Identify Project Resources: People, Hardware, and Contractors
Key Takeaways
- Data engineering and annotation are usually the largest and most under-planned effort on an AI project — not model building.
- The domain SME is not optional: the SME defines label correctness, adjudicates edge cases, and prices what each error type costs the business.
- Procurement of AI tooling is a dated schedule dependency, not a background errand: security review, legal review and vendor onboarding routinely take longer than configuring the tool itself.
- Contractor agreements need an explicit knowledge-transfer clause (documented pipelines, feature and label definitions, retraining procedure, model card, named internal staff shadowing) or the organization cannot operate the model afterwards.
- Resource demand is phase-shaped: Phases 2-3 load data engineers, annotators and the data steward, while Phase 6 loads MLOps, support and change management. Flat staffing plans are wrong.
Resourcing an AI Project Is Not Resourcing an IT Project
Task 10 closes Domain II by turning an approved solution concept into a staffing, infrastructure, and procurement plan. Two things make artificial intelligence (AI) work different from conventional IT resourcing: the largest effort is usually data work, not model building, and the team profile changes shape at every CPMAI phase. Plans that ignore either fail predictably — a highly paid data scientist idle for eight weeks waiting on a pipeline, or a working model with nobody to operate it after go-live.
Assessing Skill Requirements
| Role | What they do | Where they load | Consequence of omitting |
|---|---|---|---|
| Data engineer | Build ingestion, joins, pipelines and feature stores; move data from source systems into a usable workspace | Phases 2-3 heaviest, continuing through 6 | The most common under-plan; data scientists hand-wrangle files and the schedule slips silently |
| Data scientist / ML engineer | Select techniques, engineer features, train and tune models, run evaluation | Phases 4-5 | Rarely omitted, frequently over-weighted relative to the data roles that feed it |
| MLOps engineer | Deployment pipelines, model registry, monitoring, drift detection, retraining automation | Phase 6, designed from Phase 4 | The model works in a notebook and never reaches production, or ships with no monitoring |
| Domain subject-matter expert (SME) | Define what a correct answer is, adjudicate edge cases, price each error type, sanity-check outputs | All phases; heaviest in 1-3 and 5 | Labels encode the wrong thing and the model optimizes a target the business does not value |
| Data steward / data owner | Grant access, confirm lineage and permitted use, approve data movement | Phases 2-3 | Access requests stall for weeks; unapproved data use surfaces at the compliance review |
| Legal / privacy | Consent basis, retention, cross-border transfer, contract terms, GDPR and CCPA obligations | Phases 1-3, again before deployment | Rework or a blocked launch late, when it costs the most |
| UX / service designer | Design how a prediction reaches the human and what they can do with it | Phases 1 and 6 | An accurate model whose output is ignored — the classic adoption failure |
| Product / business owner | Own priorities, accept or reject the solution, own the benefit after go-live | All phases | No decision authority; scope drifts and benefits go unclaimed |
| Annotators / labelers | Produce and quality-check labeled training data against a labeling guide | Phase 3, plus refresh cycles | Supervised learning has no ground truth; "we will label it as we go" plans collapse |
Two judgments earn marks. First, data engineering and annotation are usually the largest and most under-planned effort on an AI project — when a resource plan looks thin, that is where to look. Second, the SME is not optional and not a part-time courtesy reviewer: the SME defines label correctness and the cost of each error type, and no amount of modeling talent substitutes for that. Note also that roles are not headcount — one person may cover two on a small project — and that shared internal people have day jobs, so their time is negotiated with their line manager, never assumed.
Evaluating Hardware and Infrastructure
The resource plan has to name and fund two distinct infrastructure lines: a training environment for the data science team, and the production inference capacity the solution will consume for as long as it runs. Each needs a budget owner, a procurement route, and a date by which it must exist. The technical sizing of both — what kind of compute, how much of it, burst or reserved — is settled when the AI workspace is provisioned in CPMAI Phase 2 (Data Understanding); what belongs in this plan is the money, the lead time, and the named owner.
Cloud versus on-premises turns on four drivers, not preference: data residency and regulatory limits on where data may physically sit; cost predictability (variable consumption versus a depreciating owned asset); existing commitments such as a committed cloud spend agreement or recently purchased hardware; and elasticity for bursty training. Insist on environment separation — development, test and production with controlled promotion, and production data used in lower environments only under masking or an approved exception. Budget storage for datasets, versioned features and model artifacts; teams routinely under-size it.
Contractors Versus Internal Capability
Bring in external help when a skill is genuinely scarce, needed briefly, or needed now: a first model where you have no in-house machine learning capability, a narrow specialty, or a short spike to prove feasibility. Build internally when the capability will be needed repeatedly, or when it encodes domain knowledge you cannot outsource. The deciding question is whether the model becomes a permanent operational asset — if it does, you will be running, monitoring and retraining it long after the consultancy leaves.
That is why the knowledge-transfer clause is the contract provision that matters most. Specify documented code and pipelines, feature and label definitions, the training and retraining procedure, a model card and evaluation record, environment configuration, and named internal staff shadowing the build, with a formal handover and written acceptance criteria. Settle ownership of data, trained model weights and derived artifacts at the same time, plus whether your data may be used to improve the vendor's other models. A model you own but cannot retrain leaves you dependent anyway.
Coordinating with Procurement
Specialized AI tooling — labeling platforms, feature stores, ML platforms, model-hosting or foundation-model application programming interface (API) access — moves slowly through procurement. Start early and plan for:
- Lead times. Security review, legal review and vendor onboarding routinely take longer than configuring the tool. Put procurement on the schedule as a dated dependency, not a background errand.
- Licensing models. Per-seat, per-node and consumption pricing (per token, per prediction, per compute-hour) behave very differently as volume grows; consumption pricing that looks trivial in a pilot can become the largest run-cost line in production.
- Security review. Where data travels, who can see it, certification evidence, breach notification obligations.
- Data-processing terms. Processing location, sub-processors, retention, deletion on exit, and an explicit position on vendor training against your data.
Phase-Shaped Resource Loading
A staffing plan showing the same team at the same allocation from Phase 1 through Phase 6 is wrong on its face. Phases 2 and 3 are dominated by data engineers, annotators, the data steward and SME time. Phases 4 and 5 shift weight to data scientists, with MLOps designing deployment alongside them. Phase 6 loads MLOps, support, security and change-management effort while modeling demand falls away — and that demand continues past go-live for monitoring and retraining. Build the plan phase by phase, hold the ramp-up and ramp-down explicitly in the schedule, name where shared people are committed so their managers can plan, and present the resource plan alongside the phase-gated ask from the business case so the organization commits people only to the phase it has actually funded.
You are staffing a clinical-coding AI project with a data engineer, two data scientists, an MLOps engineer, a coding subject-matter expert and a product owner. The sponsor asks you to cut one role to fit the budget and suggests dropping the coding subject-matter expert, arguing that "the coders' output is already in the training data." What is the strongest response?
Your resource plan for a demand-forecasting project shows the same seven-person team allocated at 100% from Phase 1 Business Understanding through Phase 6 Model Operationalization. What should you change first?
A consultancy will build the first model for your demand-forecasting project because your organization has no in-house machine learning capability yet, and the model will run permanently in operations afterwards. Before signing, which contract provision matters most to the organization's ability to run the solution once the consultancy leaves?