AI projects look like software projects, and much of standard project management still applies: scope, schedule, budget, RAID logs, status reports. But AI adds new questions. Is the data good enough? How accurate is accurate enough? Who checks the answers after go-live? What happens when the model is wrong?
This article sets out the PMO toolkit Egisedge uses on AI and AI-on-Oracle projects. Each template is in our AI PMO library and can be adapted to your own methodology.
The AI project lifecycle at a glance
| Phase | Key question | Templates used |
|---|---|---|
| Idea | Is this worth doing? | Use-case scorecard |
| Discovery | Can we do it with our data? | Charter, data readiness checklist, RACI, risk register |
| Pilot / PoC | Does it work on real data? | PoC plan, evaluation plan |
| Build and test | Does it meet the thresholds? | Evaluation plan, UAT plan, responsible AI checklist |
| Go-live | Are we ready to switch on? | Go-live readiness checklist |
| Hypercare | Is it still working? | Hypercare and monitoring plan, status report |
1. Use-case intake and prioritisation
Before a charter, score each candidate on value, feasibility, data readiness, integration effort, risk and time to value. Our use-case scorecard uses a simple formula: priority = (value × 2 + feasibility + data readiness) − risk. The point is not the formula. It is making the trade-offs explicit, so the sponsor chooses between options with the same information.
2. The AI project charter
The AI project charter fits on one page. It records:
- Business problem – what is slow, costly or error-prone today, and who feels it.
- Objective and value – hours saved, error rate, cycle time or cost.
- Scope – processes, entities and systems in and out of scope.
- AI approach – agent, classification, extraction, generation or forecasting; the platform; and where people stay in the loop.
- Data – sources, owners, volumes, sensitivity and residency.
- Success metrics – accuracy targets and the acceptance threshold for go-live.
- Milestones – discovery through hypercare.
The success threshold is the most important line. Agree it before build starts, not after the results are in.
3. RACI
AI projects bring in roles that standard IT projects may not have: data owner, AI engineer, InfoSec for model and data review. Our RACI template assigns each activity, from data access approval to responsible AI review and the go-live decision. Make sure the sponsor, not the delivery team, is accountable for the responsible AI review and go-live.
4. Data readiness
Use the data readiness checklist at the end of discovery:
- Data sources listed with owners
- Access approved in writing by the data owner
- Sensitive data classified and masking agreed
- Data residency confirmed
- Enough history available for the use case
- Quality profiled: completeness, duplicates, invalid codes
- Labelled examples or ground truth available for evaluation
- Refresh pipeline defined
- Retention and deletion rules agreed
- Test data separated from build data
If ground truth is missing, you cannot measure accuracy. Fix that first.
5. The AI risk register
The AI risk register includes AI-specific risks alongside delivery risks. Standing entries include:
| Risk | Typical mitigation |
|---|---|
| Wrong or fabricated answers | Ground in approved sources; cite sources; human review for financial postings |
| Sensitive data exposed to an external model | In-tenancy models; mask personal data; data processing agreement |
| Poor data quality | Profile and cleanse early; accuracy gates |
| Low user trust and adoption | Co-design, training, visible sources, phased rollout |
| Oracle quarterly update breaks an integration | Regression suite before each update |
| Model drift after go-live | Monitoring with defined triggers |
Score each by likelihood and impact, name an owner, and review it weekly.
6. Stage gates
The stage-gate plan has five gates:
- Gate 0 – Idea approved: scorecard above threshold; sponsor named.
- Gate 1 – Discovery complete: charter signed; data readiness passed; risks logged.
- Gate 2 – Pilot ready: design approved; evaluation plan and thresholds agreed; security review passed.
- Gate 3 – Go-live ready: evaluation results meet thresholds; UAT signed off; responsible AI checklist complete; cutover plan approved.
- Gate 4 – Hypercare exit: 30 days stable; monitoring live; handover done; benefits tracking started.
Gates are decisions, not ceremonies. Each one should end with "proceed", "rework" or "stop".
7. Evaluation plan
The evaluation plan is where AI projects differ most from standard testing:
- Golden set: real, labelled examples kept separate from build data, including edge cases and bad inputs.
- Metrics: accuracy, precision and recall per task; groundedness; refusal correctness ("I don't know" when data is missing); latency and cost per transaction.
- Test types: functional, regression after any prompt or model change, adversarial (prompt injection), bias where people are affected, security, load.
- Human review: sample size and named reviewer for each release.
Re-run the evaluation every time the prompt, model, retrieval logic or source data structure changes.
8. Responsible AI, UAT and go-live
Before Gate 3, complete the responsible AI checklist, the UAT plan and sign-off and the go-live readiness checklist. Record the model and prompt version that goes live, so you can compare later.
9. Hypercare and monitoring
The hypercare plan sets daily, weekly and monthly checks: error logs and failed integrations, sampled human review of AI outputs, drift checks, cost, adoption and benefits. Define triggers in advance, for example a fall in accuracy beyond an agreed number of points or rising user overrides.
10. Reporting to sponsors
A one-page weekly status report with RAG status adds AI metrics to the usual scope, schedule and budget lines: evaluation accuracy, test cases passed and open defects. Sponsors should see model quality as often as they see budget.
How Egisedge helps
Egisedge Technologies is an Oracle Partner delivering AI agents, AI/ML and data science, Oracle APEX + AI applications and Oracle Fusion Cloud implementations. Our AI products include the patented AI Financial Compliance Reporting platform, with AI Agents for Finance, AI Agents for HCM and Consol360 in development. Every template in this article is available in the AI PMO library in our knowledge repository.
Starting an AI programme and want a governance structure that works from day one? Email info@egisedge.com or book a demo with our team.
Frequently asked questions
How is PMO for an AI project different from a normal IT project?
AI projects need extra controls for data readiness, model accuracy, hallucination, bias, drift and data privacy. The PMO adds an evaluation plan with thresholds, a responsible AI review and ongoing model monitoring to the usual scope, schedule and budget controls.
What should an AI project charter include?
The business problem, objective and measurable value, scope, AI approach with human-in-the-loop points, data sources and sensitivity, success metrics with go-live thresholds, milestones and key risks.
What are stage gates in an AI project?
Stage gates are decision points with entry and exit criteria, for example idea approved, discovery complete, pilot ready, go-live ready and hypercare exit. A project only moves forward when the gate criteria are met.
How do you measure whether an AI model is ready for go-live?
Test it against a golden set of real, labelled examples kept separate from build data, and compare results with agreed thresholds for accuracy, groundedness, refusal correctness, latency and cost.
Talk to Egisedge about AI PMO
Tell us where you are and we'll reply within one working day with next steps.