Crowdsource AI ideas, but build the management process too

Why crowdsource at all

The people closest to the work spot the highest-value AI opportunities first. Contact centre agents know which questions repeat fifty times a day. Finance analysts know which reconciliations eat their week. Field engineers know which forms exist only to satisfy a system nobody trusts any more.

If your AI programme only surfaces ideas from the executive team or a central data function, you miss most of the good ones. Worse, you send a signal that AI is something done to the organisation, not with it. People who could be your best source of ideas learn to keep quiet.

Why crowdsourcing alone fails

Open the submission box without a management process behind it and three problems show up fast, usually within a month.

1. Volume overwhelms triage. Every idea feels urgent to the person who submitted it, and none get a timely answer. A queue of forty ideas with no owner is worse than no queue at all.
2. Duplicate and near-duplicate ideas pile up. Three teams independently submit "automate the weekly status report," phrased three different ways, and nobody notices until someone builds it twice.
3. Contributors lose trust. Submissions vanish into what looks like a black box. People stop contributing within a couple of cycles, and the ones who stop first are usually your sharpest submitters, because their time is worth the most.

The fix is not to close the funnel. It is to build a lightweight process behind it, before you open it, not after the first backlog forms.

A worked example

Say a claims handler submits: "We spend hours chasing missing documents from customers before we can process a claim." No cost estimate, no technical detail, just the problem as they experience it.

A good process does the following in order:

1. Capture it against a fixed template: problem, who's affected, rough value, rough effort. The handler fills in the first two fields and leaves the rest blank, which is fine.
2. Enrich it before anyone scores it. AI drafts a plausible value estimate from claims volume data already in the system, tags the likely affected teams, and flags similar past submissions. A human checks this, they don't just accept it.
3. Triage it on the next fortnightly review, alongside everything else submitted that cycle, not in an ad hoc call the day it arrives.
4. Score it on value, feasibility, strategic fit, and risk, using the same four criteria applied to every other idea in the queue that fortnight.
5. Communicate the decision back to the claims handler directly, with the reasoning, whether it's promoted or parked.
6. Promote, if it scores well, into a brief specific enough that a delivery team can start without going back to the handler for basics.

Six weeks later, that handler sees a document-chasing assistant piloted on their team. They tell three colleagues. That is what a working funnel buys you: not just one good idea, but a reputation that submitting is worth the effort.

Pitfalls to watch for

• Scoring drift. If the group scoring ideas changes from cycle to cycle, or scores loosely one fortnight and strictly the next, contributors notice the inconsistency faster than you'd expect, and it reads as favouritism even when it isn't.
• The silent decline. A decline with no reason attached is worse than a slow yes. It costs you nothing to explain and costs you a contributor if you don't.
• Enrichment that overwrites the submitter's intent. AI-drafted fields should sit alongside the original submission, clearly marked as AI-generated, not replace what the person actually said. If the enrichment reframes the problem, you triage the wrong thing.
• A queue with no fixed cadence. "We'll review it when we get to it" is how queues die. Weekly or fortnightly, on the calendar, non-negotiable.
• No cap on scope. Some ideas are really strategy questions in disguise. Route those out of the idea queue and into the strategy conversation directly, rather than scoring them against tactical fixes.

Next steps

If you're running an idea funnel with no fixed review cadence, pick one now, weekly or fortnightly, and put it in the diary before you take another submission. If you have a backlog already, spend one session before your next review clearing duplicates and declining anything that's sat unscored for more than a month, with a reason attached to each decline.

The role Florence plays

Florence lets any member submit an idea and uses AI to enrich the non-mandatory fields, so submitters aren't blocked by having to estimate value or effort themselves. Editors get a clear triage queue rather than a scattered inbox: new submissions are flagged, scored against consistent criteria, and either promoted into a brief or archived with a stated reason. That reason is visible to the submitter, so the funnel stays honest and contributors keep coming back.