For a 500-person org, the cost story is not employees asking questions. It is the organization repeatedly gathering, re-explaining, and re-sending the same documents and context before the model can do anything useful — and any automation set loose without a cap.
A normal assistant turn is roughly 15 input + 500 output tokens. At frontier list pricing (~$3 / $15 per million in/out), that turn costs about three-quarters of a cent. Round up to a penny and walk it forward.
The number only makes sense as unbounded context, repeated ingestion, and automation with no guardrails — reportedly an org that never set usage limits. That is a governance failure, not ordinary employee behavior. It is also the exact failure mode worth designing out before an SMB scales.
Pricing basis is frontier list price (Sonnet-class ~$3/$15, Opus-class ~$5/$25 per M in/out); exact figures and sourcing live in the companion pricing notes. Turn count assumes ~260 working days. The $500M figure itself is anecdotal — used as a failure pattern, not a citation.
These bands describe a real organization with duplication and friction — not a clean "everyone chats efficiently" world. The value is in showing where the meter runs, so leadership can manage the few behaviors that dominate spend. Confidence tags below follow the same logic as the benchmark hub: what's derived, what's directional, what's inferred.
| Band | Share | Behavior | $/user/day |
|---|---|---|---|
| L1 · Chat | 55% | Short prompts, Q&A, light drafting | $0.50–2 |
| L2 · Docs | 25% | Uploads, synthesis, repeated context setup | $3–12 |
| L3 · Formal | 15% | Long artifacts, redrafting, meetings-to-docs | $8–25 |
| L4 · Power | 5% | Heavy uploads, large contexts, proto-agents | $20–100+ |
The top two bands drive most of the spend and almost all of the surprise.
| Mode | Per day | Annualized |
|---|---|---|
| Mostly sane | $1.8k–4.5k | $0.47M–1.17M |
| Messy / bloated | $6k–18k | $1.56M–4.68M |
The difference between these two rows is not headcount and not model choice. It is whether context gets rebuilt constantly or captured once. That gap — roughly 3× — is the whole argument.
Scenario envelopes for a non-tech 500-person org, ~260 working days. Directional planning ranges, not a forecast or quote.
Vendors price like SaaS — a seat, a flat feel. The behavior is cloud: every additional ingest, retry, and output keeps the meter moving. Treating it as seat-spend is exactly why the bill surprises people.
Once it's read as utility compute, the right instinct follows automatically: don't pay to move the same data through the meter twice. Cache the expensive part. Govern the loops. Reserve the premium engine for the work that actually changes a decision.
The ingest layer, in one sentence: a centralized, indexed store of SMB's own knowledge that the model queries — instead of staff pasting raw files into a chat box every time. Build the working set once; reuse it everywhere.
It is not "build our own GPT" and it is not a GPU program. In phase one it is a small managed cloud service that prepares documents before Claude ever sees them.
A managed cloud service runs OCR, cleanup, classification, extraction, chunking, and first-pass summaries on small / cheap models.
It emits one standard payload: document type, key fields, short summary, relevant excerpts, links back to source.
Claude sees only the request plus the compact payload — and spends its tokens on reasoning, comparison, synthesis, and drafting.
Cost: extraction is done once and reused instead of re-billed per user, per chat. Accuracy: a model handed a tight, relevant payload is far less likely to get distracted or hallucinate than one fed fifty pasted documents. The discipline that saves money also improves the answers — which is why "we'll just be careful" is not a substitute for it.
How many times are we paying AI to remember what we already know?
Keep that question at the center and the spend follows organizational memory, not license counts. An SMB does not need to become an AI infrastructure company to get the benefit of AI infrastructure discipline.
Compact machine-readable summary for loading this memo into another model. House format — not RAG scaffolding.
{
"artifact": "decision_memo",
"subject": "SMB AI cost, context, ingest-layer architecture",
"audience": "COO + flat exec/VP leadership; single room",
"thesis": "Enterprise AI cost scales with repeated context reconstruction, not headcount; it behaves like cloud compute, not SaaS seats.",
"falsification": "~$0.01/turn implies ~50-67B turns to reach $500M, i.e. ~385k turns/person/workday over 500 staff -> impossible for humans -> spend is ingestion + unguarded automation.",
"cost_model": {
"user_bands": { "L1_chat":"55% / $0.50-2", "L2_docs":"25% / $3-12", "L3_formal":"15% / $8-25", "L4_power":"5% / $20-100+" },
"modes": { "sane":"$0.47M-1.17M/yr", "messy":"$1.56M-4.68M/yr" },
"basis": "messy-org scenario envelopes, ~260 workdays, frontier list pricing; directional not forecast",
"confidence": { "math":"derived", "bands":"directional", "modes":"scenario", "500M_anecdote":"inferred" }
},
"solution": {
"definition": "indexed knowledge store the model queries vs users pasting raw files",
"tiers": { "1":"cheap-model OCR/extraction/summary", "2":"structured payload handoff", "3":"Claude Enterprise for judgment" },
"levers": ["cost: preprocess once, reuse","accuracy: tight payload reduces distraction/hallucination"]
},
"ask": "one document-heavy workflow, 30-60 day instrumented pilot, expansion gated on cost-avoided + quality + rework",
"decision_question": "How many times are we paying AI to remember what we already know?"
}