Skip to content
BlackOak AgencyBlackOak.agency

// AGENTIC COMPANIES

The agentic organisation

AI agents that don’t just answer questions but pick up work: on a schedule, with a role, a budget and a manager. Here’s what that looks like, which tools exist for it, and how we help an organisation set it up.

Research updated on September 11, 2026

NLEN
40%

of large organisations are scaling AI agents, up from 27% a year earlier[2]

85%

of organisations want to be agentic within three years; 76% say current operations can’t support it[4]

>40%

of agentic AI projects will be cancelled before the end of 2027, Gartner expects[3]

28%

of Dutch companies under 50 staff use AI structurally; mid-sized 42%[8]

Sources at the bottom of the page.

// WHAT IT IS

From assistant to colleague

An agentic organisation is a company where AI agents have their own place in the org chart. They get goals instead of prompts, work through multiple steps with real tools, and report to a human or to another agent. The human shifts from doing to steering: setting goals, drawing boundaries, judging outcomes. Most companies sit on rung 1 or 2 today. What separates rung 3 and 4 isn’t the model, it’s the organisation around it.

  1. 01

    Assistant

    A chat window. Someone asks a question, gets an answer and pastes it somewhere themselves.

    The humandoes all the work, AI helps with one step.
  2. 02

    Workflow

    Fixed steps, drawn out in advance. AI fills in one or two of them: summarising, classifying, drafting an email.

    The humandesigns the flow and steps in when it breaks.
  3. 03

    Autonomous agent

    One goal, its own tools, several steps in a row. The agent picks its own route, runs on a schedule and reports what it found or did.

    The humansets the goal, judges the result.
  4. 04

    Agentic organisation

    Several agents with roles, reporting lines and budgets. They hand work to each other, ask for approval where required, and everything is logged.

    The humansits on the board: goals, boundaries, approvals.

// THE BUILDING BLOCKS

What an agentic organisation needs

Whether you use Paperclip, our AI-Office or something custom, these eight parts always come back. Leave one out and you get a demo that never reaches production.

Org chart & roles

Every agent has a title, a role instruction and someone it reports to: a human or another agent. Delegation follows those lines, not criss-cross.

Goals & tasks

Work hangs off a goal, not a prompt. Tasks are checked out atomically so two agents never do the same thing, and blockers are visible.

Heartbeat & schedule

Agents don’t run continuously. They wake on a schedule or on an event, check what’s pending, do their work and go back to sleep. That keeps cost and risk bounded.

Tools & MCP

Access to email, CRM, GitHub, accounting or your own software via MCP servers. Ticked per agent, with keys stored encrypted.

Skills & work instructions

Reusable instructions in plain language: how we write a quote, what a good review looks like. Write it once, every agent that needs it gets it.

Budget & cost

A monthly budget per agent with a hard stop. Cost per run, per goal and per model visible, so you know what a task really costs.

Approvals

Human-on-the-loop: not every action, but the consequential ones. Hiring a new agent, an email to a customer, a change in production. Those wait for a human.

Log & audit

Every run, every decision, every tool call recorded and readable afterwards. That’s not just nice to have; from August 2026 the EU AI Act requires it for more and more systems.

// THE TOOLS

Paperclip, our AI-Office and the rest

The market splits into two layers. Frameworks and runtimes (Claude Agent SDK, CrewAI, OpenClaw, Microsoft Agent Framework) build one agent. A control plane such as Paperclip or our AI-Office manages several: who does what, when, for how much money and who is watching. Paperclip puts it well: if OpenClaw is the employee, Paperclip is the company.

OPEN SOURCE · MIT

Paperclip

The open-source control plane for a company of agents.

  • Launched 4 March 2026; over 80,000 GitHub stars and 14,000 forks in six months.
  • Org chart with a CEO agent, reporting lines and delegation up and down. The board (you) approves hires and strategy changes.
  • Heartbeats of e.g. 4, 8 or 12 hours; monthly budget per agent with a warning at 80% and a hard stop at 100%.
  • Bring your own agent: Claude Code, Codex, Gemini CLI, Cursor, OpenClaw or any HTTP bot. Node.js + Postgres, self-hosted, no account required.

