| name | saas-agent-unit-economics-and-cogs |
| description | Use when producing or reviewing the saas agent unit economics and cogs component of a business plan; applies its specialist evidence, decisions, and acceptance tests instead of neighbouring pipeline skills. |
| metadata | {"portable":true,"compatible_with":["claude-code","codex"]} |
SaaS Agent Unit Economics & COGS Skill
Overview
Generic AI unit economics treats one user query as one LLM call. Agent unit economics is fundamentally different: one user request triggers a planner step, N worker steps, M tool invocations, possible retries, possible parallel branches, possible critic / supervisor calls, possible human-in-the-loop (HITL) escalations. A single resolved customer-service ticket can be 15-40 LLM calls and 5-15 tool invocations. A single "deep research" outcome can be hundreds of LLM calls. The cost waterfall is multiplicative, not additive, and investors will not accept generic AI cost models for agent products.
This skill installs the agent COGS discipline:
- Agent direct COGS — LLM tokens at each step (planner + workers + critic), tool invocations, external API cost, retry overhead, branch cost, HITL escalation cost.
- Agent indirect COGS — eval-loop cost, observability / tracing infrastructure, audit-log retention, sandbox / staging cost, supervision UX runtime.
- Agent reserves — irreversibility-incident reserve, regulator-engagement reserve, model-migration reserve.
- Agent overhead allocation — share of Agent Architect / Tool Engineer / Eval Engineer / AI Safety Lead / HITL Designer payroll.
Together these produce cost-per-task, cost-per-resolved-task (the true unit because unresolved tasks still cost money), agent gross margin, intervention-cost overhead, and the agent-cost-as-%-of-agent-revenue diagnostic.
Use When
- A SaaS / ICT plan includes an agentic or multi-agent product (single agent, planner-worker-critic, or vertical agent)
- Section 10 is being built for an agent-product plan
- Pricing must be set on per-resolution / per-outcome / per-step basis and needs cost floor
- Investors / DFIs have asked for cost-per-resolved-task specifically
- An existing AI-SaaS is moving up the autonomy ladder (assist -> suggest -> supervise -> agentic) and per-query economics no longer model reality
- The plan must pass
meta-agent-bankability-and-investor-readiness
Do Not Use When
- The AI feature is a single-shot completion (translate, summarise, classify) with no multi-step planning or tool use — use
saas-ai-unit-economics-and-cogs
- The product is internal-efficiency only with no customer-facing agent
- The business is not SaaS / subscription / usage-based recurring
Required Inputs
- Agent architecture description: single-agent or multi-agent; planner / worker / critic decomposition; tool registry (which tools, which providers, per-invocation cost)
- Per-agent-step model assignment (which step uses which model — frontier vs cheap-router)
- Average step count per task by task class (resolved, escalated, failed, abandoned)
- Average parallel-branch fan-out where applicable
- Retry policy and observed retry rate
- HITL escalation rate (% of tasks escalated to human; cost per HITL minute)
- Tool-invocation rate and per-tool cost (ERP, CRM, payment, KYC, mapping, OCR, internal API)
- Eval-loop spend per month (offline + online evals)
- Observability / tracing spend (LangSmith, LangFuse, Arize, Helicone, Weights & Biases, internal)
- Sandbox / staging compute spend
- Audit-log retention cost (storage + retrieval)
- Currency exposure (USD-denominated LLM + tool cost vs local-currency revenue)
Workflow
- Document the agent architecture — one paragraph + a step-and-tool diagram (or text decomposition: planner step -> N worker steps -> tool calls -> critic -> finaliser). Without this, cost modelling is fiction.
- Define the task taxonomy — by class (resolved, escalated-to-HITL, escalated-to-human-final, failed, abandoned, looped-and-killed). Each class has different cost.
- Build the per-task cost decomposition using
references/saas-agent-unit-economics-template.md:
- LLM cost per task = sum over steps of (input tokens x input rate + output tokens x output rate) x model-mix factor
- Tool cost per task = sum over tool invocations of per-invocation tariff (some tools are flat per call, some metered)
- External API cost per task (separate from tools where vendor lock-in or auth differs)
- Retry overhead = LLM cost x retry rate x average retry depth
- Branch overhead = (parallel-branch factor - 1) x base step cost (where applicable; multi-agent products)
- HITL escalation cost = (HITL minutes per escalated task x fully-loaded HITL hourly rate / 60) x HITL escalation rate
- Supervision overhead = supervisor / critic call cost per task
- Cost per task = sum of above. Compute weighted average across task classes.
