What we build
Every layer of anAI Operating System.
Vulc builds custom AI agents and the layers that make them dependable: workflows, company knowledge, live data, integrations, approvals and a control layer, inside accounts you own. You do not need every layer. The examples below lean on e-commerce, because that is the operation we know in our hands, and each one says what the same build looks like for an agency, a software company or a service business. What you sell changes the details, not the shape.
AI Operating System
Seven layers.Take the ones you need.
An AI Operating System is what several agents become when they share one body of company knowledge, one set of live data and one approval layer. Each layer below is a thing we build; none of them is mandatory. The sections that follow show what each one looks like in practice.
AI agents
Specialist roles with defined responsibilities, coordinated by an orchestrator where the work needs one.
Workflows and automations
How each job gets done, and how work moves between agents, systems and people.
Company knowledge
Policies, tone, product facts and the rules in people's heads, written down where an agent can use them.
Live business data
Orders, tickets, leads, spend and margins, read from the source rather than an export.
Integrations
The systems agents read from and act inside, with credentials scoped per integration.
Approvals and governance
Gates where a person sees the draft and its context before anything moves.
Control layer
The screen where you see status, outputs, approvals and system health. Yours, or one we build.
AI agents
Agents work the queue.You sign the sends.
Support, retention, ads, listings, research and the rest, built on your tone and your edge cases. They read the queue, draft the work and prepare every action with full context attached. In a store that is the where-is-my-order queue, the refund thresholds and the review replies; in an agency or a software business it is client threads and the questions your docs almost answer. Anything that touches a customer or moves money stops at an approval gate first. You review in one place and ship with one click.
Approvals view · illustrative
Refund request over threshold
Waiting on you
Hi Lisa, you are right, that took too long. The refund of €184 is approved on our side and lands in 2 to 3 working days.
#8391 · support agent · policy: refunds above €150 need sign-off
- 14:31 · reply sent for #8409 · standing approval
- 14:28 · tracking link sent for #8412 · standing approval
approval-first. nothing ships without you.
Typical build
- Support agent on your helpdesk, trained on your real tickets
- Approvals view: one place to review and sign off
- Escalation rules for the cases that stay human
- Audit log of every draft, decision and send
- Runs in your environment, on your accounts
Workflows and automations
The busywork, wired.The judgment, yours.
The repeating work between your tools, connected end to end: tagging, routing, follow-ups, reconciliation. In a store that is the path from order to fulfilment, the review requests and the daily books; elsewhere it is onboarding, renewals and the reporting nobody wants to assemble on a Friday. Every run is logged, and every step that moves money or emails a customer waits for a yes. When something fails, it fails loudly, with a paper trail.
Flow runs · illustrative
- 09:14 · run #312 · follow-up sequence · 3 drafts prepared
- 09:02 · run #311 · review requests · 14 sent on standing approval
- 08:47 · run #310 · daily reconciliation · clean, no variances
Discount above threshold
#312 · held at the gate, draft ready
Needs your call
approval-first. nothing ships without you.
Typical build
- Flow mapping from your real week, not a template
- Gates wherever money or customers are touched
- Run log with a full audit trail
- Failure alerts that name the exact step
- Handover documentation for every flow
Company knowledge
The rules in people's heads,written down.
An agent is only as good as what it knows about your company. During the build, the knowledge it needs is gathered from your own material, tested against real cases, and tightened until the agent applies it the way your best person would. It is kept where you can read and change it.
Policies and rules
Return windows, refund lines, discount limits, escalation paths: the rules an agent must apply the same way every time.
Tone and past replies
How you actually talk to customers and clients, taken from what you already wrote rather than from a style guide nobody reads.
Exceptions already solved
The judgment calls your team has made before, so the agent recognises the shape of a case it should not decide alone.
Live business data
Read from the source,not from an export.
Orders, tickets, leads, spend and margins are read from the systems they live in. An agent acts on what is true now, and the number you check at 8am is the same number the agent used at 7:45.
Direct connections
Agents read orders, tickets, leads and spend from the systems they live in, not from a nightly export that is already stale.
Scoped credentials
Each integration gets the narrowest access the job needs, granted by you and revocable by you.
One source of truth
The number an agent acts on is the number you see on the control layer, because both read the same source.
Integrations
Wired into the stackyou already run.
A system is only as useful as what it can reach. The platforms below are what we have built against directly, on real production work, and they lean on commerce stacks for the obvious reason. Beyond them, another 101 tools publish documented public APIs, from helpdesks and CRMs to finance and project tools; a representative slice is below and the complete catalogue is on the integrations page.
If the required access and endpoints are available, we can integrate the system into the build. We verify that before scope is locked, so nothing surfaces halfway through.