Fits: a team that wants several agents collaborating on product development, content or research, with a real hierarchy.

Caveat: six months old, moving fast and without a managed cloud edition. You host, secure and monitor it yourself. We’re happy to do that with you.

Book a call about Paperclip →Compare with alternatives →
EIGEN BOUW · IN PRODUCTIE

BlackOak AI-Office

Our own office for scheduled Claude agents. Small, self-hosted, in daily use.

  • Agents with a role instruction, a cron schedule and a set of allowed MCP servers and skills. Runs in Docker on your own server: Next.js, Postgres and a worker.
  • An inbox that only shows runs needing attention. Agents can call on colleagues and reply, and you read the whole chain.
  • Cost per run and per conversation visible; connection keys stored encrypted. Built-in tools for video analysis and Outlook mail.
  • Our own roster: Sentinel (security), Beacon (Lighthouse scores of our sites), Forge (build and fix), Foreman (nightly review loop) and a Code Reviewer.

Fits: an organisation that wants to start with a few agents that monitor or deliver something, and wants to own the code.

Deliberately small: no org chart with dozens of agents and no model families other than Claude. If a team outgrows it, we move it to Paperclip.

Ask for a demo →See our approach →

Side by side

AspectPaperclipBlackOak AI-OfficeFrameworks & SDKs
What it isControl plane: a company of agentsControl plane: a small team of scheduled agentsBuilding blocks for one agent inside your own software
Org chart & delegationYes, with a CEO agent and reporting linesFlat: agents call on each other via handoffBuild it yourself
Schedule / heartbeatHeartbeats + triggers on assignmentCron per agent, multiple schedulesBuild it yourself
Budget & hard stopPer agent per month, auto-pauseCost per run visible; per-run capBuild it yourself
ApprovalsBoard approval for hires and strategyInbox: findings wait for a humanBuild it yourself
Agents run onClaude Code, Codex, Gemini CLI, Cursor, OpenClaw, HTTPClaude Agent SDKAny model, your code
Hosting & dataSelf-hosted, Postgres, no accountSelf-hosted in Docker, EU server of your choicePart of your application
LicenceMIT, free; you pay model usage and serverSource from us, managed by youMostly open source

Our advice in one sentence: start with one or two scheduled agents (AI-Office or Paperclip) and only move to a full org chart once agents need to hand work to each other.

Why we choose AI-Office and Paperclip

We put Multica, Gas Town, Cofounder, OpenAI Frontier, Microsoft Agent 365 and the frameworks side by side. Our default lands on these two because:

  • both run on your server, with your data and without a per-seat licence;
  • Paperclip is the only open-source platform that ships org chart, hard-stop budget and approvals in one package;
  • AI-Office gets the first agent live within two weeks and ports one to one to Paperclip once agents hand work to each other;
  • you are not tied to one model: Claude, Codex, Gemini or OpenClaw, whichever is best or cheapest at the time;
  • we run it ourselves every day and know the pitfalls from experience.
See the full landscape: Multica, Gas Town, Cofounder, Agent 365, Frontier and more →

// WHY IT FAILS

Four in ten projects won’t make it to 2028

Gartner expects over 40% of agentic AI projects to be cancelled before the end of 2027. The reasons are always the same, and they are all organisational, not technical.

Escalating cost

An agent that runs continuously and calls a large model at every step costs more after three months than the employee it was meant to relieve.

Our answer

Schedules instead of always-on, a budget cap per agent from day one, and the smallest model that can do the job.

Unclear value

A proof of concept that impresses in a demo, but nobody can say which number in the organisation it changes.

Our answer

We start with one measurable process and agree the outcome metric up front: lead time, error rate, share handled without human escalation.

Inadequate risk controls

The agent can email, order and change things, but nobody defined what it may not do, and afterwards nobody can reconstruct what it did.

Our answer

Approval points for anything that goes outside or costs money, a stop button, and a log per run. Built in, not bolted on.

Agent washing

Gartner’s term for chatbots and workflow tools sold as “agents”. They don’t deliver what was promised, and the disappointment taints the whole subject.

Our answer

We say honestly which rung of the ladder something belongs on. Often rung 2 is the right answer, and that’s fine.

