The direct answer: do not budget your first AI workflow as a software license. Budget a bounded operating change with a baseline, data, integration, design, evaluation, adoption, consumption, monitoring, and improvement. Without volume, quality requirements, systems, permissions, SLA, and an owner, any price is a sales assumption. A useful budget shows initial cost, recurring cost, cost per completed unit, and the decision each phase unlocks.
For a midsize company, the question is not “what does AI cost?” but “what will it cost to improve this workflow without losing control?” Choose one handoff, such as approved brief to launched campaign or eligible lead to sales acceptance. Lock the current process and the simplest non-AI alternative. Only then compare buy, configure, or build.
Use the First-Workflow Investment Ledger
- 1. Decision and baseline — volume, cycle time, rework, error, conversion, contribution, and current cost.
- 2. Data and integration — sources, permissions, cleaning, identity, APIs, latency, and fallback.
- 3. Design and delivery — discovery, simple rules, configuration, prompts, code, and documentation.
- 4. Evaluation and safeguards — tests, acceptance, human review, security, privacy, and proportionate red teaming.
- 5. People and adoption — owner, in-workflow training, process change, and launch support.
- 6. Software and technical operations — licenses, models, retrieval, storage, observability, and consumption.
- 7. Improvement and reserve — monitoring, fixes, versions, provider change, contingency, and shutdown.
The seven lines are a decision ledger, not universal percentages. Separate implementation from recurring cost and label each item required, evidence-conditioned, or deferrable. The FinOps Foundation recommends workload-level estimates that balance cost, quality, and performance. Google and Microsoft likewise treat returns, consumption, dependencies, operations, and architecture as cost components—not just model price.
Start with unit economics
Choose a unit marketing and finance both recognize: approved brief, quality asset launched, accepted lead, qualified conversation, or reactivated customer. Calculate cost per eligible unit completed, not cost per generation or token. Track human time saved and reinvested, corrections, rejects, integration failures, and opportunity cost. Productivity becomes value only when released capacity changes useful volume, speed, quality, or revenue.
Hypothetical example: a workflow handles 400 briefs per month. Automation reduces assembly time but review grows because inputs arrive incomplete. The ledger reveals that fixing intake and acceptance criteria costs more up front yet reduces rework. A license-only view looks cheap on paper and operates expensively. Use company data; this is not a benchmark.
Fund through three gates
- Gate 1 — problem proof: two weeks to measure baseline, exceptions, data, and a simple rule. Output: scope, comparator, and uncertainty.
- Gate 2 — workflow proof: four to six weeks in shadow mode and a limited cohort. Output: quality, time, failure, adoption, and observed unit cost.
- Gate 3 — operating proof: eight to twelve weeks with an owner, monitoring, rollback, and improvement cadence. Output: recurring cost, durable benefit, and a scale, hold, or stop decision.
Do not commit the annual budget in Gate 1. Release the next tranche only when prior evidence survives the comparator. Keep a simple rule when it is sufficient. Do not build when packaged software works without critical integration. Configure or build when the workflow depends on proprietary knowledge, volume, differentiation, and control. Revisit the choice as volume and requirements change.
Include the cost of operating
Name who owns failure, updates, quality, and use. Without an owner, a pilot decays; without observability, consumption and error grow silently; without fallback, work stops; without in-workflow enablement, adoption becomes a presentation. NIST organizes risk into govern, map, measure, and manage. Translate those outcomes into tasks and hours instead of a vague governance line.
Approval criteria for the CFO and CMO
- Scope — one unit, owner, population, SLA, and explicit exclusions.
- Comparator — current process and simplest sufficient alternative to test incremental value.
- Quality — criteria by error type, reviewed sample, and abstention rule.
- Economics — initial, monthly, unit, and switching cost; benefit as a range.
- Operations — support, alerts, human capacity, continuity, portability, and rollback.
- Decision — date, accountable person, and thresholds to stop, correct, or scale.
In the United States, include security review, privacy commitments, data location, sector obligations, procurement lead time, accessibility, and vendor insurance when the workflow touches them. Usage pricing changes with volume and model choice; fixed fees can hide caps, hours, or dependencies. Review only the surfaces the workflow actually uses.
Connect the investment to a real delivery
Use the 90-day starting plan at https://makinai.co/insights/en/where-start-ai-marketing-90-day-first-workflow, compare buy, configure, and build at https://makinai.co/insights/en/buy-configure-or-build-ai-solution, plan operations at https://makinai.co/insights/en/ai-marketing-pilot-production-operating-scale, and form a small delivery cell at https://makinai.co/insights/en/ai-marketing-implementation-cell-before-center-of-excellence. MAKINAI supports AI strategy and transformation at https://makinai.co/services/en/ai-strategy-transformation-consulting. Bring a workflow map, volumes, baseline, stack, and constraints; we can bound the smallest fundable scope without requiring a new platform.
There is no universal price or guaranteed return. The sources support workload-level cost estimation, value alignment, dependency-aware planning, and continuous risk operation. The ledger, gates, and criteria are MAKINAI editorial recommendations to validate against real data, architecture, and contracts.