This guide provides a comparison framework and five illustrative projects. It addresses investment choices; the AI roadmap then sets out the programme and timing.
Key points
A blocking prerequisite must remain visible even when expected value is high.
An unsupported estimate remains an assumption; assigning a score does not strengthen it.
Dependencies, budget and team capacity determine launch order.
What information do you need?
Prepare a short brief for each project: the specific problem, users, activity volume, result to verify, business owner, alternatives and initial resource estimate. Include the data needed and possible consequences of errors.
Compare AI with clearer rules, existing features, conventional automation and leaving things as they are. An incomplete brief calls for preparation. A small exploration budget may be justified with a defined result; it need not authorise full deployment.
Identify blocking prerequisites first
Examples include no available business owner, inaccessible essential data or insufficient control over actions that commit the company. Distinguish what prevents deployment from what can be explored in a limited test.
Missing historical data may prevent reliable forecasting at scale without preventing preparation. A small pilot, however, does not justify unauthorised exposure of data.
The NIST AI Risk Management Framework, Map function supports examining context and whether AI is appropriate before proceeding. It does not supply the commercial scoring scale used here. An AI assessment can establish prerequisites currently spread across business teams, IT and management.
Compare seven criteria
The following framework is a Spentia methodological proposal, not an externally validated standard.
Criterion | Question | Evidence needed |
|---|---|---|
Expected value | What useful result, at what full cost and by when? | Baseline, volume and benefit assumptions |
Technical feasibility | Can the solution be built and operated in this context? | Technical trial, integrations and skills |
Data | Does the required information exist and can it be used? | Access, quality, freshness and permitted use |
Risk controls | Which errors or consequences would be unacceptable? | Failure scenarios, checks and fallback |
Adoption | Can people integrate it into daily work? | User testing and support workload |
Dependencies | Which decisions or deliverables must come first? | Prerequisites, owners and realistic dates |
Strategic priority | Which agreed management objective does it serve? | Explicit objective and accountable owner |
For each criterion, distinguish established facts, estimates and unknowns. Do not automatically turn missing information into a poor score: a focused test may be needed.
Should you use a weighted score?
A score can make decision-makers’ preferences explicit if the scale and weights are defined before ranking. On a 1–3 scale, a higher score must always mean a more favourable situation. For risk, score the strength of controls rather than severity.
Define levels using company-specific reference points: expected value or capacity, effort and integrations. A score of 2 describes an intermediate situation; it must not hide missing information.
With weights totalling 100%, the weighted result remains between 1 and 3. This ordinal ranking supports discussion; it is not ROI. Keep blocking conditions separate and test whether reasonable weight changes reverse the priorities. An unstable ranking calls for more information.
You can also compare documented criteria without producing a total. The example below takes this approach.
Organisation and assessments
Make organisational decisions with a clear assessment
Our assessments identify practical changes to help your teams and processes work together.
Five illustrative projects
This is a fictional teaching exercise. The situations illustrate possible decisions; they are not a client portfolio or observed gains.
Candidate project | Assumed situation | Suggested decision |
|---|---|---|
Identify unusual clauses in customer requests | Accessible documents, limited scope and business review available | Test representative cases before production use |
Assist incoming lead qualification | Potentially valuable, but qualification criteria and CRM data incomplete | Prepare these prerequisites before a pilot |
Answer routine HR questions | Low request volume; a document library may suffice | Compare the simpler option and defer AI investment |
Forecast sales demand | Significant capacity issue; history available but predictive performance unknown | Fund a test against a simple baseline method |
Negotiate commercial terms autonomously | Concessions cannot be controlled or actions stopped in time | Rule out this autonomous scope; consider preparation support |
The first project can be tested because controls are available, without assuming effectiveness. The second needs a preparation budget. The third calls for a simpler alternative. The fourth needs performance evidence. The fifth requires a different scope before it can be considered in this context.
These decisions reflect needs, evidence, alternatives and delivery conditions, rather than a single average. AI agents require particularly careful definition of permitted actions and controls.
Turn priorities into a launch sequence
List dependencies and team availability. Projects may compete for the same experts, data or training period. Data preparation may need to precede several use cases despite limited direct benefit.
Do not add benefits from projects that handle the same cases or release the same hours. Account for overlap in the portfolio value.
For each selected project, state the next decision, budget authorised for this stage, owner and review date. Permission to run a trial is not funding for every later stage. After the test, extend, adjust or stop.
Frequently asked questions
How many projects should run at once?
How many projects should run at once?
As many as the teams can genuinely prepare, control and support. Estimate user, expert and support time. The number of available ideas does not measure delivery capacity.
Should quick wins take priority?
Should quick wins take priority?
They can provide evidence and sometimes reduce spending quickly. They do not automatically fund the next stage: released time is not necessarily cash. A prerequisite may take priority if several significant benefits depend on it.
What about projects without reliable data?
What about projects without reliable data?
Identify the information actually needed and the work required to obtain or improve it. Decide whether limited exploration is possible or the project should wait. Avoid a broad data programme without a defined use case.
Who decides the priorities?
Who decides the priorities?
Management decides with business owners and the relevant functions. The framework makes the evidence explicit. Identify who is accountable for each result before launch.
When should priorities be reviewed?
When should priorities be reviewed?
At the end of a test, when a prerequisite changes or at an agreed review. Update changed assumptions, available resources and observed results. Funding decisions should be open to reassessment.
Sources
NIST, AI Risk Management Framework 1.0 — Core, 2023, Map and Manage functions.
The comparison framework, scoring scale and five examples are Spentia methodological tools. No success statistics or client portfolio are claimed.
