Continuous improvement, run by AI
Your company already knows how to fix itself.
Every employee who sees a broken process has an idea. Almost none of them survive the trip to a decision. kAIzen turns a two-sentence observation into a reviewable business case, checks it against work already in flight, and tracks it to a verified KPI.
No signup. Loads with a seeded demo organisation.
An example of what kAIzen produces: the observation “Loan approvals still involve printing, signing, and re-scanning.” becomes a structured case with the problem stated as “Wet-signature step forces a paper round-trip inside a digital approval chain”, an affected population of 240 people · Retail Credit, an estimated impact of ~140 hrs/month, a priority of High · Medium complexity, and a suggested owner of Operations, Payments.
Why enterprise suggestion schemes stall
Large companies have run suggestion schemes for a century. The box on the wall became a portal, and the portal became a backlog. The failure was never idea collection. It was that processing an idea — structuring it, sizing it, checking it against everything already underway — cost more than the idea was worth.
That cost just collapsed.
One sentence in. One number out.
The same idea, at four points in its life. Nothing here is a mock-up of a process — it is the loop the product runs.
Step 1
You describe what's broken.
Two sentences, in the words the person actually used.
Submitted by a credit officer
Loan approvals still involve printing, signing, and re-scanning.
That is the entire input. No form, no template, no category picker.
Step 2
AI structures it against your company's own context.
Roles, systems and sites it already knows about.
Structured case
- Problem
- Wet-signature step forces a paper round-trip inside an otherwise digital approval chain.
- Current process
- Officer prints the packet, collects the signature, re-scans, re-uploads to the case file.
- Affected population
- 240 people · Retail Credit · 3 sites
- Proposal
- Extend the qualified e-signature already licensed for mortgage annexes to consumer loan approvals.
Step 3
It checks what's already in flight.
Vector search across every idea ever submitted.
Already in flight
- E-signature rollout for mortgage annexesIn progress · Payments82%
- Remove the paper archive step from credit filesApproved · Retail Credit61%
- Overlaps with work already underway. Merge or narrow the scope before this goes to a review board.
Step 4
And it doesn't stop until the KPI moves.
Ownership, status and the realised number at the end of it.
Idea #4128 · Retail Credit
- Submitted
- In review
- Approved
- In progress
- Done
Verified against the KPI
142 hrs/month realised
Measured after the change went live, not estimated at submission.
The loop, end to end
Step 1
You describe what's broken.
Two sentences, in the words the person actually used.
Step 2
AI structures it against your company's own context.
Roles, systems and sites it already knows about.
Step 3
It checks what's already in flight.
Vector search across every idea ever submitted.
Step 4
And it doesn't stop until the KPI moves.
Ownership, status and the realised number at the end of it.
Submitted by a credit officer
Loan approvals still involve printing, signing, and re-scanning.
That is the entire input. No form, no template, no category picker.
Structured case
- Problem
- Wet-signature step forces a paper round-trip inside an otherwise digital approval chain.
- Current process
- Officer prints the packet, collects the signature, re-scans, re-uploads to the case file.
- Affected population
- 240 people · Retail Credit · 3 sites
- Proposal
- Extend the qualified e-signature already licensed for mortgage annexes to consumer loan approvals.
Already in flight
- E-signature rollout for mortgage annexesIn progress · Payments82%
- Remove the paper archive step from credit filesApproved · Retail Credit61%
- Overlaps with work already underway. Merge or narrow the scope before this goes to a review board.
Idea #4128 · Retail Credit
- Submitted
- In review
- Approved
- In progress
- Done
Verified against the KPI
142 hrs/month realised
Measured after the change went live, not estimated at submission.
What you actually get.
Guided intake
Seven adaptive steps that skip themselves. Only the questions that matter for this idea, with live completeness feedback as you go.
Impact estimates with visible assumptions
Every number the AI produces shows its working. Managers argue with the assumption instead of distrusting the output.
Semantic deduplication
Vector search against every idea ever submitted. The overlap meeting stops being a meeting.
Owner suggestions
Ranked candidates drawn from a skills graph of your organisation, not a guess.
Goal focus
Type a KPI in plain language. Every idea in the portfolio is scored against it in seconds, and the ranking is yours to interrogate.
CI Cockpit
Funnel status, realised versus pipeline impact, review queue, ownership, activity.
Your compliance officer will ask what the model saw. You'll have the answer.
Full audit trail
Every status change, review, proposal edit and ownership transfer is written with actor and before/after state.
Prompt version pinning
Prompts, pipelines and output schemas are versioned data with one-click rollback. Reconstruct any decision on any date.
Model-provider abstraction
Anthropic, OpenAI-compatible, or self-hosted behind one interface. Procurement dictates the provider; the product doesn't care.
Schema-validated output
The model returns JSON validated against a typed schema, with automatic repair-retry. Nothing unstructured reaches your database.
Why this works now.
Idea-management SaaS was built pre-LLM.
HYPE, Brightidea, Wazoku. Their architecture assumes a human evaluates every submission. That assumption is now the bottleneck, and it is load-bearing in their codebase.
Process mining tells you where, not why.
Celonis and Signavio read system logs. The fix usually lives in the head of the person doing the work, and never reaches a log.
The suggestion box is a graveyard, and everyone knows it.
Employees stop submitting because nothing happens. Managers stop reading because they cannot triage the volume. Both are behaving rationally.
The questions buyers actually ask.
Does this replace our existing suggestion scheme or sit alongside it?
Alongside it, to begin with. kAIzen ingests submissions from wherever they already arrive — a form, a shared mailbox, an incumbent tool — and takes over the structuring, deduplication and tracking that your current scheme leaves to people. Most pilots run in parallel with the incumbent for a quarter so you can compare throughput on identical intake. Replacing anything is a decision you make afterwards, on evidence.
Where does our data go, and can we self-host?
The model provider is an interface, not a dependency: run against Anthropic, any OpenAI-compatible endpoint, or a model inside your own network. EU-hosted deployment is the default for European customers, and a fully self-hosted install is supported where procurement requires it. Ideas, KPI data and the audit log live in the database you nominate.
What stops the AI from inventing impact numbers?
Two things. Estimates are decomposed rather than asserted — headcount, frequency and duration are each shown and each editable — so an estimate you disagree with is an assumption you can correct rather than a number you have to trust. And the estimate is never the figure of record: an idea is only credited once the KPI it targeted has been measured after go-live. Submitted and realised are stored separately and always shown side by side.
Can employees submit anonymously, and how does that square with an audit trail?
Anonymity covers the author's identity, not the record. An anonymous submission is stored without an author reference, while every subsequent action on it — review, edit, approval, ownership transfer — is still written with its actor and before/after state. The trail stays complete about what was decided and by whom, and stays silent about who raised it.
How long does a pilot take to stand up?
A seeded environment carrying your KPI structure and one division's org data is typically live within two weeks. The pilot is then scoped to 90 days, because that is roughly how long it takes for an approved change to go live and produce a number you can actually measure rather than project.
What happens to ideas that get rejected?
They keep a reason and stay searchable. A rejected idea is written with the rejecting actor, the rationale and the date, and it stays in the vector index — so when the same problem is raised again eighteen months later, the new submission surfaces the earlier decision instead of reopening the same debate from scratch. The submitter sees the reason, which is the single biggest determinant of whether they ever submit again.
Start with one process. Prove one number.
A pilot scoped to a single KPI, in one division, in 90 days. If the number doesn't move, you'll know exactly why.