A typical AI consulting engagement for a mid-market company runs between $25,000 and $250,000 depending on scope, phasing, and who is doing the work. What separates a $40K pilot from a $200K implementation is not the vendor's brand — it is what each phase requires your team and theirs to deliver. This article breaks that down.
The Three Engagement Types and Their Typical Scope
Most AI consulting work falls into one of three structures. Understanding which one a proposal represents is the first thing a CTO or Head of Data should confirm before discussing price.
Discovery and Assessment (4–8 weeks)
This phase answers a single question: Is this problem ready to be solved with AI? The output is not a system — it is a prioritized use-case map, a data readiness assessment, and a rough architecture proposal. A discovery engagement is time-boxed and fully self-contained; it does not commit you to a build phase.
Discovery engagements are where gaps surface early. If your data is fragmented across five systems with no consistent identifiers, that finding alone is worth the cost — it prevents a failed build at ten times the price. See our piece on AI readiness assessment for companies for the specific criteria that determine whether a team is ready to build.
Proof of Concept / Pilot (8–16 weeks)
A pilot builds one scoped feature or workflow against your actual data and infrastructure. The goal is a working demonstration in production-like conditions, not a polished product. A well-run pilot has a binary verdict: this works and the path to production is clear, or it does not and here is why.
The key risk at this stage is scope creep. A pilot that starts as a forecasting model for inventory and grows into an integrated planning dashboard is not a pilot anymore — it is an unbudgeted build. Locking scope before kickoff is the single most effective cost-control measure in this phase.
Full Production Implementation (4–12 months)
A full implementation takes an approved architecture from the discovery phase and builds it to production quality — with monitoring, failover, documentation, and a trained internal team to maintain it. This is where data infrastructure decisions made earlier have the largest downstream cost. The difference between a clean data pipeline and a messy one can shift total engagement cost significantly. The analysis in Snowflake vs. BigQuery cost for companies covers how platform selection affects long-term spend.
Phase by Phase: What You Are Actually Paying For
Across all three engagement types, the work breaks into four categories of effort. A proposal that does not itemize these is not a proposal — it is a ballpark.
Architecture and System Design
Decisions made here lock in cost for the rest of the engagement. Which models? Which vector store or data warehouse? Streaming or batch? These choices affect not just the build cost but the monthly infrastructure bill after the consultants leave.
Data Preparation and Integration
This is consistently the most underestimated line item. Cleaning, normalizing, and connecting data sources to the AI system can account for the largest single block of effort in organizations without a mature data layer. Proposals that do not include this explicitly are assuming your data is already production-ready — an assumption worth challenging in writing before signing anything.
Model Development and Integration
This covers selecting or fine-tuning models, building the application layer — APIs, agents, workflows — and integrating the system with your existing tools. In most mid-market engagements, model selection is less important than integration quality. A well-integrated general-purpose model outperforms a custom-built one that lives in a disconnected environment.
Testing, Handoff, and Enablement
A production system that your team cannot maintain is a liability, not an asset. Handoff covers technical documentation, runbooks, and deployment procedures. Enablement covers training the people who will operate and evolve the system. This phase is where engagements most commonly get cut short — and where the cost of that decision shows up six months later.
The Line Items Nobody Puts in the First Quote
Four cost categories appear in nearly every engagement but are absent from most initial proposals:
- Cloud infrastructure: Model inference, vector storage, and data pipelines generate cloud spend from day one of the pilot. This is typically excluded from consulting quotes and billed directly to you by the cloud provider.
- Model licensing: If the engagement uses a licensed frontier model, the API cost is yours, not included in the consulting fee.
- Data cleanup: If your source data requires significant normalization before it can be used, this work adds time and cost. Ask for a line-item estimate after the discovery phase, not before.
- Internal team time: Subject-matter experts, IT staff, and managers who support the engagement are not free. A realistic budget accounts for a meaningful share of the consulting fee in internal resource time.
Ten Questions to Ask Before You Sign
- Does this proposal include a discovery phase, or does it assume the architecture is already defined?
- What does data readiness mean for this engagement — and what happens if we do not meet it?
- What deliverables mark the end of each phase, and who approves them?
- Which costs are excluded from this quote: cloud infrastructure, model licensing, data cleanup?
- Who owns the code, models, and documentation at the end of the engagement?
- What is the escalation path if a phase runs over scope?
- How is the engagement structured — sprints with weekly demos, or milestone-based delivery?
- What does the handoff process include, and how long does it take?
- What does a successful pilot look like — what is the binary test that says this works?
- What has to be true on our side before work begins?
If you want to explore how Rankea structures AI engagements for mid-market companies, the software development and AI services page covers the model and what to expect from the first conversation.
Frequently Asked Questions
Is a discovery sprint always required before a larger engagement?
Not always, but skipping it increases risk. If your team already has a documented architecture, clean data, and a clear use case, a pilot can start directly. If any of those three are uncertain, a discovery sprint is almost always the cheaper path — the alternative is discovering the problem during a large implementation.
What is the difference between a pilot and a production implementation?
A pilot is time-boxed, scoped to a single workflow, and designed to produce a verdict — this is feasible or it is not. A production implementation builds from that verdict into a maintained, documented system that real users interact with daily. Treating a pilot as a production build is one of the most common reasons AI projects run over budget.
Who owns the code and models at the end of the engagement?
This depends entirely on the contract. The standard expectation is that you own all work product, but fine-tuned models built on proprietary tooling may include licensing restrictions. Confirm this explicitly in the statement of work before signing.
How do weekly sprints and demos work in practice?
A sprint-based engagement delivers a working increment every one to two weeks that your team reviews against real use cases. This structure surfaces misalignments early — when they cost one sprint to fix rather than a full re-architecture. Each demo should produce a written summary of what shipped, what did not, and why.
When should you ask for fixed-price versus time-and-materials?
Fixed-price works when scope is fully defined and the data layer is clean — typically after a completed discovery phase. Time-and-materials is appropriate when scope is exploratory or data readiness is unknown. The risk in fixed-price with undefined scope is that the vendor protects their margin by narrowing deliverables; the risk in open-ended time-and-materials is that there is no forcing function on scope.
