Draft the AI Solution
Key Takeaways
- The Phase 1 solution draft is not a detailed design: it is enough architecture to size, budget, and de-risk the work, and it feeds the feasibility assessment and the business case.
- Every AI solution contains the same building blocks — data sources, ingestion, a feature or data store, a training environment, model serving, integration points, and monitoring with feedback capture — and the draft must answer one question for each.
- Choose batch scoring or real-time inference from when the decision is made, not from ambition: a weekly purchasing decision does not need a millisecond endpoint, and a card authorization cannot wait for tonight's batch.
- Name the AI pattern and a candidate model class with the evidence that would settle it, and leave the algorithm to CPMAI Phase 4 (Model Development).
- For third-party APIs and pre-trained models, data residency and vendor data-use terms are a gate, not a preference, and exit cost belongs in the build-versus-buy-versus-fine-tune comparison.
What the Solution Draft Is — and Is Not
At the end of CPMAI Phase 1 (Business Understanding) you produce a high-level solution draft: enough architecture to size the work, budget it, and expose the risks, and no more. It is not a detailed design, not a data model, and not a vendor selection. Its purpose is to be an input to the feasibility assessment and the business case — and to be revised. Phase 2 (Data Understanding) will find that data you assumed exists does not; Phase 3 (Data Preparation) will change the pipeline; Phase 4 (Model Development) will pick a different algorithm. A draft that cannot survive being changed was over-specified. On the exam, committing in Phase 1 to a named neural architecture, a graphics processing unit (GPU) cluster, and a vendor before anyone has looked at the data is a trap.
Walk the Data Flow End to End
Take a claims-triage solution and trace one case. A claim is created in the policy administration system; an ingestion job picks it up on submission, with a nightly batch for corrections; the claim's attributes are joined with policy history and prior-claim aggregates held in the feature store; the scoring service returns a fraud-risk score and the top three contributing factors; the score is written back to the claim record and shown as a priority band in the adjuster's existing queue; the adjuster's decision and the eventual outcome — paid, denied, confirmed fraud — are logged against the claim ID; monitoring compares live score distributions and outcomes to the training period and raises drift alerts. One paragraph like that, written for your project, exposes more risk than a page of boxes and arrows, because every hand-off in it hides a system owner, an interface contract, or a missing outcome label.
The Components Every AI Solution Has
| Architecture component | Question the draft must answer | Consequence of leaving it blank |
|---|---|---|
| Data sources | Which systems hold the raw records and the outcome labels, who owns them, and how much history exists? | Phase 2 discovers the label you need was never recorded, and the project restarts or dies |
| Ingestion and data flow | How does data move, how often, and how are late corrections handled? | Freshness assumptions collide with reality and the model scores stale inputs |
| Feature or data store | Where do prepared inputs live, and will the same features be available at scoring time as at training time? | Train/serve skew: the model performs in the lab and fails in production |
| Training environment | Where does training run, on what compute, with what access to sensitive data? | Compute cost and privacy approvals surface after the budget is fixed |
| Model serving | Batch or real-time, at what volume, within what latency budget and availability target? | The workflow cannot consume the prediction in the time it actually has |
| Integration points | Which host systems, which interface contract, whose team, which release calendar? | Integration effort is discovered late and then dominates the schedule |
| Monitoring and feedback capture | What is logged, what triggers an alert, and how are outcomes captured for retraining? | No evidence of business impact and no path to retraining in Phase 6 |
| Operations and cost | Who runs it, what does inference cost per thousand predictions, and how often is it retrained? | Total cost of ownership is understated and the ROI case collapses in year two |
Batch Scoring or Real-Time Inference
Choose by when the decision is made. Batch scoring runs on a schedule against a set of records and writes results to a table or the host system. It tolerates minutes or hours of latency, is cheap and simple to operate, and can be rerun after a bad run — but it needs a stated freshness rule and a run window that fits the business calendar. Real-time inference computes the score inside the user's or system's transaction. It demands a latency budget in milliseconds, an availability target, features that can be looked up or computed at request time, and a defined fallback when the service is down. Weekly purchase orders do not need a 200-millisecond endpoint; a card authorization cannot wait for tonight's batch. Choosing real-time by default imports high-availability engineering, on-call rotas, and cost that the decision cycle never required.
