AI Token Optimization

# 01 — what we audit

Every token your product ingests or emits is a line item. We read the line items.

• prompt architecture

  System prompts accrete. Instructions repeat. Three people added "be concise" in three different places. We deduplicate, modularize, and delete the paragraph nobody remembers writing.

• context hygiene

  Sessions that never end. Histories that never compact. Ninety-turn transcripts riding along on turn ninety-one. Fresh context per task; carry state, not scrollback.

• prompt caching

  Most teams cache a fraction of what they could and pay full price for the same prefix ten thousand times a day. Stable prefixes, cache-aware ordering, hit rates that justify the feature's existence.

• model routing

  Frontier model as the default for every request is a choice. Usually the wrong one. Cheap-first cascades, escalation on failure, verified by evals — not vibes.

• output controls

  max_tokens is not decorative. Structured outputs, stop sequences, and answers that end when they are done answering.

• retrieval efficiency

  Top-k set to 20 out of caution. Chunks overlapping 60%. Your RAG pipeline is buying the same paragraph five times per query.

• agent loops

  Tool definitions the length of novellas. Retry storms. Agents that re-read the file they just wrote. We cap the sprawl.

• spend observability

  Per-feature cost attribution, so "the AI bill went up" becomes a ticket with an owner instead of a mystery with a Slack thread.

# 02 — what we automate

Once the tokens are cheap, put them on payroll. We build, run, and audit agents for the back office; your headcount moves up the value chain instead of off the org chart.

• support

  Tier-one tickets are a language problem. Agents triage, resolve, and escalate with full context; the repetitive two-thirds gets handled, your humans keep the interesting third. CSAT holds or it doesn't ship.

• admin

  Scheduling, invoicing, expense chasing, data entry, the follow-up email about the follow-up email. Agents run it end to end, with audit logs your auditors can actually read.

• hr operations

  Onboarding paperwork, leave requests, policy questions, interview scheduling. The agent files the forms. It does not decide who gets hired, promoted, or let go. That line is load-bearing.

• what stays human

  Every workflow ships with scoped permissions, an escalation path, and a kill switch. Consequential calls — money above a threshold, anything touching a person's job — wait for a person. By design, not by accident.

# 03 — process

1. we talk

   One call. Your goals, your traffic, your bill. We ask for read access to the prompt-owning code and 2–4 weeks of usage logs. Redacted is fine; we audit tokens, not customers.

2. we open issues

   In your repo, with explicit code and prompt suggestions. Each issue names the offending tokens, the fix, and the measured delta. Diffs, not slideware.

3. we stay in communication

   Your engineers ship the changes on their own schedule; we review, re-measure, and re-run evals until quality parity holds. Ongoing support until the numbers settle.

4. we report

   Benchmarked comparisons showcasing the token spend before and after our recommendations — per request, per feature, per month — alongside the eval results proving nothing got dumber in the process.

Automation engagements add one step: shadow mode. The agent works your live queue alongside your team for thirty days. Nobody downstream notices. It cuts over only where it matches the humans — measured, then trusted.

# 04 — deliverables

• Annotated issues and pull requests with explicit prompt and code changes. Yours to merge; the reasoning travels with the diff.
• Benchmarked before/after: tokens per request, cost per request, cache hit rate, projected monthly spend.
• Eval report with a quality-parity gate. No recommendation ships without one. A cheaper wrong answer is not an optimization.
• Cache and routing configuration, documented well enough to survive the next reorg.
• Spend observability setup: per-feature attribution, budget alerts, a dashboard someone will actually open.
• Working agents for the workflows we automate — support triage, admin ops, HR paperwork — each with scoped permissions, an audit log, an escalation path, and an off switch.
• Shadow-mode report: agent versus human on your live queue — resolution rate, error rate, cost per task — before anything cuts over.
• A summary report your CFO can read in one sitting. Short. Like your new prompts.

# 06 — estimate

input tokens        ≈ 70% of spend           (inputShare = 0.70)
cached input reads  bill at 10% of list      (cacheDiscount = 0.90)
cheaper tier        bills at 20% of default  (routeDiscount = 0.80)

newSpend = S × [ 0.70 × (1 − trim) × (1 − 0.90 × cacheUplift)
               + 0.30 ] × (1 − 0.80 × routedShare)

savings display-capped at 70%.
anything above that, we'd rather prove than promise.

support savings    = tickets × cost × deflected
busywork savings   = hours × 4.33 wks × rate × automated
agents' token bill ≈ 10% of gross savings  (audited, so it shrinks)

net = support + busywork − token bill
consequential actions stay human. no slider changes that.

# 07 — pricing