The problem
Nothing breaks on the last day. It breaks months later.
A renewal notice nobody knew about goes unanswered. A successor follows a procedure the last person knew was dangerous. By then, the institutional knowledge is unreachable at any price.
42%
of institutional knowledge is held exclusively by one individual
92%
of organisations fail to capture it before those people leave
50–200%
of annual salary is the cost of replacing an employee
Why the current handover cannot work
The standard artefact is a two-page document written by someone with one foot out of the door, reviewed by a manager who cannot know what is missing — because the missing thing is precisely what was never visible to them. Its completeness is unmeasurable by construction.
No function owns the question "did the knowledge actually move?" HR owns the offboarding checklist. Engineering owns the successor. Nobody owns the knowledge.
How it works
Four stages. Nothing installed.
No integrations, no OAuth, no access to your live systems at any point — which removes the security review that stalls this kind of work.
Ingest
You provide the offboarding export your IT team already produces for legal hold: mail, chat, tickets, documents, repository history. Files, not connections.
Map
We reconstruct who actually knows what, isolate the knowledge concentrated in the departing person, and rank it by the damage its loss would cause.
Interview
Short daily sessions — around twenty minutes — aimed only at the ranked gaps, quoting the person's own words back to them.
Verify
Every answer is scored against your records. Contradicted claims are flagged with the evidence attached, and gaps are re-interviewed while time remains.
What you receive
Four artefacts, before the last day
Delivered while the person is still on payroll — which is the only window in which any of it can still be fixed.
Knowledge Risk Map
Every area where they are the sole holder, ranked by business damage, with the evidence behind each finding.
Successor Handbook
How each area works, why it works that way, what breaks, who to call, and what never to do.
Verification Report
Every answer scored for specificity and coverage; contradicted claims flagged and cited.
Residual Gap List
What is still missing on the last day — and the scope for a paid consulting agreement if you want one.
Proof
What the engine found in a blind test
A synthetic company: a nine-year senior engineer, 62 artefacts, six pieces of undocumented knowledge deliberately buried in ordinary noise. The engine saw none of the answers in advance.
6 of 6
planted knowledge areas surfaced and correctly ranked
3 of 3
evasive answers caught, each citing the artefacts that refute them
0
false alarms on planted noise
7 → 2
at-risk areas when the same knowledge was shared evenly — the signal tracks concentration, not noise
Synthetic data — no real company information. We report each client's own numbers rather than recycling these.
What this is not
We will not claim more than we can prove
The boundaries we hold, stated before you ask.
| The claim we don't make | What we actually do |
|---|---|
| Not a lie detector | We detect claims your records contradict, answers too vague to be usable, and topics never covered at all. A human decides what a flag means. |
| Not surveillance | We analyse work artefacts your company already owns and retains. No activity tracking, no screen capture, no observation of anyone working. The departing employee sees what was extracted and can redact it. |
| It does not invent findings | Every flag must cite specific artefacts by ID, author and date. A flag that cannot cite is suppressed and never shown. The system is tuned to miss a real gap rather than manufacture a false accusation. |
Security
You are handing us sensitive knowledge. We designed for that.
Single-tenant isolation
A dedicated environment per engagement. No shared datastore across clients.
Scheduled destruction
The raw corpus is destroyed on a contractual clock after the engagement closes, with a certificate.
Your keys, your endpoint
Encryption under customer-managed keys, and analysis can run through your own enterprise model deployment.
No model training
Your data is never used to train models and never crosses between engagements — contractually.
In-tenant option
For the strictest requirements, the processing environment deploys inside your own cloud.
Stated plainly
We are early-stage. SOC 2 Type II is in progress and not yet complete. Better you hear it now than in week three.
Questions
The questions people ask first
Is this really AI, or just search?
Both, deliberately. The ranking layer is deterministic arithmetic, because you need to audit why we flagged one of your people. The AI does the language reasoning nothing else can: reading an answer, working out what it actually claims, and checking that against years of your own records.
Can it tell when someone is lying?
No, and we will never claim it can. It detects three things: claims your records contradict, answers too vague to be usable, and topics never covered at all. Each one arrives with the source attached, and a human decides what it means.
Is this employee surveillance?
No. We analyse work artefacts your company already owns and already retains — the same material an eDiscovery vendor would receive. No activity tracking, no screen capture, no observation of anyone working. The departing employee sees what was extracted and can redact it.
What if the AI invents a contradiction that is not real?
Every flag must cite specific artefacts by ID, author and date. A flag that cannot cite is suppressed and never reaches you. The system is tuned to miss a real gap rather than manufacture a false accusation.
Why not just run an exit interview with ChatGPT?
You can run an interview with a chatbot today. What you cannot do is rank what to ask about, or know whether the answer was true — both require holding the corpus as ground truth and reasoning against it. That grounding is the engineering, and it is what turns an exit interview into a verified handover.
Do we need to connect our systems?
No. You provide the offboarding export your IT team already produces for legal hold — mail, chat, tickets, documents, repository history. There are no integrations, no OAuth, and no access to your live environment at any point.
Get in touch
We are onboarding design partners now
If someone senior has recently given notice, that is the moment this is worth most. A first conversation takes twenty minutes.