- Cost per resolved task (the true unit) = total agent operating spend / number of resolved tasks. This is materially higher than cost-per-task because unresolved / failed / abandoned tasks still consume cost and must be amortised across the resolved outcomes. Investors will quote this number.
- Agent gross margin = (agent-attributable revenue - agent COGS) / agent-attributable revenue. Use revenue attribution discipline (per-resolution pricing tracks directly; tier-bundled agent revenue must be attributed honestly).
- Agent contribution margin per pricing model — for per-resolution, per-outcome, per-step, per-agent, hybrid. Identify which pricing models leave thin or negative margin.
- Apply mitigation levers in scenarios:
- Step compression — collapse planner + worker into single-shot for routine tasks (saves 30-50%)
- Model downshift on routine steps — route low-stakes steps to cheap-router model (saves 40-70% on those steps)
- Tool result caching — cache deterministic tool outputs (saves 20-40% of tool cost)
- Branch caps — hard-cap parallel branches and loop depth (prevents runaway; saves variable %)
- HITL escalation re-tuning — better triage reduces unnecessary HITL escalations (saves HITL minutes)
Quality Bar
- Agent architecture stated explicitly; planner / worker / critic / HITL decomposed
- Task taxonomy enumerated (resolved / escalated / failed / abandoned / looped-killed) with proportions
- Cost-per-task and cost-per-resolved-task both computed; the latter quoted as headline
- Agent gross margin separately computed from blended AI / SaaS margin
- Tool-cost line distinct from LLM-cost line (auditors and investors expect both)
- HITL escalation cost included; not assumed zero
- Sensitivity covers volume, pricing, intervention rate, failure rate, FX
- Stress scenarios cover provider 5x, tool outage, intervention 2x, irreversibility reserve, model deprecation migration
- Mitigation playbook explicit with expected impact per lever
- Living-plan cadence assigned with owners
- A sceptical Series A AI-agent-fund partner would not laugh at the model
Anti-Patterns
- "Agent cost is roughly the LLM cost" — ignores tools, retries, branches, HITL, supervision
- "We charge per-resolution so cost-per-task doesn't matter" — wrong; per-resolution requires knowing the cost floor or you set price below margin
- Cost computed on resolved tasks only — must amortise failed and abandoned across resolved
- "Tool calls are free" — most enterprise tools (CRM, ERP, payment, KYC) have per-call cost
- No retry-overhead line — production agents retry 10-30% of the time
- Single-model assumption when product needs router (frontier for plan, cheap for worker)
- "HITL is rare" — production agents typically escalate 5-25% depending on irreversibility class
- USD-only modelling when revenue is local currency
- No branch cap modelled — multi-agent products can blow cost budgets when uncapped
- "We'll optimise costs after launch" — optimisation is the business model in agents, not an afterthought
- Treating supervisor / critic calls as overhead, not COGS — they are COGS
Outputs
- Agent architecture description (one paragraph + step decomposition)
- Task taxonomy with class proportions
- Per-task cost decomposition (LLM + tool + API + retry + branch + HITL + supervision)
- Cost per task (weighted)
- Cost per resolved task (headline diagnostic)
- Agent gross margin
- Agent contribution margin by pricing model
- Sensitivity + stress analyses
- Mitigation playbook with prioritised levers
- Living-plan cadence assignment
SLA-COGS Treatment Subsection
When the agent product carries SLA commitments (uptime, accuracy, response time, definition-of-done), the cost waterfall must classify SLA-related cost lines correctly. The full discipline lives in saas-agent-sla-cogs-treatment/SKILL.md; the summary classification:
| Cost line | Classification | Rationale |
|---|
| HITL labour deployed to defend SLA (escalations from SLA-relevant cases) | COGS | Direct cost of delivering at SLA |
| SLA-monitoring infrastructure (uptime telemetry; accuracy telemetry; alerting) | COGS | Infra required to operate at SLA |
| Eval cost on SLA-relevant metrics | COGS | Quality cost of SLA |
| Retraining amortisation when retraining is SLA-driven | COGS | SLA upkeep |
| Redundancy / hot-standby compute for SLA uptime | COGS | Direct delivery cost |
| SLA credits issued to customers | Contra-revenue | Concession against subscription / usage / outcome revenue |
| Outcome-pricing refunds | Contra-revenue | Reversal of recognised revenue |
| Customer-success cost related to SLA defence (account management, comms) | S&M | Customer-success function |
| Legal cost defending SLA disputes | G&A | Indirect, dispute-driven |
| Insurance premiums covering SLA exposure | G&A (or COGS allocation) | Risk transfer; usually G&A |
Implications for agent gross margin:
- Agent revenue (gross) = booked agent revenue
- Less: SLA credits issued + outcome refunds (contra-revenue)
- = Agent revenue (net of contra-revenue)
- Less: Agent COGS including HITL-for-SLA + SLA infra + SLA evals + SLA-driven retraining amortisation + redundancy
- = Agent gross profit / agent gross margin
Common error. Classifying SLA credits as G&A or as a below-the-line item. This overstates gross margin and revenue; auditors will adjust.
