How to prioritise an enterprise portfolio of agent opportunities

An enterprise opportunity list can grow faster than its ability to learn. Every function proposes assistants and agents. Sponsors attach optimistic savings. A small platform and governance team becomes the constraint. If every idea receives a pilot, none receives enough operational evidence.

Portfolio prioritisation should decide where the next unit of scarce attention produces the most useful business decision. That is different from predicting which idea will create the largest eventual return. Early estimates are uncertain; good portfolio governance makes that uncertainty visible and funds its reduction deliberately.

Normalise proposals into workflow opportunities

Do not compare “a sales agent” with “an HR copilot.” Require every proposal to describe the same minimum unit:

• workflow and bounded case set;
• current outcome and problem evidence;
• users and people affected;
• decisions and actions involved;
• data and systems required;
• proposed degree of autonomy;
• accountable workflow owner;
• plausible non-agent alternative;
• material consequences and constraints;
• baseline and next decision.

This turns a technology wishlist into comparable investment hypotheses. The GOV.UK discovery guidance treats understanding the problem, users, constraints, wider context and success measures as prerequisites to moving forward. A portfolio should require the same discipline before it funds experimentation.

Use six lenses, not one magic score

Assess each opportunity across six lenses. Use a simple scale with evidence notes; precision beyond the evidence is false confidence.

Decision value

What business outcome could change, for whom and how materially? Is the current problem frequent, consequential or strategically important? Is there an observable baseline? Distinguish value to the enterprise from value to users and affected people.

Workflow suitability

Does the work need interpretation or adaptive planning, or would process repair and ordinary automation do better? Are inputs, outputs, normal paths and exceptions understood? Can the workflow be bounded for learning?

Delivery feasibility

Are the necessary data, integrations, identity, evaluation and operating skills available? What critical dependency is unproven? Feasibility is local to the proposed scope, not an enterprise maturity score.

Consequence and controllability

What can go wrong, who is affected and can authority be enforced? Can the system stop, hand off, reconcile and provide redress? High consequence calls for more evidence and narrower progression, not an arbitrary deduction that hides the issue in a total score.

Economics and capacity

Estimate end-to-end cost: integration, review, evaluation, operation, model usage, incidents, support and change. Identify which scarce specialists and business operators are required. Compare with the current process and credible alternatives.

Reuse and learning

Will the work create a reusable control, dataset, integration, evaluation method or operating insight? Does it test an uncertainty relevant to several workflows? Do not inflate weak opportunities merely because they mention a common platform; require a concrete reusable asset.

Add non-compensable gates

Some conditions should not be averaged away:

• no accountable owner;
• no legitimate and usable data route;
• inability to enforce necessary authority boundaries;
• no observable outcome or state;
• no safe handling for affected people;
• no viable operating path; or
• a simpler alternative clearly dominates.

An opportunity with exciting value cannot compensate for missing authority by scoring well on five other dimensions. Mark it as blocked, redefine the scope or fund discovery specifically to resolve the gate.

NIST says risk tolerance is contextual and that risk-management resources should be prioritised according to assessed impact (Framing Risk). Portfolio governance should therefore expose consequence and uncertainty directly rather than burying both inside an overall attractiveness rank.

Allocate by stage and uncertainty

Place opportunities in a staged portfolio:

Stage	Decision being funded	Typical evidence
Intake	Is the workflow problem worth discovery?	owner, problem signal, rough boundary
Discovery	Is there a viable intervention and test?	workflow map, baseline, alternatives, constraints
Experiment	Does the proposed design work under bounded conditions?	representative cases, outcome and control results
Prepare for live	Can the service operate safely and sustainably?	architecture, ownership, support, release evidence
Live	Should scope continue, expand, change or retire?	monitoring, impact, incidents, economics

Fund the next decision, not the imagined final program. A discovery candidate may receive a small allocation to test data accessibility or user need. A live service may deserve more investment than new pilots because it produces real outcomes and evidence.

This approach also makes stopping legitimate. A candidate that resolves its key uncertainty negatively has produced value by releasing capacity.

Balance the portfolio deliberately

A healthy portfolio is not simply the top ten scores. Check concentration across:

• business outcomes and functions;
• low- and high-consequence workflows;
• assistive and action-taking modes;
• near-term operational value and longer-term learning;
• common platforms or suppliers;
• shared data and identity dependencies;
• workforce groups and affected populations; and
• delivery and assurance capacity.

Too many similar opportunities can create hidden correlated risk. Ten agents relying on the same retrieval platform or provider are not ten independent bets. Conversely, a strategically useful shared control may justify sequencing several related workflows so the first builds the foundation and the next tests reuse.

Use a portfolio evidence card

For each opportunity, keep a one-page card:

1. Decision now: advance, hold, reshape, stop or operate.
2. Evidence: observed facts and source quality.
3. Inference: what the team concludes from those facts.
4. Uncertainty: what could change the conclusion.
5. Next test: the smallest credible evidence step.
6. Owner and capacity: accountable person and required specialists.
7. Dependencies: data, platform, policy, supplier and organisational change.
8. Review trigger: date or event for reconsideration.

Separating evidence from inference prevents a sponsor’s confidence from becoming a fact. Preserve negative results. A failed data-access test may de-prioritize several opportunities and improve portfolio decisions.

Make benefits comparable without flattening them

Use outcome families—time, quality, capacity, risk, user experience, revenue support or mission effect—but preserve the original measure. “Hours saved” based on a survey estimate is not equivalent to observed reduction in end-to-end handling time. Label forecast, modelled and observed values.

The GOV.UK benefits guidance recommends comparing observed evidence with forecasts and using it to decide whether work should continue or change. Apply that at portfolio level: update value estimates as pilots and live services produce data, and reallocate accordingly.

Do not let avoided-risk claims dominate without a plausible scenario and control link. Do not count model-generated drafts as completed value if people discard or substantially rework them.

Run a decision meeting, not a pitch contest

Send evidence cards in advance. Use a multidisciplinary group that can challenge domain value, architecture, data, workforce impact, security, privacy, evaluation and economics. The accountable portfolio authority should decide and state why.

Ask the same questions of every candidate:

• What changed since the last decision?
• Which claim is observed and which remains assumed?
• What does the non-agent alternative achieve?
• What is the most consequential unresolved condition?
• Which scarce capacity will this consume?
• What will the next allocation let us decide?
• What will cause us to stop?

GAO’s framework connects governance, data, performance and monitoring in AI accountability (GAO). A portfolio review should connect them too, rather than approving value narratives separately from operating evidence.

Re-rank as evidence changes

Priority is not a permanent label. Reassess when a dependency matures, a regulation or policy changes, a supplier updates capability, a live incident reveals correlated risk, or outcome evidence challenges the business case.

NIST’s AI RMF Core treats mapping, measurement and management as continuous across the lifecycle. Portfolio allocation should be equally dynamic. Protect live-service improvement and retirement work from being displaced by a constant supply of new ideas.

The best portfolio is not the one with the most agents. It is the one that can explain why each active investment is the best next use of scarce evidence-building and operating capacity.