Audit your current operating layer.
Inventory every AI surface, locate where decision logic lives, rank candidate self-improvement loops.
Pilots never reach production.
Demos look incredible. Almost none of them end up running the business.
Spark / OS/ Client intro/ v1.0/ 2026
Company
Intelligence
for the AI era.
Owned company operating systems for businesses that want to compound, not just adopt.
RING 03 / HUMANSRING 02 / AGENTSRING 01 / WORKFLOWSCORE / INTELLIGENCE
This deck is for one kind of buyer.
Best-fit · operator
Building company intelligence.
You see AI as structural, not tactical. You want to own the operating layer, not rent another tool.
Compounding leverageGovernance + evalsFewer, deeper tools
Wrong-fit · feature-buyer
Shopping for an AI workflow.
You want a chatbot, a dashboard, a single integration. Real needs, but other vendors solve them better.
"Connect our tools""A copilot for us""AI without changing anything"
AI is not a feature. It is a new company structure.
Last 3 years · productivity
How do we make employees faster?
Faster writing, faster coding, faster summarising. Real gains. Structure unchanged.
Next 5 · company intelligence
How does the company itself become smarter?
Work moves through memory, rules, workflows, evals. The winners reorganise around AI, they don't just use it.
What we hear in every intake
Three years in. AI still does not compound.
Companies have more AI tools than ever, and most of the value evaporates. The root cause is the same in every diagnosis: the intelligence layer is not owned.
D / 01
POC TO PROD< 8% conversion observed
D / 02
Business logic lives inside vendor prompts.
Your lead-qualification rules sit inside someone else's prompt template. You do not see it, version it, or leave with it.
OWNED IP0% of decision logic
D / 03
Context resets every Monday.
The model knows your business as well on day 400 as on day 1. Nothing is persisted in a place you control.
SYSTEM MEMORY0 days retained
D / 04
Agents stay read-only because nobody trusts them.
Without policy, approvals, and evals, no operator gives an agent real access. The agent stays a slightly better search box.
ACTION LAYERRead-only by default
For thirty years, tools were at the center. Now intelligence is.
The current model · humans around the edge
SaaS tools at the center.
Business logic scattered across vendors. Humans copy state between them. Every tool is an island with its own schema and politics.
Tools hold logicHumans act as couriers
The Spark model · intelligence at the center
Company intelligence at the core.
Memory, goals, rules, and schemas live in one owned layer. Agents and workflows orbit it. Tools become instruments.
Intelligence is coreAgents and workflows orbitTools are instruments
06 / The Differentiator / Company as codebase
Your company should be
readable, versioned,
and executable.
01 / Explicit
Rules written down. Workflows mapped. Schemas defined. Decisions logged. Nothing load-bearing left as oral tradition.
02 / Inspectable
Anyone on your team can read how the company decides, version a change, and review what shipped. Reviewers are a feature, not a bottleneck.
03 / Executable
Agents operate from versioned instructions. Improvements are traceable. If the company is not written down, agents cannot improve it.
A company that improves itself, week over week.
Sense
Capture every important business signal as Markdown in one store.
Decide
Policy and operating rules determine what the system may do.
Act
Governed agents execute through thin tool connectors.
Check
Evals and approvals catch drift and gate high-risk actions.
Learn
Failures and feedback update memory, retrieval, and instructions.
Improve
The operating system updates itself. Next week's decisions are better.
Examples in production
- A failed knowledge query updates docs and retrieval and the agent's instructions.
- A repeated support issue becomes a documentation or product update.
- Sales-funnel friction triggers an experiment, scored against last quarter's funnel.
- New expert advice updates the operating manual and every agent that depends on it.
What Spark builds into your company.
P / 01
Signal Store
Every email, message, meeting, ticket, note, normalised to Markdown in your database.
P / 02
Memory
A durable, structured record of what the company knows, decides, and remembers.
P / 03
Workflows
Explicit, executable representations of how work moves through your business.
P / 04
Agents
System prompts, tool grants, and policy boundaries for each agent.
P / 05
Approvals
Human-in-the-loop gates with scope, signers, and timeouts.
P / 06
Schemas
Typed structures for customers, contracts, decisions, and other business objects.
P / 07
Connectors
Thin, owned adapters into the SaaS you keep. Tools become instruments.
P / 08
Evals
Production telemetry for every agent, workflow, and decision.
Source systems are sensors. The Signal Store is memory.
N.B./Gmail is not memory. Slack is not memory. HubSpot is not memory. They are sensors. Memory is the layer where agents reason and act.
F / 01Raw source + normalised Markdown
F / 02Extracted entities (people / companies / deals)
F / 03Facts + action items + summary
F / 04Linked artifacts (contracts / decks / tickets)
F / 05Action history (what agents and humans did)
10 / Doctrine / The three Spark stances
*Less* software.
Yours at handover.
Humans up the stack.
01 / Use fewer tools, make them more agentic
Every extra SaaS layer adds permission models, hidden workflows, and places where context drifts. We keep the tools that pull weight. The rest moves into the operating layer.
02 / You own what we build
SparkOS is not a seat. It is your memory, schemas, workflows, rules, connectors, and instructions, in an environment you control. Do not rent your company's intelligence.
03 / Humans move up the stack
Agents handle routing, drafting, reconciliation, monitoring. Humans handle direction, judgment, relationships, taste, accountability. The goal is replacing coordination drag, not people.
Map. Instrument. Operationalize. Hand off.
W 01 - 02Phase 01 / Map
P / 01
You leave with
— Spark Map
— Ranked intervention targets
— First-loop shortlist
W 03 - 06Phase 02 / Instrument
P / 02
Install the Signal Store + core primitives.
Signal Store, memory, schemas, policies, and approvals stood up in an environment you own. Tools connected as governed instruments.
You leave with
— Owned intelligence layer
— Logic your team can read & edit
— First evals running
W 07 - 10Phase 03 / Operationalize
P / 03
Move workflows to autopilot.
Two or three load-bearing workflows migrate to L3. Evals in production. First self-improvement loops measurably closing.
You leave with
— 2-3 workflows at L3
— Approvals as exceptions
— Improvement loops live
W 11 - 12Phase 04 / Hand off
P / 04
Operate without us.
Runbooks, eval dashboards, extension patterns, and the 90-day autonomy roadmap transfer to your team. Spark stays optional.
You leave with
— Full ownership
— 90-day L3 → L4 plan
— We're optional
12 / The ask
Map your company intelligence layer.
Start with the €5,000 Foundation Track: a fixed-scope, 1–2 week diagnostic that ends with your Spark Map, ranked intervention targets, and a clear next step. With us, or without us.
Start the Foundation Track →/hello@sparkcollective.be
SPARK · COMPANY OPERATING SYSTEMS · For the AI era