Skip to main content
News
Platform11 min read

From Cocoa to Chocolate: Use Any AI, Stay in Control

Intelligence is abundant. Control is not. Chocolate Factory is the control plane for all your AI.

From Cocoa to Chocolate: Use Any AI, Stay in Control

Chocolate Factory is the control plane for all your AI. It is where a business builds and orchestrates its AI agents, and where two gateways govern the models and tools those agents use. Every model sits behind one door, every tool behind one set of rules, and every run leaves a full record of what happened. It runs wherever the business’s data has to stay.

Intelligence is abundant. Control is not.

The leap in AI over the last few years has been staggering. Frontier models that once belonged to a handful of labs are now a credit card away, and open-weight models run on hardware a business can own. Raw intelligence is no longer scarce, and it no longer sets one business apart from another.

What is scarce is the ability to put that intelligence to work safely. A generalist model knows everything about the internet and nothing about the business it is meant to serve: its policies, its systems, its customers, or the lines it must never cross. On its own it cannot be trusted with real work, and it cannot be held to account for what it does.

Think of it like cocoa. Cocoa is valuable, and many suppliers grow it. But nobody walks into a shop and asks for cocoa mass. They ask for chocolate: made to a recipe, from approved ingredients, on a line someone is accountable for, inspected, and recorded batch by batch. Between the cocoa and the chocolate sits a factory, and the factory is where the value is made.

Model output is the cocoa. Every provider sells it, and the price keeps falling. What a business needs is the chocolate: work done correctly, by the rules, with a record it can stand behind. Chocolate Factory is the factory in between.

AI is already in every team. Nothing is in charge of it.

Most businesses did not decide to adopt AI. It arrived, team by team. One group signed up for Claude agents, another for OpenAI’s, a third built its own, and a growing number of the company’s apps now call a model somewhere in the background.

Each of those decisions made sense on its own. Together they leave AI that nobody is in charge of. Provider keys sit in every app. No one can cap what a team, an app or an agent spends until the invoice arrives. Any agent can reach any tool it has been given a connection to, often with more access than its job needs. When something goes wrong there is no trace to follow, and an agent caught in a loop keeps calling, and keeps billing, until someone notices.

The result is a business that cannot answer three basic questions: which models are in use, what they cost, and what each agent can touch. So AI stays in pilots.

Around 95% of enterprise AI pilots never reach production (MIT NANDA). The model was rarely the reason. Control was.

Building AI is easy. Operating it is hard.

Today a team can connect a model, write a prompt, add some company knowledge and have an agent answering questions. The demonstration looks impressive. Then someone asks when it can go into production, and the easy part ends.

80%
20%
The first 80% proves an agent can work
The last 20% decides whether the business can trust it

That last fifth is the real world: ambiguous requests, incomplete or conflicting information, rules that change, integrations that fail, and users who do things nobody tested for. A fix for one case breaks another, and teams swap prompts, models and tools without knowing which part caused the problem. Behind every response, an agent works through a chain of steps.

  1. Understand the request
  2. Retrieve knowledge
  3. Select tools
  4. Call systems
  5. Follow a process
  6. Apply business rules
  7. Check its output
  8. Respond

Each step is a place where things can go wrong, and the final answer can look convincing even when the work behind it was not. What separates a dependable system from an impressive demo is whether anyone can see how the answer was produced. And AI is never finished: models, prompts, knowledge and rules keep changing. It needs an operational layer, a place where AI is run, not just built.

A factory, not a pile of projects

Most businesses start one project at a time, each with its own prompts, models, tools and way of keeping things safe. Governance is reinvented every time, and an improvement in one project never reaches the next. A factory works differently. When something goes wrong, the engineers do not rebuild the factory; they find the stage responsible and improve it.

Assembly

Compose each agent from its model, instructions, knowledge and tools.

Inspection

Observe every step of every run, not just the final answer.

Testing

Try real questions against an agent and compare results as things change.

Quality assurance

Human approval on the steps that carry risk, with guardrails to follow.

Continuous improvement

Refine the one part that needs attention, not the whole system.

