Skip to content
Playbooks

AI workflows driven by
internal knowledge.

A playbook defines the precise steps your assistant executes alongside the governance boundaries it cannot cross. Team members capture their subject-matter expertise in plain language, enabling the assistant to carry out routine operations with complete autonomy tailored to your needs.

Refund order 4417Overnight refund queue

Playbooks

Your way of doing the work, written down — steps an agent follows, rules it cannot break.

v4
Refund a customer

Look up the order, check it is inside the returns window, refund it and tell the customer.

24 runs · 92% clean · 2dteam
v2
Onboard a new supplier

Collect the documents, check them against the list, open the record and hand it to purchasing.

11 runs · 100% clean · 5d
v3 · unpublished changes
Weekly support themes

Read last week's tickets, group what they are really about, write the Friday summary.

9 runs · 89% clean · 4d

The playbook shelf in the app. The line under each one answers the only question worth asking: how much has it run, and how much of that went the way it was written.

In conversation

Via AI assistant or a plain instruction. Transform your workflows.

The same job, “refund this order and tell the customer”, done two ways. The difference is not how clever the agent is. It is whether you want it to follow certain instructions every single time.

Asking

It works it out each time

Refund order 4417 and let the customer know.

  • A different order every time
  • Every tool it has, at every moment
  • Sends when it decides it is ready
  • Your rules are a sentence it might recall

Fine while you are watching. Less fine at 3am.

Following a playbook

It follows what you wrote

  1. 1

    Find the order

    look up order· Stripe
  2. 2

    Check the returns window

    check policy· Notion
  3. 3

    Refund it

    issue refund· Stripe asks you first
  • Never put card details in a reply

    checked on the action itself, before it happens

Autonomous background processing

Start a complex playbook and close your screen. Tasks complete in strict sequence across your defined schedule, routing results straight to the team channels you already use.

Mandatory approval checkpoints

Flag sensitive actions to require explicit human authorisation. Unattended operations wait securely until approved, which rules out unauthorised actions and guesswork.

Enforced system guardrails

Controls execute directly on system calls before they are fulfilled. Governance stays immutable, operating entirely independently of language model behaviour.

The steps

Knowledge-driven workflows.

Playbooks capture how work gets done, combining team expertise with AI capabilities to execute context-aware operations.

Plain text instructions
Skip visual builders, complex diagrams and IT tickets. Define conditional logic naturally by putting the condition at the start of the sentence.
Granular tool scope
Restrict integration access strictly to the steps that need it. If an early step only reads data, the agent never holds the access that would let it move money.
Explicit approval checkpoints
Pause workflows at the steps you designate as critical. Overnight background runs wait securely for human confirmation rather than making unauthorised choices.
Deterministic exception handling
Codify edge cases explicitly, so execution halts immediately with clear diagnostic feedback instead of being improvised in the moment.

Refund a customer

What this playbook does

  1. 1

    Find the order and note what the customer paid

    look up order· Stripe
  2. 2

    Check it is inside the returns window

    30 days from delivery, not from the order date. A gift order runs from the day the recipient got it.

    If the item was marked damaged on arrival, hand it to the claims desk the run stops here

    check policy· Notion
  3. 3

    Refund the amount they paid

    issue refund· Stripe asks you first
  4. 4

    Tell the customer what happened

    reply to customer· Zendesk

Refund a customer

The rules it follows

  • Never refund more than the customer paid

    checked on the action itself, before it happens · watches step 3

  • Never put card details in a reply

    checked on the action itself, before it happens · watches step 4

  • Refunds over €200 need a person

    the process stops and a person approves · watches step 3

Refund order 4417

Stopped by a playbook ruleNever refund more than the customer paid

The refund was for €900. Stripe says the order was €240.

The action was not taken. This rule is checked outside the assistant, on the action itself, so it cannot be talked past.

The rules

Immutable policy guardrails.

Policy enforcement takes place within the console, evaluated directly against the underlying execution layer rather than inside agent prompts. Enterprise administrative rules and local playbook policies both apply, so a denial from either authority stops the execution instantly.

