The pitch you hear most often about AI in portfolio planning is that it makes the plan for you. You feed it your strategy, your projects, your budget, your constraints, and it returns the optimal portfolio. The work that used to take three months of arguments in spreadsheets is reduced to a single conversation.
It is a good pitch. It is also not what serious organisations actually want.
What they want is something more boring. They want the spreadsheet arguments to be fewer and sharper. They want the parts of the work that are repetitive to stop consuming senior people. They want the parts that need judgement to stay with the people accountable for the outcome. And they want to be able to defend the plan to a board, line by line, when it lands.
That is the boring part of AI in portfolio planning. It is also the useful part.
Start with the bit that should not be AI at all
The DemandFlow Medium Term Plan is scored deterministically. The scoring engine is dull on purpose, because the rules need to be defensible in a CFO conversation.
Each project is scored across five dimensions. Strategic initiative alignment, network platform urgency, benefit, success probability and customer impact. The composite is a weighted sum. Weights live on the plan, sum to one hundred, and can be edited with the change recorded.
The initiative alignment score uses a max plus bonus formula. The highest contribution counts in full, and the others nudge it up. That removes the most common distortion in portfolio scoring, where a project that touches several strategic themes ends up looking more important than a project that genuinely delivers one.
Benefits are de-duplicated across projects and initiatives. If two projects claim the same revenue uplift, the scoring engine recognises it and refuses to count it twice. Most organisations, if you summed every claimed benefit in their plan, would be delivering several times the actual achievable improvement. Removing that arithmetic on its own raises the quality of the conversation.
Weight defaults are hybrid. They inherit from the tenant settings, but the plan owner has to confirm or edit them at the start of each cycle. No silent reuse of last year’s weights.
Deferred projects auto-roll into the next cycle as proposed, carrying their prior composite score as a starting point. Nothing quietly falls through the cracks. Nothing quietly moves either. The roll is proposed, not committed.
None of that is AI. It is structure. The point of the structure is that the AI does not need to do the work the structure already does.
Where the AI actually earns its keep
There are two places.
The first is sizing project proposals from network and software lifecycle data. The auto-planning engine sweeps the platforms, instances and software versions in the estate, identifies upgrade needs by end-of-support dates, folds in in-flight and deferred work, and gathers cost references from closed projects. That context goes to a large model which proposes sized projects with rationale attached. Those proposals land in the plan as draft entries that a human planner accepts, edits or rejects.
The model is not setting strategy. It is doing the kind of work that a senior planner can do, given two weeks and a pile of spreadsheets, in a few minutes instead. The output is auditable and labelled as AI-sourced, so it never gets confused with a human commitment.
The second is watching projects after they have been funded. Once a project is on the roadmap and into delivery, a daily inference layer reviews its current state. Schedule, budget, risk register, blocked work, gate status. It produces a confidence-rated health summary, a narrative explanation and a small set of recommendations. The portfolio view shows the heat map. The exec sees the same data the project manager sees, at the same time, instead of two days later in a slide pack.
It is advisory output. It carries a date, a confidence note and a refresh option. It is built to be useful between steering meetings, not to replace them.
The handoff that ties it all together
When the plan is approved, the selected projects flow automatically into the financial roadmap as line items, year by year. The plan, the project and the roadmap share the same data. The approval is the handoff, and the handoff is the same record.
That matters because the most common failure mode in portfolio planning is the gap between the approved plan and the funded plan. A plan that produces roadmap entries on approval cannot be silently overwritten by a later spreadsheet.
What we deliberately did not build
It is worth being clear what is not in scope. The platform does not auto-approve projects. It does not generate strategic themes. It does not decide which initiatives matter to the organisation. It does not predict benefit realisation that the project sponsor has not committed to.
Those are decisions, and decisions belong to people. The platform’s job is to make the decisions easier, not to make them.
The bigger point
A plan is a thinking tool. The interesting question is not how clever the AI in the plan can be. It is whether the plan still tells the truth six months after it was approved.
A deterministic scoring engine, AI applied narrowly to the work that genuinely benefits from it, and a direct line from approval to the financial roadmap, gets you closer to that than a more impressive demo would.
Boring, in this case, is the feature.
