Enterprise Compute Initiative MVP: set a hypothesis and exit test

An Enterprise Compute Initiative MVP should test one valuable AI use case under real operating conditions. Before the project starts, write the hypothesis, baseline, users, data boundary and exit test. This keeps the work focused on a business decision rather than a polished demonstration.

The result can be proceed, revise or stop. Stopping after a well-run test is useful when the evidence shows that the workflow, data or economics do not justify scale.

Understand the current programme structure

The official Enterprise Compute Initiative page says ECI is a short-term programme for Singapore-based companies. Participating companies work with cloud service providers for tools, cloud credits and training, and with selected partners for AI strategy, technical implementation, MVP development and change management.

The page currently states that supportable consulting cost is capped at S$150,000, with DISG support of 70% of actual cost capped at S$105,000. Companies pay the remaining 30% and are expected to commit manpower and technical resources for a programme of up to six months.

These are programme ceilings and expectations, not a promise of selection or a substitute for the live cohort terms. The official application page asks companies to provide their AI ambition and use case and allows them to indicate a preferred cloud provider.

Write a falsifiable hypothesis

Use one sentence with a workflow, user, measurable improvement, quality boundary and time period. For example, the company might test whether an assisted review reduces first-pass handling time while keeping verified accuracy above an approved threshold.

Avoid a hypothesis such as AI will improve productivity. It has no baseline, task or failure condition. Choose a measure the company can collect without changing the definition after results appear.

State what would disprove the case. A high error rate, unavailable data, unacceptable review effort or lifetime cost above the business value can all be valid stop conditions.

Select representative work

Build a sample that includes normal, difficult and adverse cases. Exclude personal, client or regulated data until the security and legal basis are clear. If synthetic or redacted data is used, note how that limits the result.

Record current task volume, time, error, rework, delay and escalation. The baseline should be measured with the same definitions used for the MVP.

The 60-day AI trial scorecard provides a smaller trial method. ECI can support a broader MVP and change process, but the decision still needs an honest comparison.

Define the technical and data boundary

List source systems, data fields, integrations, model or service, user roles, hosting, retention and output destinations. Identify what remains outside the MVP.

Decide how results are evaluated and logged. Separate model output, business rule, human correction and final decision. This allows the team to see whether improvement comes from AI, process redesign or added manual effort.

Review provider terms, access, subprocessors, deletion, export and incident response. Keep credentials and production secrets in approved secure systems, never in the project evidence pack.

Commit the right internal team

Name a business owner, process expert, technical lead, data owner, security reviewer and affected user group. Reserve their time in the project plan.

The official programme expects participating companies to commit manpower and technical resources. A company that delegates every decision to the partner may receive an MVP it cannot operate or challenge.

Include change-management work early. Users should understand what task changes, what judgment remains human and how exceptions are handled.

Build the evidence plan

Use a table with hypothesis measure, baseline, target, sample, data source, reviewer and decision rule. Add cost, adoption, control exceptions and failure cases.

For an output-quality test, use independent review rather than asking the builder to grade its own examples. Preserve unsuccessful cases and corrections. A curated demo set can hide the risks management most needs to see.

Track cloud and consulting usage against budget. Credits can reduce project cost, but the scale decision should use expected post-programme cost and internal support effort.

Hold two formal decision reviews

At the midpoint, check whether the data, workflow and evaluation remain valid. Stop or narrow the work if a core assumption has failed instead of spending the remaining budget to complete the original slide deck.

At the end, compare evidence with every exit condition. State limitations, unresolved risks, lifetime cost and workforce impact. Approve only the next proportionate stage.

If the decision is stop, export permitted records, remove integrations, revoke access and complete contractual deletion steps. A revise decision should define one unanswered question and a new limit. Proceeding requires approved production ownership, monitoring and incident response.

Plan ownership after the programme

Identify who maintains prompts, models, data pipelines, evaluations and user training after partner support ends. Record vendor dependencies and exit options.

The enterprise AI readiness guide covers organisation-wide preparation. The business support decision guide remains the shared cornerstone.

An Enterprise Compute Initiative MVP creates value when it answers a defined operating question with evidence. The cloud platform, partner and funding can enable the test, but management still needs a clear hypothesis and the discipline to stop when the exit conditions are met.

## Use an evidence table, not a demonstration narrative

For every hypothesis, record the baseline, test population, model or system version, result, limitation and decision. Keep failed tests. They often provide the clearest evidence that the team changed direction for a reason rather than merely polishing a preferred demonstration.

Define who may approve continuation when a threshold is missed. The approver should see the cost already incurred, the remaining cloud or consultancy budget, the next learning objective and the point at which another experiment stops being proportionate.

For AI use cases, add data ownership, personal-data handling, security, human oversight and output-error controls. A technically successful prototype may still fail the business or governance test. The exit decision should consider all three dimensions: measurable benefit, deployability and responsible operation.