Action-level evaluation
Security policies validate every operation before it executes. Nothing unauthorised is granted access, and only precisely screened data is let through.
Flexible approval workflows
Configure policy evaluations to trigger an automated system approval or to require an explicit manual sign-off, based on the risk involved.
Dynamic scope protection
Policy rules evaluate the surrounding context dynamically, so new workflows stay covered automatically without extra manual setup.
Hierarchical enforcement
Playbooks may apply stricter local conditions, but they can never override or loosen the baseline enterprise controls set in the console.

TBAC engine

Dynamic execution control.

Access permissions are enforced dynamically at runtime, protecting sensitive systems and keeping you in complete operational control.

  • Just-in-time tool provisioningIntegration access is restricted strictly to what an individual step requires, keeping unnecessary capabilities entirely hidden from the assistant.
  • Comprehensive parameter screeningThe tool, the incoming arguments, the external connection and where the data is headed are all weighed in a single integrated evaluation.
  • Inherited enterprise governanceOrganisation-wide defaults establish the core boundaries that departmental policies build upon, so the most conservative restriction always governs execution.

Policy check

The rule

Customer data does not leave the company. Payments above €10,000 wait for a person.

Three calls, one rule

  • create_invoice€4,200, for an existing customerAllowed
  • transfer_payment€24,000, above the team's thresholdNeeds approval
  • send_emailCustomer list attached, to an external addressBlocked

The same check runs before every call and afterwards stands as a step in the decision chain.

The run

AI that follows your lead and learns from your team.

An action is not a black box you read about afterwards. The steps tick off visibly and a skipped step is shown. If an action has to stop, the system holds it until someone approves it, and records that.

Transparency
Above the message box you see the playbook, the step and how far along it is. That goes away when the action ends.
Actions cannot simply be skipped
“Didn't apply: the item was not marked damaged” stays on the record, so you can see why this case came out the way it did.
Every fact names its source
Looked up in Stripe, provided by you, or the agent's own words. The verbatim line it was lifted from sits underneath it.
Only real facts can be a yardstick
A rule may measure against something looked up in a real system, never against a number the agent wrote itself. A rule that trusts the thing it is checking is not a rule.

Refund order 4417

  1. Find the order and note what the customer paid
  2. Check it is inside the returns window
  3. Hand it to the claims deskdidn't applythe item was not marked damaged on arrival
  4. Refund the amount they paid
  5. Tell the customer what happened
What it established
order totalStripe · looked up
240.00

Order total: €240.00

delivered onStripe · looked up
14 July 2026

Delivered 14 Jul 2026

customeryou provided this
Renate Böhm, order 4417
why they askedthe agent's own words
The size ran small and she has already reordered.
Refund a customerRefund the amount they paidstep 4 of 5

Publish and share

Internal processes which improve based on your needs.

A draft stays private until you publish it. Only the latest published version is ever active, and every release carries a note saying what changed.

Refund a customer

The record

24runs

finished
19
stopped
3
blocked
1
running
1

Releases

  1. v4

    Refunds over €200 now ask a person

    2 Aug · Mara K.

  2. v3

    Added the claims-desk stop for damaged items

    17 Jul · Mara K.

  3. v2

    The reply moved to Zendesk

    3 Jul · Mara K.

Runs on its own

Every weekday, 8am

The overnight refund queue

Next Mon, 08:00

Drafts are private
Nobody else has access. Once you publish, colleagues read exactly what the assistant carries out.
Every release keeps its note
The note is written from what changed between versions. You keep it, edit it, or write your own.
The status is fact
How often the playbook ran, was checked or was stopped: everyone with access to it gets the same figures.
Unattended does not mean uncontrolled
Start a playbook on a schedule and it is attributed to you and checked. Exactly as if you had started it by hand.
Get started

Transform a playbook and convince yourself.

Bring a use case you would like to hand over in future. We will show you how it works.