Reserve link. SLA-credit reserve (saas-agent-deferred-revenue-and-credit-reserves) lives on the balance sheet; the P&L recognises the period's SLA-credit accrual as contra-revenue.
Cross-reference. saas-agent-sla-cogs-treatment/SKILL.md and saas-agent-sla-cogs-treatment/references/saas-agent-sla-cogs-policy.md.
Living-Plan Agent Cadence Defaults
| Element | Cadence | Owner | Variance threshold |
|---|
| Cost per resolved task | weekly | CFO + Head of Agent | +15% WoW |
| Cost per task (weighted) | weekly | CFO + Head of Agent | +10% WoW |
| Intervention rate | weekly | HITL Designer + Head of Agent | +3pp absolute |
| Task success rate | weekly | Eval Engineer | -3pp absolute |
| Tool-invocation reliability per tool | weekly | Tool Engineer | tool error rate >2% |
| Branch / loop ceiling breaches | weekly | Eval Engineer | any breach |
| HITL escalation cost share | monthly | CFO | >25% of agent COGS |
| Agent gross margin | monthly | CFO | -3pp MoM |
| Provider pricing watch | monthly | Head of AI / CTO | any change |
| Retry-overhead share | monthly | Tool Engineer | >15% of LLM cost |
| Tool-cost share | monthly | Tool Engineer | >30% of agent COGS |
| Model-migration reserve | quarterly | CFO + Head of AI | reserve drawdown |
| Irreversibility-incident reserve | quarterly | CFO + AI Safety Lead | any drawdown |
References
references/saas-agent-unit-economics-template.md — formulas, worked example, COGS waterfallreferences/saas-agent-cost-per-task-calculator-spec.md — calculator specskills/10-financial-projections/saas-ai-unit-economics-and-cogs/SKILL.md — parent AI unit economicsskills/07-marketing-sales-strategy/saas-agent-pricing-strategy/SKILL.md — pricing that protects agent marginskills/12-risk-analysis/saas-agent-risk-and-stress-test/SKILL.md — stress-test consumerskills/meta-agent-bankability-and-investor-readiness/SKILL.md — bankability scorecard consumerskills/meta-living-plan-governance/SKILL.md — governance disciplinebook-extractions/agent-products-business-plan-audit-2026.md — agent auditbook-extractions/ai-on-saas-business-plan-audit-2026.md — AI-on-SaaS auditbook-extractions/agent-sla-commercial-business-plan-audit-2026.md — SLA + commercial auditbook-extractions/mersch-hacking-saas-extraction.md — CFO-grade SaaS disciplineskills/10-financial-projections/saas-agent-sla-cogs-treatment/SKILL.md — SLA-COGS classificationskills/10-financial-projections/saas-agent-deferred-revenue-and-credit-reserves/SKILL.md — SLA-credit reserve, refund reserve, deferred revenueskills/10-financial-projections/saas-agent-revenue-recognition/SKILL.md — ASC 606 / IFRS 15 per pricing primitiveskills/10-financial-projections/saas-agent-sla-economics-in-projection/SKILL.md — SLA performance as projection driver
Africa / Uganda Application Notes
- Agent LLM + tool cost is USD-denominated; revenue often local currency. Per-resolution pricing must include FX corridor; the agent unit economics must hold across +/-20% FX swing (UGX 3,500-3,900/$, NGN 1,500-1,800/$, KES 128-145/$ as 2025/26 ranges).
