | 285 | 74% | $68.4K | $7.9K | eng.ops@initech.com | Oct 1, 2026 | Engineering |
 Glean1 agent | AI search | 445 | 58% | 620 | 72% | Est. $149K | $18.1K | knowledge.ops@initech.com | Apr 12, 2027 | |
 Writer | AI writing | 120 | 31% | 220 | 55% | $52.8K | $14.5K | marketing@initech.com | May 20, 2027 | Marketing |
 MidjourneyDiscovered | AI image | 64 | 22% | 120 | 53% | $14.4K | $4.2K | design.lead@initech.com | Jul 1, 2026 | Design |
 Fireflies.aiDiscovered | AI meeting notes | 340 | 21% | 520 | 65% | Est. $62.4K | $11.2K | sales.admin@initech.com | Sep 30, 2026 | Sales |
 Harvey | Legal AI | 28 | 36% | 44 | 64% | Est. $132K | N/A | legal.ops@initech.com | Jun 15, 2027 | Legal |
Each card is one specific change that reduces spend. Sorted by impact. Click to drill in — most are one-click apply.
94% of outputs would pass eval on Haiku 4.5, based on a 2,000-call shadow run. Latency would drop 41%, accuracy holds within 1.2pp.
70% of queries are single-table SELECTs that don't need a frontier model. A SQL-specialized router saves ~$2.3k/mo.
P50 useful content is 800 tokens but agent averages 6.2k tokens out — verbose JSON wrapping + repeated metadata.
Sends 18.2k tokens of system prompt every turn. ~9.4k bytes are stable across calls — eligible for prompt caching with no behavior change.
12 agents haven't been invoked in 90+ days but remain warm. Combined idle cost is from health-check pings + reserved capacity.
summarize_email, draft_summary, and email_recap do nearly identical work — Marketing, Sales, and Customer Ops each maintain their own copy.
Retries 4.2× on average due to upstream Stripe schema mismatch on `metadata.order_id`. Each retry is a full LLM call.
Runs 24/7 but 92% of consumers are 9–5 ET. Off-hours runs aren't read until next business day.
Sudden cost changes that warrant a look. Detected by comparing each agent / tool against its 14-day baseline.
Marketing Operations has already spent $18.7k against a $28k monthly budget and is running at $1.5k/day vs. the planned $920/day pace. If this continues, the team will exceed budget by $16.8k before month-end.
Switched from Haiku 4.5 → Sonnet 4.5 on deploy at 14:32 UTC. Previous behavior: $94/day. Now: $396/day. No corresponding increase in invocations.
Caller `lead-scorer-42` retried tool 11,420 times between 09:14–10:14 UTC. Probable retry loop on schema mismatch. Charged $238 for the hour vs. $14 baseline.
Agent ID `agt_8f2_88af` appeared, ran 4,200 invocations against Opus 4.7, and stopped. Owner: ad-hoc deploy by @diego.v. Not in inventory.
No new agents added. Existing `drafts_blog_posts` switched from caching enabled → disabled in deploy fbc23. Avg cost/call up from $0.014 to $0.052.
Detected for the 4th week running — Saturday 02:00–05:00 UTC sees 3.2× normal load. Likely an external scheduled trigger nobody owns.
At current burn rate, you'll exceed the March 2026 budget by $25,130. Apply the recommendations above to cut $12,790 recurring — that brings forecast EOM to $196,340, back inside budget.