Specify the Approach Without Over-Committing
State the AI pattern — Pattern and Anomaly Detection for fraud, Predictive Analytics and Decision Support for maintenance, Conversational and Human Interaction for a service assistant — and a candidate class of approach: supervised binary classification, time-series forecasting, unsupervised anomaly detection, ranking, or a language model with retrieval over your own documents. That is enough for sizing, skills planning, and a first cost estimate. Write it as a candidate with the deciding evidence named: "supervised binary classification if Phase 2 confirms at least eighteen months of reliable fraud outcome labels; unsupervised anomaly detection if the labels prove unusable." The algorithm decision belongs to Phase 4.
Build, Buy, or Fine-Tune
This is where the guide settles the sourcing decision: whether the capability is bought, adapted, or built. Three routes exist, and a foundational-level answer justifies the choice on time-to-value, data sensitivity, unit economics, and exit cost — never on architecture.
| Sourcing route | What it is | Strongest when | The caveats that decide it |
|---|---|---|---|
| Buy — call a third-party API | Pay per call for a vendor's hosted model | You need a working capability fast, the task is generic, and you hold no training data | Per-call charges for the product's life; vendor model updates can change behavior without notice; limited explainability; data residency and vendor data-use terms are a gate, not a preference |
| Fine-tune a pre-trained model | Adapt an existing model to your domain inside your own environment | The work is recognition or language work and you hold a modest labeled set | Base-model license must permit commercial use; adds MLOps capability you may not have; settle who owns the tuned weights |
| Build custom | Train your own model on your own data | The capability is genuinely differentiating, or the data cannot leave your control | Longest timeline and largest data, skills, and compute requirement; hard to justify when a comparable model already exists |
For regulated or personal data, establish the processing location, the retention terms, and whether the vendor trains on your inputs against the privacy and security plan from Domain I before the option is costed — an option the organization cannot lawfully run is not a cheap option. Then compare the three on differentiation, data sensitivity, unit economics at expected volume, and exit cost: how hard would it be to move off this vendor in two years, what would have to be rebuilt, and who owns the fine-tuned weights and derived artifacts? A team proposing to build from scratch owes you that comparison written out in money and weeks.
Document Integration Points by Name
"We will integrate with the CRM" is not a documented integration point. "Read customer contact history from the CRM through the existing REST service owned by the CRM platform team; write the propensity score back to the account record nightly; requires change-advisory-board approval and avoids the quarter-end freeze" is. Record system, direction, interface contract, owning team, and required approval. This list is where the real schedule lives and where cross-team dependencies become visible to the sponsor.
Outline Deployment and Operational Considerations Early
Name, at draft level, who operates and monitors the solution, the expected retraining cadence, the human review path for low-confidence outputs, the logging retained for the audit trail, and the run cost per thousand predictions. This work belongs to Domain V (Operationalize AI Solution), but if it is missing from the Phase 1 draft the business case counts build cost only — and understated total cost of ownership is one of the most common ways an AI project gets approved and then canceled.
Keep It Deliberately Revisable
Version and date the draft, and record the assumption each component rests on. When Phase 2 shows the outcome label is only 60% populated, you change the draft, re-cost it, and take the change back through feasibility — which is CPMAI iteration working exactly as intended, not a failure of planning.
A retailer wants a demand forecast that category buyers will use when they place purchase orders every Monday morning. The solution architect's draft proposes a real-time inference service with a 200-millisecond latency budget and a 99.9% availability target. What should the project manager challenge first?
A European insurer's draft AI solution proposes sending free-text claim narratives that contain policyholder names and medical details to a third-party large language model API hosted outside the EU. What should the project manager add to the draft before the option is costed?
During Phase 1 the executive sponsor asks the project manager to commit in the business case to a specific deep-learning architecture and a named vendor's GPU platform. Data Understanding has not begun. What should the project manager do?