How to Estimate Tokens Burned by Slack/Teams AI Agent Mentions Before Deployment

Why Pre-Deployment Estimation Matters

Anthropic's launch of Claude Tag in Slack on June 23, 2026, plus Microsoft's Copilot integration in Teams, makes "AI agent at the chat layer" the default deployment shape for B2B AI tooling. The catch: these integrations have an unusual cost surface. Per-mention costs feel small ($0.10-$1.00) but fan-out across an org makes the monthly bill arrive faster than expected.

Before you flip the deploy switch, you want a defensible estimate. This framework gets you within 15-25% of actual usage on day one and improves rapidly with real data.

Step 1: Estimate Per-Mention Token Anatomy

Every @bot mention in Slack/Teams expands into roughly the same five inputs:

A. Thread/channel context. 5K-50K tokens. Smaller for short threads, larger for long-running channels with rich history.

B. User profile + permission context. 1K-3K tokens. Username, role, channel permissions.

C. Tool descriptions. 2K-10K tokens. The agent's available tools (search, ticket creation, code execution, etc.).

D. Attached files (if any). 0-200K tokens. The biggest variable in the equation. A single attached PDF can dwarf the rest.

E. Output. 500-3K tokens for a typical reply. Higher for tool-call sequences.

Step 2: Compute Median Per-Mention Cost

Sum the bands above and apply your model's pricing. For Claude Sonnet 4.6 ($3 input / $15 output per M tokens), a median mention with no file attachment:

• Input: 25K + 2K + 5K = 32K tokens × $3/M = $0.096
• Output: 1.5K tokens × $15/M = $0.023
• Total: ~$0.12 per median mention

With a median file attachment (~50K tokens added):

• Input: 82K × $3/M = $0.25
• Output: 1.5K × $15/M = $0.023
• Total: ~$0.28 per file-heavy mention

Step 3: Estimate Mention Volume per User

The volume side has more variance than the per-mention side. From observed usage of similar Slack/Teams AI integrations:

• Light users (40% of team): 0-2 mentions/day
• Moderate users (40% of team): 3-8 mentions/day
• Heavy users (15% of team): 10-25 mentions/day
• Power users (5% of team): 50+ mentions/day (often async / pipeline-driven)

For a 20-person engineering team, that aggregates to roughly 80-150 mentions/day, or about 2,200-3,300 mentions/month.

Step 4: Account for File Attachment Mix

Across teams that have rolled out Slack/Teams AI agents, the typical mix is 80% no-attachment / 20% with-attachment. Plugging in:

• 2,640 mentions/month × 80% × $0.12 = $254 (no-attachment)
• 2,640 mentions/month × 20% × $0.28 = $148 (with-attachment)
• Subtotal: $402/month in synchronous mentions

Step 5: Add Async Task Cost

Modern Slack/Teams AI agents support async assignments — "@Claude, look into this and report back." These cost 5-10x a sync mention because the agent runs unsupervised and re-reads context multiple times.

For a 20-person team, expect roughly 1-3 async tasks per day, weighted heavily toward the 5% of power users:

• 2 async tasks/day × 22 working days × ~$5/task = ~$220/month

Step 6: Add Fan-Out Multiplier

Slack channels have an organic fan-out effect: one user's question triggers others to ask follow-up questions in adjacent threads, each costing a fresh agent run. Empirically, this adds 25-40% to the raw mention count, especially in support-style channels.

Apply a 30% multiplier to the synchronous mention subtotal: $402 × 1.30 = $523/month.

The Defensible Estimate

For a 20-person team rolling out a Slack/Teams AI agent on Sonnet 4.6:

• Sync mentions (with fan-out): $523
• Async tasks: $220
• Total monthly estimate: $700-$800/month

Add 25% buffer for the first month to account for novelty-driven over-usage and you should budget $900-$1,000/month.

Scaling to Larger Teams

The math scales roughly linearly with team size, with one nonlinear effect: fan-out grows faster than headcount because more people generate more cross-thread discussion. For a 100-person team, expect:

• Linear scaling baseline: $700-$800 × 5 = $3,500-$4,000/month
• Fan-out adjustment: add 15-25% on top = $4,200-$5,000/month

Three Estimation Pitfalls

Pitfall 1: Using launch-week data. Mention volume in week 1 is 2-3x typical because of novelty. Use week 4 data for forecasting, not week 1.

Pitfall 2: Forgetting context length growth. Channels that exist for months accumulate context. By month 3, your average input tokens per mention may be 30-40% higher than month 1. Build truncation into your context-fetching layer or budget for the drift.

Pitfall 3: Not accounting for model upgrades. Most teams plan to "upgrade to Opus for hard tasks." That upgrade typically affects 10-15% of mentions and increases their cost 5x — adding 35-65% to total spend. Factor it in if you anticipate using Opus for any subset.

Build-In Cost Controls Before Launch

Three controls every Slack/Teams AI deployment should ship with on day one:

• Per-workspace daily/monthly token cap (set at 1.5x your estimate)
• Per-channel cost dashboard (find the heavy users early)
• Smart routing: trivial queries to Haiku/Flash, complex ones to Opus

With these in place, a $1,000/month forecast can absorb 30-50% over-budget months without surprises. Without them, the same forecast can blow out 3-5x in the first quarter.