// HOW WE HELP

Five steps to a working agent organisation

We don’t just build this for clients, we run it ourselves: our own AI-Office watches our sites, reviews our code and writes our briefs. We bring that experience with us. No slide deck, but an agent doing real work after six weeks.

  1. 01

    Scan

    1 to 2 weeks

    We walk through your processes and find the candidates: repetitive, well defined, measurable and with room for a review step. We work out what they cost now and what an agent would cost.

    Outcome

    Shortlist of 3 processes with business case and risk profile.

  2. 02

    Design

    1 week

    The org chart for the agents: roles, who they report to, what they may and may not do, which connections they get, what budget, and where a human must approve.

    Outcome

    Design document with role instructions, tool permissions and approval points.

  3. 03

    Pilot

    4 to 6 weeks

    One or two agents in AI-Office or Paperclip, on your server, on real data. At first a human reviews every outcome; we measure the agreed metric every week.

    Outcome

    Working agent in production, with numbers on cost and quality.

  4. 04

    Governance

    alongside the pilot

    An owner per agent, logging and retention, a stop button, and the EU AI Act: the transparency duty applies since 2 August 2026, high-risk obligations follow from December 2027. We sort it now, not then.

    Outcome

    Agent register, policy rules and an audit trail a regulator can read.

  5. 05

    Scale & operate

    ongoing

    More agents, handoffs between agents, cost monitoring and a monthly review of what they did. We host and manage, or hand over to your team.

    Outcome

    An agent organisation that grows under its own steam, with you at the wheel.

Frequently asked questions

Does this replace employees?

Usually not. The first thing to go is the work nobody wanted: nightly checks, chasing anomalies, first drafts. McKinsey finds in 2026 that most leaders expect AI to act mainly as support over the next two years. The roles that emerge, such as agent owner and outcome reviewer, are new work for existing people.

What does it cost?

The tooling is open source or ours, so that’s not where the cost sits. The cost is model usage (capped per agent with a monthly budget), a server, and our hours for scan, design and pilot. We budget those per step up front, so you can stop after the scan with no obligations.

Paperclip or AI-Office?

If you start with one to three agents that monitor or deliver something, AI-Office is simplest: fewer parts, up quickly, code in your own hands. If you want a real org chart with agents managing each other and multiple model families, Paperclip is the better base. We run both and choose per situation.

Where does my data live?

On your server, in the EU if you want. Both tools run self-hosted with their own Postgres database. What goes to the model you control per agent via tool permissions; connection keys are stored encrypted.

What about the EU AI Act?

Since 2 August 2026 an agent talking to people must make clear it is AI, and the AI Office in Brussels can enforce. The heavier high-risk obligations (risk management, logging, human oversight) were moved by the Digital Omnibus to December 2027 and August 2028. Most business agents don’t fall under them, but we set up logging, an owner and a stop button regardless: it’s simply good management.

How fast is something running?

A first scheduled agent, with one connection and an inbox where you read its findings, is up within two weeks. A pilot with measured results takes four to six weeks. An org chart with several agents handing work to each other is a matter of months, not years.

Curious which of your processes qualifies first? A thirty-minute call usually tells us.

Book an intro call →

Sources

  1. McKinsey & Company, 2025: The agentic organization: contours of the next paradigm for the AI era
  2. McKinsey & Company, 2026: The State of AI: Global Survey 2026
  3. Gartner, 25 juni 2025: Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027
  4. MIT Technology Review, 26 mei 2026: Rethinking organizational design in the age of agentic AI
  5. California Management Review, maart 2026: Governing the Agentic Enterprise: A New Operating Model for Autonomous AI at Scale
  6. GitHub, geraadpleegd 11 september 2026: paperclipai/paperclip — The open-source app everyone uses to manage agents at work
  7. Paperclip, geraadpleegd 11 september 2026: Paperclip Documentation
  8. AI Platform MKB, juni 2026: AI in het Nederlandse MKB: de stand van zaken in juni 2026
  9. EU Artificial Intelligence Act, bijgewerkt 31 augustus 2026: High-level summary of the AI Act (incl. Digital Omnibus-wijzigingen)
  10. VentureBeat, mei 2026: Anthropic says 80% of its new production code is now authored by Claude
↑ Back to top

← To the blog