- HITL cost in Uganda / Kenya / Nigeria / Rwanda is materially lower than US benchmarks (UGX 4,000-8,000/hour fully-loaded vs USD 30-60/hour US); this shifts the agent-vs-human economics — agents in Africa must beat a much lower human cost floor, and HITL fallback is cheaper to operate. Model both directions.
- Tool integration costs in Africa skew higher because of fragmented enterprise SaaS adoption — many tools are not yet API-first, requiring custom connectors. Model integration build cost as part of agent product cost, not as overhead.
- For agents serving WhatsApp / USSD / SMS / IVR channels, channel costs are real per-task line items (WhatsApp Business API per-conversation tariff; USSD aggregator per-session fee; SMS per-message; IVR per-minute). Add channel-cost line to per-task decomposition.
- In-region GPU for fine-tuned worker models (af-south-1, africa-south1, Liquid, Cassava, Raxio, Ethiopian AI Institute) prices 1.5-3x US/EU; model in-region inference premium if data residency is required.
- Cache-hit ratios in African vertical agents tend to be higher than US benchmarks because user task distributions are more concentrated (e.g. agri-extension agents see top 30 questions account for 60% of volume); model 40-60% cache.
- Audit-log retention costs in Africa can become binding when regulators (KE ODPC, NG NDPC, UG NITA-U, ZA Information Regulator) require multi-year retention; cost the retention period explicitly.
- Sovereign-AI procurement (KE Talanta, RW innovation, NG NITDA, ZA Presidential 4IR, EG infra) may require local hosting that raises per-task cost 30-80%; reflect in pricing if applicable.
July 2026 Portable Contract
Required Inputs
| Input artefact | Source/provider | Required | Behaviour when absent |
|---|
| Approved commercial assumptions, contracts, usage/cost evidence, accounting framework, opening position, and projection horizon for saas agent unit economics and cogs | Client records, approved operating model, finance owner, and accounting doctrine | Yes | If absent, contract terms, usage evidence, framework, or cost drivers are unavailable, isolate the affected schedule, label it unassessed, and do not force the model to balance with a plug. |
| Finalised business brief, target reader, country, and stage | Client intake and engagement owner | Yes | Stop section decisions and route the missing context to client intake. |
| Reconciled upstream assumptions that this section consumes | Named pipeline owners | Conditional | Record the dependency, affected claim, owner, and recovery step; do not substitute an invented value. |
Outputs
| Artefact | Consumer | Observable acceptance condition |
|---|
| Unit-economics model with task or tenant cost bridge | Plan author and target decision-maker | The artefact answers the section decision and traces each material conclusion to the supplied evidence. |
| saas agent unit economics and cogs exception and handoff note | Downstream section owners | Every blocked or conditional item names its consequence, owner, evidence request, and restart condition. |
| saas agent unit economics and cogs release record | Reviewer or plan assembler | Records the checks completed, failures, unassessed items, professional review required, and release state. |
Evidence Produced
| Evidence | Format | Acceptance condition |
|---|
| Formula trace, source/assumption register, three-statement or schedule reconciliation, and finance-gate record | Source-linked table, calculation, or annotated prose | The evidence is reproducible from named inputs and distinguishes verified fact, management assumption, and inference. |
| saas agent unit economics and cogs decision record | Decision note | States the selected action, rejected credible alternative, countercase, rationale, and risk accepted or avoided. |
| saas agent unit economics and cogs review trace | Gate entry | Identifies the date, input versions, reviewer role, failed checks, recovery owner, and any check that remains not assessed. |
Capability and Permission Boundaries
For saas agent unit economics and cogs, the controlling focus is cost per completed task, attempted-versus-completed usage, tool costs, human review, and contribution margin. This skill may inspect records and calculate planning scenarios in read-only mode; it may not post entries, change ledgers, set accounting policy, certify IFRS treatment, or release statutory values without authorised professional review. Its normal mode is read-only analysis and drafting. Any mutation, external communication, spending, certification, or professional conclusion outside that boundary requires explicit authority and must remain traceable to the approving role.
