Executive summary (read this first)

AI SaaS pricing fails when it sells compute instead of outcomes. In 2025 the winning model is: price on value (tasks, workflows, SLAs), cap risk with clear entitlements and rate limits, control COGS with routing/caching/prompt hygiene, and prove ROI with dashboards that show cost per successful answer/task and time saved.

What to do:

1. Choose your pricing primitive by use‑case: seat + usage for horizontal assistants, task‑based for agents that do work, or plan credits (messages/tasks) with soft warnings and hard caps.
2. Build a bottom‑up cost model that includes tokens, retrieval, rerankers, vector queries, logging, and support. Track cost per successful answer/task—not just tokens.
3. Package value, not knobs: Good/Better/Best plans with entitlements (regions, BYOK, approval workflows, longer retention, premium SLAs).
4. Instrument everything: per‑tenant usage meters, alerts at 70/85/95%, and an in‑app usage page.
5. Run safe experiments: canary new prices, grandfather existing customers, and raise list prices only when your outcome metrics improve or COGS drops sustainably.
6. Aim for healthy margins: short‑term target ≥65–75% gross margin, expanding to ≥75–85% as routing, caching, and retrieval quality mature.

This guide gives you a copy‑and‑paste framework—formulas, plan structures, regional notes (AU/US/UK/NZ), and negotiation tactics—so your AI SaaS is profitable, predictable, and buyer‑friendly.

What you are really selling (it is not tokens)

• Assistants: trusted answers with citations, fast.
  Sell: seats + message/answer credits, accuracy/latency SLAs, and admin controls.
• Agents: completed tasks (refunds raised, cases resolved, records updated) with approvals and audit.
  Sell: task bundles/credits, role‑based approvals, audit exports, and outcome guarantees.
• Platforms: governed AI capabilities (RAG/Graph‑RAG, tool registry, evaluation) for internal builders.
  Sell: workspace limits, model routing options, region pinning, BYOK, and observability.

Rule of thumb: If the value is human time saved, use task credits or assistant answers with caps. If the value is access to governed capabilities, use seat + platform limits.

Pricing primitives (choose one primary; mix sparingly)

Pick one primary driver (seats / answers / tasks) and add 1–2 secondary levers (retention, residency, SLAs). More levers = more confusion.

The unit‑economics model (your single source of pricing truth)

Define your cost per successful answer (assistants)

COGS_answer=Tokens+Retrieval+Rerank+Vector DB+Logging/Obs+Support\textbf{COGS\_answer} = \text{Tokens} + \text{Retrieval} + \text{Rerank} + \text{Vector DB} + \text{Logging/Obs} + \text{Support}COGS_answer=Tokens+Retrieval+Rerank+Vector DB+Logging/Obs+Support

Where:

• Tokens = (prompt + completion) × $/token by route.
• Retrieval = vector queries × $/query + metadata/keyword cost.
• Rerank (optional) = reranker calls × $/call.
• Vector DB = memory/IO amortised + per‑query charges.
• Logging/Observability = trace storage + analytics share.
• Support = proportion of human review time and tier‑1 support load.

Price floor per answer = COGS_answer ÷ target gross margin (e.g., 0.7 for 70%).
Price per 1,000 answers = 1000 × price floor, rounded sensibly.

Define your cost per successful task (agents)

COGS_task=Planning tokens+Retrieval stack+Tool calls+Approvals+Logging/Obs+Support/Exceptions\textbf{COGS\_task} = \text{Planning tokens} + \text{Retrieval stack} + \text{Tool calls} + \text{Approvals} + \text{Logging/Obs} + \text{Support/Exceptions}COGS_task=Planning tokens+Retrieval stack+Tool calls+Approvals+Logging/Obs+Support/Exceptions

• Tool calls include third‑party API fees and your own compute.
• Approvals cost is human minutes × loaded rate / expected approval pass rate.
• Support/Exceptions accounts for rework and “route to human”.

Price floor per task = COGS_task ÷ target margin.
Bundle as task credits per plan (e.g., 1,000 tasks/month with soft caps).

Track reality weekly: cost per successful answer/task by feature and by tenant. This is the KPI your pricing levers must protect.

Worked examples (planning‑grade numbers)

The numbers below are illustrative to show the method. Replace with your own telemetry.

Example A — Knowledge assistant (per‑answer economics)

• Average tokens (including retrieval context): 1,200
• Retrieval: 3 vector queries + 1 keyword query
• Reranker: 0.4 calls per answer (only on risky queries)
• Vector DB amortised per answer: A$0.002
• Logging/Obs per answer: A$0.001
• Support share per answer: A$0.0015

Assume blended token rate yields A$0.006 for 1,200 tokens; retrieval A$0.0015; rerank A$0.001 on average.

