AI in the Idea Bank: Helping Quiet Employees Speak Up

New to AI agents in HR? This post is part of our guide to AI HR agents — start there for the full picture, or keep reading for the deep dive.

Launching an idea bank from scratch? Categories, owners, response SLAs, and metrics are covered in our guide to running an ideas hub. This post is only about the AI layer on top of it.

"Every month at our company a dozen and a half genuinely good ideas come up. Two reach the idea bank, and usually from the same two people. The rest stay silent. Not because they don't think — but because they don't know how to phrase it."

That's what the HR director of a 600-person company told me. Six months earlier they'd launched a classic idea bank: a form with three fields, a "Under review" status, a feed to browse. A quarter later the portal held 47 ideas, 31 of them from three activists. Everyone else opened the page, saw the form, and closed it. It's a textbook scenario.

The idea bank is the most deceptively simple HR feature. It seems like all you need is a form + a list. In practice, most idea banks die within a quarter for three reasons, and AI helps solve two of them.

Let me share how we rewrote the idea bank in our product, and what changes when AI becomes not "a generator of ideas for the person" but an assistant with phrasing — a fundamentally different role.

Why an Idea Bank Stops Getting Submissions After a Quarter

Three causes of death, in order of importance:

First — the blank-form barrier. The same one as in a recognition program. The employee opens it, sees a white field that says "Describe your idea," and realizes they now have to formulate something structured, clear, non-trivial. That's too high a bar for an everyday work gesture. For most people, it's easier to close it.

Second — no status visibility. The employee submits an idea, and... silence. Two weeks later — silence. A month later — silence. It works like "no, but we won't tell you" — the most culture-destroying format. After one or two such experiences, the person never submits again.

Third — no closed loop. Even if an idea is approved, the employee often never learns it was actually implemented. The idea turns into a line of history, gets detached from its author, and six months later no one remembers it was proposed by Ivan from accounting. Ivan draws the obvious conclusion: "my participation doesn't matter."

As a result, the idea bank degenerates into two stable groups:

• Activists — 2–3 people with a naturally low phrasing barrier who don't need feedback. They always submit. They're the ones keeping it formally "alive."
• Everyone else — silent. And it's exactly their ideas the company loses.

AI helps the first group write cleaner and faster, but the main thing is that it removes the barrier for the second. That's the core point of AI in the idea bank: not "generation" of ideas, but access to the thoughts already in people's heads that don't make it out.

Inline Suggestions — Gray Text in the Idea Form

The same mechanic as in recognition. The employee starts writing an idea, and gray text appears in the field — a continuation of the sentence you can accept with Tab, edit, or ignore.

The context the AI relies on for ideas is different from recognition:

• Similar ideas already submitted. So the person doesn't repeat (or, conversely, can deliberately reference a similar one and build on it).
• Department specifics. Ideas from sales and ideas from developers look different. The AI factors in the role context and suggests appropriate language.
• The company's strategic priorities, if formalized. If the company's OKRs include "improve new-hire retention" and an employee writes an idea about onboarding, the AI connects them.

The employee begins: "I propose improving the onboarding process..." — and the AI continues in gray: "...by automatically assigning a buddy based on overlapping skills." The employee either accepts (Tab) or rewrites it their own way. In both cases they're no longer on a blank page but in editing mode — and that's a completely different cognitive load.

What's important: the AI doesn't suggest whole ideas. It completes what the person already started writing. This isn't "AI thought it up for you," it's "AI helped you not stumble on the phrasing." A fundamental boundary — without it, AI becomes an autogenerator that kills the value of the whole system.

And a separate function that works specifically for the idea bank: the AI checks for similarity at submission time. If your idea resembles five already submitted, the system offers: "there are similar ones, take a look — maybe comment there, or is this a new angle?" This reduces duplicates and at the same time raises the quality of discussion.

The Brainstorm Agent — a Dialogue Before Submission

Inline suggestions work when the employee already has an initial thought. But what if there isn't one, just a general sense of "something's off with process X on our team, it could be improved"?

For such cases — brainstorm mode. The employee clicks "Help me shape an idea," and a dialogue opens with an agent that asks 4 structural questions.