The complete integration catalogueBuilt against directly
Commerce and payments
- Shopify
- Stripe
- Recharge
- Loop
- Chargeflow
Advertising
- Meta
- Google Ads
- TikTok
Email and retention
- Klaviyo
Analytics and search
- Google Analytics
- Search Console
Reviews and social proof
- Yotpo
- Loox
Documented public APIs · a slice of 101
Helpdesk and support
- Gorgias
- Zendesk
- Intercom
- and more
CRM and sales
- HubSpot
- Salesforce
- Pipedrive
- and more
Work and project management
- Asana
- Monday.com
- ClickUp
- and more
Finance and accounting
- Xero
- QuickBooks
- Sage
- and more
Email, SMS and automation
- Mailchimp
- Brevo
- ActiveCampaign
- and more
Engineering and product
- GitHub
- GitLab
- Sentry
- and more
How integrations connect to agents
Support agent · illustrative
- Shopify order
- Support agent
- Policy check
- Draft reply
- Your approval
- 14:31 · order #2201 read from Shopify · refund above your threshold
- 14:31 · return policy checked · reply drafted in your tone
- 14:32 · routed to you · waiting on approval
approval-first. nothing ships without you.
Research agent · illustrative
- CRM prospect
- Research agent
- Qualification
- Personalised draft
- CRM update
- 09:04 · new prospect read from the CRM · six sources gathered
- 09:06 · qualified against your criteria · outreach drafted
- 09:06 · CRM record updated · draft held for your review
reads, checks, drafts, updates. you send.
Two flows, illustrative. Every category, every tool and a third flow are on the integrations page.
Approvals and governance
Approvals keephumans in control.
Autonomy is decided per action, not per system. Routine work can earn the right to run alone; anything that touches customers or money stops at a gate with your name on it. You set the thresholds at the start and move them as trust builds, and every action leaves a record.
Per-action autonomy
Each action is either allowed on its own, allowed within a threshold, or held for a person. You decide which, per action.
Standing approval
Routine, low-risk work earns the right to run without a gate once it has proven itself against real cases.
Thresholds
Everything above your line arrives as a prepared draft with its context attached, one decision from done.
Audit trail
Which agent acted, on what, with which permission, and who approved it. Every action on record.
Control layer
See the system.Direct it.
The control layer is where you see status, outputs, approvals and system health, and where the numbers you run on live, down to true profit after every cost: margin per product once ad spend, fees, shipping and returns are out for a store, profit per client once the hours are counted for an agency or a service firm. If you already have a dashboard that does this, we build the agents and workflows around it. If you need one, we build it around your decisions, so the number you check at 8am is the first thing you see.
Control layer · illustrative
refreshed 5m ago
€8,412
revenue today
€2,180
acquisition spend
31%
net margin
3
waiting on you
net margin, last 7 days
illustrative layout. yours is built around your numbers.
Typical build
- Approvals view: one place to review and sign off
- Profit per product, service or client, per channel, per day
- Direct connections to your stack, no exports
- Views per role: owner, ops, marketing
- Only built when nothing you have does the job
Modular by design
You do not needevery layer.
Vulc does not force you into a fixed product. A build takes the shape your operation needs, and the next layer joins the same approval surface when you are ready. Four shapes we build often:
Your dashboard, plus an AI workforce
Agents, workflows and integrations built around the interface you already run.
Three specialist agents
A small AI team on the roles that eat the most hours, sharing one approval layer.
One high-value workflow
A single process, end to end, from the moment work arrives to the moment it is done.
A full AI Operating System
Every layer, from the AI team to the control layer, in infrastructure you own.
The full AI Operating System
The whole operation.One owned layer.
Agents, workflows, knowledge, data, integrations, approvals and the control layer working as one system, with an orchestrator deciding what runs when. The one we built for ourselves runs a commerce operation, so a store is the version we can show you today. The orchestrator, the specialists and the gates are the part that carries, which is why the same layer fits an agency roster or a software company just as well. It arrives as working software in your environment, with the keys.
Orchestrator · illustrative
- support · queue clear · 2 drafts in your inbox
- email · winback campaign drafting
- ops · inventory sync clean
- reporting · weekly margin report ready
Weekly ad budget shift
prepared by the media agent, context attached
Needs your call
approval-first. nothing ships without you.
Typical build
- Orchestrator plus the specialists your operation needs
- One approval surface across everything
- Runs in your cloud, on your accounts, with your keys
- Documentation and runbooks for the whole system
- Optional retainer to tune and extend it
Ownership
Built in your accounts.Owned by you.
Every layer above is built in your repository, on your accounts, with credentials you grant and can revoke. After handover the system keeps running if Vulc's access is removed. That is the test we build to.
- Your code
- Your database
- Your AI accounts
- Your integrations
- Your credentials
- Your data
- Your infrastructure where practical