COGS_answer (AU) ≈ 0.006 + 0.0015 + 0.001 + 0.002 + 0.001 + 0.0015 = A$0.013
At 70% margin, price floor ≈ A$0.043/answer. Round to A$45 per 1,000 grounded answers in Pro tier.

Indicative list by region (same margin target, similar telemetry):

(Use your FX and costs; maintain one global psychological anchor per market.)

Example B — Refund agent (per‑task economics)

• Planning & validation tokens: 900 total
• Retrieval (policy + order): 4 queries
• Tool calls: 2 (create case, set refund) – your infra + payment API fee US$0.03
• Approval time: 20s average @ US$60/hr; pass rate 92%
• Logging/Obs per task: US$0.004
• Exceptions/rework: 8% at +US$0.15 each

COGS_task (US) ≈ tokens US$0.005 + retrieval US$0.002 + tools US$0.03 + approvals US$0.333 × (20/3600)/0.92 ≈ US$0.002 + logging US$0.004 + exceptions US$0.012 → US$0.055 (rounded)

At 70% margin, price floor ≈ US$0.183. Sell 1,000 tasks at US$199–$249 depending on SLA, with overage at US$0.25/task.

Packaging: plans that sell value (and protect margin)

Good / Better / Best with a clear value ladder:

Add‑ons that enterprise buyers understand:

• BYOK (bring your own key)
• Private networking (VPC peering/private link)
• Dedicated data plane (separate DB/vector index)
• Extra region replicas (latency/residency)
• Longer retention (365+ days)
• Premium SLAs (response/uptime/rollback windows)
• Custom tools/approvals (agent expansions)

Keep the menu short. The more levers you expose, the harder procurement becomes.

Entitlements, caps, and alerts (bill‑shock prevention)

• Counters: seats, grounded answers, successful tasks, vector queries, storage, regions, API calls.
• Soft warnings: email + in‑app at 70/85/95% usage.
• Hard caps: temporary throttling with upgrade path; admin override available.
• Fair usage: spell out what’s included (no scraping/public bulk jobs on Starter, etc.).
• Grace windows: small buffer for end‑of‑month spikes to avoid disruption for good customers.

Show an in‑app usage page with trends, forecast to month‑end, and the cost per successful answer/task so buyers see value, not just consumption.

Regional notes: Australia, USA, UK, New Zealand

• Australia (AU): Buyers often ask for AU data residency, clear GST on invoices, and plain‑English data handling. Consider A$ list prices with task/answer packs that map cleanly to local budgets.
• United States (US): Strong appetite for usage‑based pricing and annual prepay discounts. Be explicit about overage rates and SLA penalties.
• United Kingdom (UK): UK‑based residency for public sector; pricing typically shows ex‑VAT with VAT handling in invoices. Procurement expects plan comparators and predictable overage.
• New Zealand (NZ): Privacy expectations similar to AU; consider NZD plans or show localisable pricing. Many SMEs run Xero—task packs that align with finance workflows resonate.

Where residency matters, make it a plan toggle (Pro+): “Keep data in AU/UK/US/NZ region” with a short explainer.

Controlling COGS: seven levers that move the needle

1. Prompt hygiene – shorten and modularise system prompts; remove repeated lore; use named templates with variables.
2. Routing – send classification/formatting to smaller models; reserve frontier models for heavy reasoning.
3. Caching – response cache for common Q&A; embedding cache keyed by content hash; target ≥30% hit rate.
4. Retrieval quality – clean ingestion, one‑intent chunks, hybrid search with metadata filters; fewer, more relevant tokens in context.
5. Selective reranking – only on high‑risk queries; don’t apply universally.
6. Vector DB tuning – PQ/quantisation to cut memory; measure recall before committing.
7. Latency budgets – early cut‑offs for low‑value expansions; avoid token‑hungry wandering.

As COGS falls, choose a mix of margin expansion and price improvements to boost competitiveness.

Discounting & negotiation (without destroying margin)

• Discount bands: publish internal floors (e.g., 10% for annual prepay, 15–20% for multi‑year/volume; >25% requires CFO sign‑off).
• Trade for value: discounts in exchange for case study, reference calls, or design‑partner feedback.
• Protect your floor: if buyers want steeper cuts, adjust entitlements (fewer tasks/answers, single region, standard SLA).
• Avoid “all‑you‑can‑eat”: unlimited tokens/tasks is a margin time‑bomb. If demanded, set fair‑use and per‑minute rate limits.
• Anchor on outcomes: remind buyers you charge for successful answers/tasks with governance (approvals/audit), not raw compute.