After the employee goes through these 4 questions (in free text, not in a fixed format), the agent assembles a structured draft of the idea: "Here's your idea, put together. Want to submit it like this, or edit?"

What this gives:

• Structure. Most ideas die at the formatting stage because the person doesn't know what to say. 4 questions give them a frame.
• Quality. What reaches the committee isn't "let's improve onboarding," but ideas with an explicit value proposition, reasoned obstacles, and a rough effort estimate. This saves review time many times over.
• Confidence. An employee who went through the 4 questions claims the idea as their own, because they answered. The AI didn't invent it — it structured it.

And importantly: the brainstorm agent is an option, not a mandatory function. An employee whose idea is already formed can submit directly through a short form. Brainstorm is for those who aren't sure. Not for everyone.

Where AI Stops: Statuses and the Closed Loop

AI removes the phrasing barrier, but it does nothing about the second and third causes of death. Those are process, not software intelligence: every idea moves through Draft → Under review → Approved → Implemented, the review stage has an SLA, an approved idea gets a named owner, and the author is notified at every transition — including a reasoned answer when an idea is rejected.

We walk through setting this up step by step in the ideas hub guide. Here, only what it changes for participation.

A Closed Loop Dramatically Changes Participation

The biggest insight I've seen in idea banks over recent years: having a closed loop — feedback to the author and a public feed of what's been implemented — multiplies the number of ideas submitted. The exact multiplier depends on the starting point: the deader the old "suggestion box," the stronger the effect.

This aligns with what industry reviews describe: classic suggestion boxes fail precisely because there's no closed loop — no owner, no timeline, no answer to the author, no visible implementation (Kainexus — Why Suggestion Boxes Fail, Itonics — Bright Ideas in Your Suggestion Box). At our clients, the typical picture — after introducing a lifecycle with mandatory notifications and a public feed of implemented ideas — is that within six months the flow moves from "3 activists and 8 ideas" to "30–40 authors and dozens of ideas a month." Exact numbers depend heavily on the program's starting state, so here it's an order of magnitude, not a "hospital average."

What it changes:

• Confidence they'll be heard. If an employee sees that 12 ideas were actually implemented in a quarter and it's written publicly, they develop trust: "I can submit, it won't disappear into a black hole."
• Social proof. When a colleague brings over a cup of coffee and says "hey, I submitted an idea about X, they implemented it, it'll be in the next release," that works more powerfully than any HR message.
• A sense of impact. This is, in fact, one of the basic psychological needs in SDT, which we wrote about earlier. When an employee sees their idea in the product, they feel agency. That's a rare and valuable feeling in a large company.

The public "Implemented Ideas" feed is a separate page showing all realized ideas from the last 90 days with the author's name. It's the second most important artifact of the idea bank after the process itself. New employees who land in the portal see this feed and immediately understand: "here, ideas actually get realized, they don't just pretend."

And separately: rejected ideas should also be closed out publicly. Not the names, but the count and the general reasons. "This quarter 47 ideas were reviewed, 18 implemented, 22 deferred, 7 rejected (the main reason — conflict with architectural constraints)." This dispels the suspicion that the committee "froze" the majority and clearly shows that the system works.

The Bottom Line

An idea bank dies for three reasons: the blank-form barrier, the lack of status visibility, the lack of a closed loop. AI helps with two — it removes the phrasing barrier and structures the thought through brainstorming. The third — the closure loop — has to be built into the process: an SLA at each stage, a specific owner, mandatory notifications to the author, a public feed of implemented ideas.

You can live without AI — the activists will write. Without a lifecycle and a closed loop, you can't — the idea bank will die within a quarter regardless.

And most importantly: AI shouldn't generate ideas for the employee. That's a fundamental boundary. AI helps confidently formulate what the person already has in their head but can't get out because of the barrier. If AI writes the "idea" itself, that's spam, not an idea bank.

Next, we'll dig into the redesign of the portal's content admin — banners + pages + news in one place, and why we spent a month on WCAG 2.1 AA accessibility.

If you want to see our AI-powered idea bank live, in the demo, follow the link in the card. A 5-minute demo, no presentations.

The AI level here is a setting, not a default — see idea bank.