Direct answer: there is no responsible price for “AI consulting” until the outcome, data, integrations, risk level, usage volume and operating responsibility are defined. The real budget combines professional services, data preparation, build and integration, evaluation and security, infrastructure and consumption, organizational change, and capability transfer. To compare proposals, give every provider the same scenario and require cost by layer, phase and assumption. The final metric should be total cost per accepted outcome—not hourly rate or token price.
Two proposals with the same initial price can produce very different economics. One may include testing, observability, enablement and support; another may stop at a demonstration. A low bid can also hide dependencies: unprepared data, missing APIs, human review, licenses, model usage, production environments or maintenance. Comparison starts when the buyer makes these assumptions explicit.
The MAKINAI Total AI Delivery Cost model
Divide the budget into seven layers. Each provider should state what is included, the billing unit, the primary assumption, the range of variation and who owns the work after the engagement. This allows buyers to compare a strategy consultancy, a development company and a platform without confusing materially different scopes.
1. Discovery, strategy and product definition
Include problem framing, user research, prioritization, workflow design, initial architecture, risk assessment, business case and success criteria. This should not be a generic slide phase; it must close open decisions. Cost increases when stakeholders are numerous, rules conflict, markets differ or no process owner exists. The primary output is a testable definition of the outcome and a clear statement of what will not be built.
2. Data and knowledge
Account for inventory, access, cleaning, transformation, labeling, permissions, quality, knowledge retrieval, retention and refresh. A model may be simple while the data work is substantial. Require separate estimates for initial preparation and recurring maintenance. If documents, catalogs or histories change frequently, refresh cost belongs in operations rather than a future exception.
3. Build, experience and integration
This layer covers engineering, prompts, orchestration, interfaces, APIs, automations, identity, queues, connectors and systems of record. The number of systems, API quality and real-time action requirements often change effort more than screen count. Require a named list of integrations, environments and responsibilities. “Integration included” without endpoints, limits and tests is not a comparable estimate.
4. Evaluation, security and governance
Budget for test cases, evaluation datasets, risk-proportionate red teaming, privacy, authorization, logs, human approval and launch criteria. NIST organizes AI risk work through Govern, Map, Measure and Manage; these are lifecycle activities, not one final review. A provider that excludes evaluation can appear cheaper by transferring the cost of discovering failures to the client.
5. Infrastructure, models and consumption
Separate development and production costs. Include input and output tokens, embeddings, retrieval, storage, networking, agent runtime, external tools, observability and environments. Cloud providers support pay-as-you-go and reserved-capacity models; the choice depends on volume, predictability and service level. Do not freeze the model in a temporary price table. Estimate units of usage and refresh current prices before approval.
6. Operations, adoption and change
Include support, human review, training, communications, process updates, quality monitoring, incidents and continuous improvement. A system nobody adopts incurs cost without return. UK AI procurement guidance recommends planning for ongoing support, training, knowledge transfer and lifecycle management during procurement. Ask who responds when a model, API or source dataset changes.
7. Ownership, transfer and exit
Estimate documentation, runbooks, repositories, data export, licenses, rights to code and evaluations, internal training and transition to another provider. A lower price with an expensive exit can have a higher total cost. Define what happens to data and artifacts at contract end. Exit should be testable, not merely a legal clause.
How to estimate timeline without inventing a date
Split the schedule into four gates rather than one final deadline. Gate 1 validates the problem, data and success criteria. Gate 2 produces a proof of value with representative cases. Gate 3 adds security, integration, evaluation, observability and operations. Gate 4 scales volume, teams and adoption. Each gate ends with evidence and a decision to proceed, adjust or stop. Uncertainty becomes a managed decision rather than a hidden margin.
- For each gate request: expected duration and likely range; client dependencies; team size and composition; entry and exit criteria; deliverables; one-time and recurring costs; risks that change timing; impact of delays in data, security and integration.
Request two forecasts: a likely date when dependencies are met and a conservative date that includes known risks. Do not accept a contingency percentage without the risk it covers. A mature provider shows the critical path and distinguishes parallel work from approvals that block progress.
Three scenarios for production cost
Model low, base and high usage with the same variables: tasks per month, average input and output size, tool calls, retrieval queries, retry rate, storage, latency, human review and service level. AWS documents that aggregated cost reports and invocation records serve different allocation needs; the FinOps Foundation’s FOCUS specification normalizes cost and usage data across vendors. The practical requirement is to instrument cost by workflow and outcome during the pilot.
The proposal sheet that creates comparability
- Outcome and acceptance metric; scope and exclusions; seven cost layers; four delivery gates; provider and client effort; one-time and monthly costs; three volume scenarios; pass-through licenses and consumption; change pricing; support and SLA; ownership and exit; cost per completed task; assumptions that invalidate the proposal.
Normalize currency, taxes, duration, volume and service level. Then compare three numbers: investment to the first validated outcome, total first-year cost and unit cost in operation. The lowest first number is rarely enough. Weight quality, security and transfer capability alongside price.
Red flags
- A fixed price without data or integration assumptions; a schedule that ends at prototype but is described as production; cloud usage dismissed as immaterial; security and evaluation treated as optional; no estimate of client effort; support without an SLA; undefined ownership; discount tied to long lock-in; ROI without a baseline or adoption rate.
Next step
Before asking for prices, give every provider the same use case, volumes, systems, risk and acceptance criteria. Use MAKINAI’s RFP guide: https://makinai.co/insights/en/how-to-write-rfp-ai-services. Decide the delivery model with https://makinai.co/insights/en/in-house-ai-team-or-ai-consulting-firm. Compare capability using https://makinai.co/insights/en/how-to-choose-ai-implementation-company-brazil-scorecard. If you need to structure the budget, roadmap and delivery model, see https://makinai.co/services/en/ai-strategy-transformation-consulting.