Degraded Mode
For saas agent unit economics and cogs, loss of evidence about cost per completed task, attempted-versus-completed usage, tool costs, human review, and contribution margin activates degraded mode. If the controlling saas agent unit economics and cogs evidence is unavailable, the same boundary applies. When contract terms, usage evidence, framework, or cost drivers are unavailable, isolate the affected schedule, label it unassessed, and do not force the model to balance with a plug. Return the verified subset, label the affected decision qualified or not assessed, explain the downstream consequence, and state the smallest evidence request or authorised action that permits recovery. Do not convert the missing check into a pass.
Decision Rules
| Choice or condition | Action | Failure or risk avoided |
|---|
| For saas agent unit economics and cogs, commercial billing, cash receipt, service delivery, and accounting recognition occur in different periods | model each event separately, reconcile the bridge, and route judgemental treatment to the finance reviewer | Cash, revenue, liability, and margin can be conflated into a misleading forecast |
| For saas agent unit economics and cogs, A current legal, regulatory, tax, accounting, market, or platform claim controls the saas agent unit economics and cogs decision | Verify the controlling source, effective date, jurisdiction, and reviewer status before release | Stale external facts become permanent plan assumptions |
| For saas agent unit economics and cogs, The evidence reconciles with neighbouring sections and the countercase does not overturn the choice | Complete unit-economics model with task or tenant cost bridge, attach the evidence and release record, and hand off named dependencies | Premature release and repeated downstream rework |
Workflow
- Define the exact saas agent unit economics and cogs decision, intended reader, jurisdiction, business stage, and permission boundary.
- Collect approved commercial assumptions, contracts, usage/cost evidence, accounting framework, opening position, and projection horizon and map each material conclusion to its source; stop the affected conclusion when an input could change it.
- Apply the specialist methods and directly linked references already contained in this skill, retaining its domain thresholds, calculations, and Uganda or East Africa context where applicable.
- Compare the credible alternatives, test the countercase and failure path, and apply the decision table rather than selecting a template default.
- Produce unit-economics model with task or tenant cost bridge with the evidence, exception, and handoff records; reconcile every shared assumption with its owning section.
- Run the section quality checks, applicable finance or professional review, and anti-slop gate. If a gate fails, correct the evidence or decision and return to the responsible step.
Quality Standards
- Unit-economics model with task or tenant cost bridge must answer a real decision for the named bank, investor, DFI, grant, board, or strategic-partner reader.
- Formula trace, source/assumption register, three-statement or schedule reconciliation, and finance-gate record must be source-linked, dated where facts can change, and sufficient for another reviewer to reproduce the conclusion.
- The section exposes its countercase, stop condition, recovery action, and effect on neighbouring sections.
- No unavailable source, calculation, tool, or professional review is reported as passed; finance and statutory judgements follow the governing doctrine.
- Language remains specific to saas agent unit economics and cogs, uses British English naturally, and passes the repository anti-slop gate without promotional filler.
Anti-Patterns
- In saas agent unit economics and cogs, treating an unavailable approved commercial assumptions, contracts, usage/cost evidence, accounting framework, opening position, and projection horizon as confirmed. Correction: qualify the affected conclusion and issue the named evidence request.
- Producing unit-economics model with task or tenant cost bridge that restates the brief but makes no choice. Correction: record the choice, rejected alternative, rationale, countercase, and implication.
- Ignoring a conflicting upstream assumption. Correction: return it to its owning section and resume only from a reconciled version.
- Reporting an unavailable check as passed. Correction: mark it not assessed and narrow the release state.
- Claiming compliance, assurance, bankability, or investor readiness from narrative quality. Correction: run the applicable gate and retain its evidence.
- Copying the worked example into a client plan. Correction: use the method only and replace every fact with verified engagement evidence.
Worked Example
An agent attempts three tool calls for every completed task and sends 15% to human review. Cost attempts, completion, tool fees, and reviewer time separately before setting the contribution-margin threshold.
References
- Use the verified project evidence register and the owning upstream pipeline section for saas agent unit economics and cogs; no local deep-dive reference is declared.
- For saas agent unit economics and cogs claims involving money, tax, grants, reserves, revenue, cost, valuation, or financial statements, apply the Chwezi finance doctrine and record the required professional-review state; illustrative figures never become client facts.
