A record of every AI conversation at your company. See what staff ask AI and what your own AI does, get a warning when something looks wrong, and find anything by name, tool or date.

Most AI use leaves no record anyone can use.

• 52% of workers use AI tools their employer hasn’t approved.

  Okta survey, May 2026

• A court ordered OpenAI to hand over 20 million ChatGPT conversation logs.

  ABA Journal, January 8, 2026

• Federal government departments must keep records of decisions made with AI, and of every time a person overrules it, safe from tampering.

  Directive on Automated Decision-Making, sections 6.2.7, 6.3.4 and 6.3.11

How it works

Step 1

Every AI conversation goes on one list

halfway connects to ChatGPT, Copilot, Claude and other AI tools on company computers, plus any AI your company builds itself. Every question, answer and action is recorded, with names hidden.

Step 2

Flags for what needs a look

An AI tool you haven’t approved, private or client information typed into AI, or your own AI sending email without a person’s OK. You choose what gets flagged.

Step 3

Search, or just ask

Look things up by name, tool or date, or type a question like “What is the HR team using AI for this month?”

What each team uses AI for, and what your AI did

Teams and people

Each team’s main uses and tools at a glance. Nobody is named unless two people approve it.

Your own AI

Every action your company’s AI took, and whether a person approved it.

Screens from the halfway app, with sample data.

One record, two ways to read it

The same record in plain words, and as an engineer sees it.

{
  "record_id": "rec_01JB7Q4ZK3",
  "time": "2026-10-28T14:02:11Z",
  "source": "m365_copilot",
  "actor": {
    "type": "person",
    "role": "hr_analyst",
    "name": "[masked]"
  },
  "event": "prompt",
  "ai_system": "sys_hr_copilot",
  "files": 12,
  "personal_info": ["name", "email"],
  "flags": [{
    "type": "personal_info",
    "status": "closed"
  }],
  "review": {
    "decision": "approved_use",
    "at": "2026-10-28T15:30:02Z"
  },
  "prev_hash": "b84d0e…07aa",
  "hash": "3f9a71…c21e",
  "timestamp": "rfc3161:…"
}

Sample data.

Nobody can quietly edit the list

Each record is stamped with a code made from the one before it, and stored where it can’t be edited. If anyone changes an old record, the codes stop matching and halfway shows exactly where.

1. 1. Written

   #10413f9a

   #1042b84d

   #104307aa

   Records are linked in order.

2. 2. Changed

   #10413f9a

   #1042edited

   #104307aa

   Someone edits an old record.

3. 3. Caught

   #10413f9a

   #10429c1e ≠

   #1043breaks

   Its code no longer matches, so every record after it breaks. halfway shows where.

The plan

1. Built

   The first version

   The halfway app for Windows, the browser and desktop recorders, the gateway and agent kit for your own AI, flags, search, reports, the staff policy, and a check that proves nothing was changed.

2. Now

   Pilots

   Working with a first few teams, signing the installers, testing the company feeds from Microsoft, OpenAI and Anthropic, and the French version.

3. Later

   Rule checks

   For organizations that need them: flags tied to federal AI rules, the EU AI Act and hiring laws.

Your data

• Kept where you choose

  halfway keeps records on a computer or server you choose. The hosted version will run in Montréal, with Calgary as backup.

• Recorded openly

  Staff are told. halfway also shows you any computer or AI tool it isn’t fully recording.

• Names hidden

  Masked in search. Revealing a name takes two approvals, and every search is recorded.

• You decide how long

  You choose how long records are kept and when they’re deleted, without asking us.

• Never sold or used for training

  We won’t sell your data or train AI on it.

• Security audit

  An outside security audit (SOC 2) is planned before the first paid customers.