How it works
From your real work to skills your AI can run.
Praxeum turns how your company actually operates into verified, executable skills, with a human accountable for every one, and the evidence attached.
Prefer to see it? Walk through the demo. Best on a laptop.
The loop at a glance
Build the skill, then run it: one accountable trail.
Read-only
Connect, read-only
You grant read-only access to the exact Slack channels, Drive folders, Notion spaces, and meeting series you choose, with more sources on the way. Scoping is employee-visible, and permissions travel with everything we derive.
Conversations
Chat threads: the exceptions, the corrections, the "no, do it this way".
Docs
Wikis and shared drives: the policies and runbooks, cited back to the page.
Meetings
Call transcripts, for the meeting series an organiser chooses to include.
The instant tier
Convert what’s already written
Documented procedures become draft skills in days, cited back to the source doc. Real, day-one value while the deeper mining runs.
Finding the patterns
Find how work actually happens
On top of your policies, Praxeum finds the recurring steps, the exceptions, and the "we never do X" rules that never reached a wiki. It reads your conversations for that, because the corrections ("no, do it this way") are where the real procedure lives.
- Refund exceptionsseen 12× · 4 people · 3 wks
- Customer onboardingseen 9× · 3 people · 2 wks
- Billing exceptionsseen 6× · 2 people · 4 wks
'No, do it this way.' The corrections are where the real procedure lives.
You decide
Doc vs practice
Where your doc says one thing and your team does another, we show you the conflict with the evidence attached, and you decide on purpose. No conflict? Praxeum still builds the skill from your policy.
Doc vs practice
Your policy says 30 days. Your team has refunded annual plans up to 60 days, 12 times since January. Which is right?
annual plans → refunds up to 45 days
A test before approval
Tested on your past cases
Before anyone approves a skill, Praxeum replays it on real past cases and grades each result against a rubric drawn from how your team decided them. The approver sees the score and every miss, not a promise.
refund-exceptions@v3 Replay test
- Right amount
- Inside the policy window
- Sign-off requested before money moved
- Reply matches team tone
38 of 40 past refunds matched the team's decision
- #4102 Team refunded at 60 days, skill said no. See the conflict
- #4117 Customer on a legacy plan, not in the skill yet.
Ready for review · Sarah Chen
The skill
One human approval
A skill is the actual file your AI reads and follows. A person reviews it, edits it if needed, and approves it, with the replay results in front of them. Nothing reaches an AI co-worker until they do.
# refund-exceptions
Decide a customer refund from the plan and timing.
## steps
- Check the plan type and timing.
- Within the window? annual 45 days · monthly 30.
- Never after a chargeback. Ask a person to sign off.
## tools
stripe · refunds.create
Added to your catalog
- refund-exceptions@v3
- onboarding-runbook@v5
- tone-of-voice@v2
- billing-exceptions@v1
No new AI to buy
Served over MCP
The approved skill is served over MCP to the AI your team already runs, which carries the version it ran. Each skill sits on its own rung of the permission ladder, and at sensitive forks it asks a human to sign off, in Slack.
- Claude
- ChatGPT
- Cursor
- Copilot
- Gemini
- read
- suggest
- execute with sign-off
- execute
Markdown and code
The skill interpreter
Skills are Markdown, the format the big AI platforms are standardising on, and a model can read Markdown and still improvise. So Praxeum’s skill interpreter binds the parts that must never vary to reviewed, tested code. The replay test covers the coded parts too.
# refund-exceptions
Decide a customer refund from the plan and timing.
## steps
- Check the plan type and timing.
- Within the window? annual 45 days · monthly 30.
- Never after a chargeback. Ask a person to sign off.
## tools
stripe · refunds.create
// the parts that must never vary const WINDOW = { annual: 45, monthly: 30 } function inWindow(days, plan) { return days <= WINDOW[plan] }
- annual at 41 days → in window, sign-off as always
- annual at 60 days → outside the window, escalate
- after a chargeback → never
A named bundle with a job
Skills become AI co-workers
Approved skills group into a co-worker: a named bundle of approved skills with a job, reachable where your team already works. Each skill keeps its own rung, so Carmen can chase an invoice on her own but must ask before a refund goes out. Co-workers run on the AI you already license, on your connection.
- refund-exceptionssign-off
- billing-exceptionssign-off
- invoicingexecute
- renewal-followupexecute
- pricing-exceptionssign-off
- lead-qualificationsuggest
- onboarding-runbookexecute
- leave-requestssign-off
- escalationssign-off
- ticket-triageexecute
- incident-runbookexecute
Four examples. Yours will look like your team.
Recall, not governed action
Ask your company brain
The same reading that builds your skills builds a brain your team can just ask, in Claude, ChatGPT, Slack, or Teams. Every answer is limited to what that person can see. From an approved skill it names the version; anything else is marked unverified until a person approves it.
It stays honest
Drift detected
Praxeum keeps reading. When the way of working moves, or a cited source changes, the skill goes back for review, never silently averaged in. Until a human rules, the contested step is served as ask-a-human.
“We’ve been holding annual refunds to 30 days since September. 45 was too generous.”
seen 8× · 3 people · since Sep#support-opsannual plans → refunds up to 45 days
annual plans → refunds up to 30 days
Approved as v4. Back in service, and your co-workers pick it up.
The trust is in the details
Four artifacts do the governing.
Citations, versions, sign-off, and an audit trail: the real objects a Praxeum skill carries, so a co-worker acting on your behalf, on the AI you already run, is always accountable.
Every claim cites its source
refund-policy#support-ops · Apr 12Co-workers carry the version they ran
v3 · approved by Sarah Chen · Jun 30evidence current as of Jun 30Sensitive actions ask a human
Approve a $240 refund on an annual plan booked 41 days ago?
refund-exceptions
Every call is on the record
The details that make it safe
Evidence grade
Every draft arrives with citations and an evidence grade: seen enough times, over enough weeks, by enough people. Repeated behavior is evidence a practice exists, not proof it should be policy, so a human decides before any agent can run it. The reviewer edits, approves, assigns an owner, or rejects.
Executable skills
A served skill is executable, not just readable. It carries a machine-readable action binding (which tool, which action), its required tools, and a wired-and-verified readiness state, so an approved decision becomes the right tool call with almost no glue.
Your credentials stay yours
The agent executes on its own MCP connection. We never hold your credentials and never run the action ourselves.
Where the reasoning lives
Under the hood we assemble conversations into episodes, cluster the recurring ones, and induce the real procedure. Conversations hold the reasoning your wiki never captured. Operational sources (ticketing, CRM, and more) are complementary and growing, so the record of what actually happened can ground the record of why.
The week-one promise is staged and honest: documented procedures convert immediately; the highest-value, undocumented knowledge surfaces progressively over weeks; never a finished brain on day one.