Price changes & grandfathering

• Grandfather customers on old plans for 12–24 months where possible; offer upgrade credits to encourage migration.
• Announce changes with 60–90 days’ notice; share a why: improved accuracy, new governance features, expanded SLAs, or sustainably lower COGS re‑invested into scale/reliability.
• Never surprise: warn proactively via in‑app usage views and emails.

Billing operations (that make finance love you)

• Metering integrity: immutable event stream; idempotent counters; backfill tools.
• Reconciliations: monthly checks that answers/tasks match invoices; sample audits.
• Refund logic: credit when failures were your fault (downtime, regression) to build trust.
• Collections: automated dunning with human handover for enterprise.
• Tax: show GST/VAT as applicable; keep records per region.

Price & packaging templates (copy and adapt)

Template A — Assistant‑first product

• Starter (A$149/m): 5 seats, 5,000 grounded answers, single region, 30‑day retention, standard SLA.
• Pro (A$799/m): 50 seats, 50,000 grounded answers, 2 regions, 180‑day retention, priority SLA, approvals + audit export.
• Enterprise (Quote): unlimited seats, custom answers, multi‑region, BYOK, private networking, premium SLA, DPA.

Overage: A$0.06 per 10 grounded answers on Starter; A$0.05 on Pro. Task packs optional add‑on.

Template B — Agent‑first product

• Starter (US$299/m): 1,000 task credits, 5 seats, single region, approvals, 90‑day audit trail.
• Pro (US$1,499/m): 8,000 task credits, 25 seats, 2 regions, audit exports, premium support.
• Enterprise (Quote): custom credits, multi‑region, BYOK, private networking, custom approvals, premium SLAs.

Overage: US$0.25/task; volume packs discount to US$0.18–0.20.

(Adjust numbers to your telemetry and margins.)

Forecasting revenue (sanity checks before launch)

• Adoption curve: conservative (P50) and optimistic (P80) scenarios for active seats, answers/tasks per seat, and expansion %.
• Churn & expansion: logo churn 2–4% monthly on SMB, net revenue retention targets 100–115% with task pack upsells.
• COGS trend: assume −20–35% token spend per answer over 2–3 quarters via routing/caching/retrieval gains.
• Break‑even: find the task/answer volume where your gross margin crosses 70% and your CAC payback < 12 months.

Ethical & customer‑friendly price experiments

• Canary new plans for 5–10% of new sign‑ups; never alter price mid‑contract.
• A/B only on new prospects; clearly disclose list vs promotional pricing.
• Holdout: keep 10–20% of leads on existing pricing to sanity‑check conversion and LTV.
• Measure: activation, conversion, ARPA, gross margin, support load, refund rates.

FAQs (for buyers and stakeholders)

Q: Why not just charge for tokens?
Because tokens are inputs, not outcomes. Buyers want predictable value. Price on grounded answers or completed tasks, with transparent caps.

Q: How do we stop abuse on low‑tier plans?
Set rate limits, IP/app restrictions, and soft/hard caps with in‑app visibility. Educate users on overage options.

Q: Our tasks vary in complexity—how do we price fairly?
Use task classes (Standard vs Premium) or weightings (e.g., refund with approval = 1.2 credits). Keep to two classes maximum to avoid confusion.

Q: Should we include fine‑tuning in plan fees?
Offer fine‑tunes as professional services or Enterprise add‑ons with clear scope. Don’t bake volatile work into base pricing.

Q: What margin should we target?
Short‑term ≥65–75% gross margin is realistic for AI SaaS; target ≥75–85% as routing/caching/retrieval mature.

Q: Can we promise fixed costs?
You can promise capped costs via entitlements and fair‑use. Internally, manage variability with routing, caching, and strict evaluation to prevent regressions.

Q: What if our COGS spikes after a model change?
Freeze rollouts behind canaries; auto‑rollback if cost per successful answer/task breaches thresholds. Communicate with customers if SLAs or pricing are affected.

Where Digital One Agency fits

If you want pricing that sells value and protects margins, we can help you:

• Instrument usage meters (answers, tasks, storage, regions) and in‑app usage pages.
• Model cost per successful answer/task and design routing/caching to shrink COGS.
• Build plan logic, entitlements, caps and alerts, and billing integrations.
• Ship RAG/Graph‑RAG assistants with citations/refusal and agent workflows with approvals/audit.
• Package enterprise add‑ons: BYOK, private networking, dedicated data plane, premium SLAs.
• Run price experiments safely with canaries and clean analytics.
• Provide finance options so you can build now and pay across milestones if eligible.

CTA

Book a Free 1‑Hour Strategy Session
👉 Book a meeting
📞 Call us globally 24/7: +61 2 9057 9079
📧 Email: contact@digitaloneagency.com.au