Chocolate Factory is built as that factory: one shared foundation where every agent is configured, governed and observed the same way.

People and appsWhere agents are used

  • Your own apps
  • Microsoft Teams
  • WhatsApp
  • Schedules

Agent OrchestratorAgents · workflows · memory

AI GatewayModels · routing · metering

MCP GatewayTools · access · budgets

Model providers

  • Commercial APIs
  • Open-weight hosts
  • AI factories

Enterprise systems

  • Salesforce
  • SAP
  • Partner services
  • First-party apps
TracingEvery hop, end to end

Runs wherever your data has to stay.

  • Your own VPC
  • On-prem
  • Air-gapped
  • Hybrid
  • Any public cloud
The Chocolate Factory stack: where agents, models and tools meet.

People reach agents in their own apps, in Microsoft Teams, on WhatsApp, or on a schedule with no one asking. Those agents are built and run in the Agent Orchestrator. The AI Gateway governs the models they think with, and the MCP Gateway governs the systems they act on. Tracing runs through every hop, end to end.

The gateways are not only for agents built in Chocolate Factory. The AI Gateway serves apps and agents from any vendor: existing Claude, OpenAI and in-house agents, and any app already calling a model, can route through it and come under the same keys, budgets and metering. The Orchestrator builds and runs its own agents, and both gateways wrap them as its harness.

A gateway, in plain terms, is the single entry point between apps and agents and the services they use. It authenticates, authorises, routes and meters each request, so the business can change providers without changing its apps. Chocolate Factory has two: one for models and one for tools.

1. AI Gateway

Every factory has a single intake door, where ingredients from many suppliers are checked, weighed, priced and booked to the line that will use them. The AI Gateway is that door for models. Commercial and open-weight models sit behind one endpoint, so switching models means changing a name, not rewriting an integration. Provider keys stay in the gateway, teams, apps and members get virtual keys, and budgets, rate limits and allowed models are set per team, project or app key.

A routing rule sends each request to the lowest-cost endpoint, a preferred host or one fixed endpoint, and falls back if one is down. Every request is metered in tokens and cost, by model, team and consumer. Spend caps with a circuit breaker, guardrails, a request log, load balancing, routing by difficulty, caching and keys that expire are part of it too.

2. MCP Gateway

A model that can only talk is an advisor. A model that can act is an employee, and an employee needs to know which machines it may operate. MCP servers are the connectors agents use to act on real systems. The MCP Gateway keeps them in one curated registry and checks every call first: is the caller allowed this action, and do the connector’s budget and call rate allow it?

Access is set per project, down to single actions, so reading can be allowed while writing is denied. Circuit breakers cut a connector off when usage looks anomalous or reaches for sensitive data. Tool filtering, tool groups, an audit log and cost per tool call are part of it too.

3. Agent Orchestrator

The future of enterprise AI is not one generic chatbot. It is a managed workforce of specialists, each tuned to a distinct job and given only the tools it needs. The Agent Orchestrator is where that workforce is built, organised and governed. Each agent is assembled from instructions, a model, knowledge, tools, MCP connections, integrations, memory, a schedule and the other agents it works with.

Run limits stop an agent that would otherwise loop. Each agent has its own key, so nothing it does is anonymous, and every run can be read in order. The harness routes the Orchestrator’s agents through both gateways.

4. Business Process Studio

Some work should be left to an agent’s judgement. Other work must happen the same way every time. For that a factory uses a fixed recipe, and Chocolate Factory uses the Business Process Studio: a process drawn as a flowchart of triggers, agents, tools and decisions, built by hand or drafted by AI from a description.

Steps that carry risk wait for a named approver. Every save becomes a version that runs only when someone decides it should. An agent can call a workflow as a tool, and every run is recorded step by step.

5. Run API

Agents are only useful where people work, and much of that work happens inside the business’s own software. The Run API lets any app run any agent from one address, with the agent’s ID and a key, and every run it starts is traced like any other. The answer arrives live, word by word, or all at once, and the business chooses where conversation history lives.

6. Observability

A factory that cannot see inside its own line can only guess why a batch went wrong. In Chocolate Factory, monitoring is always on. Every conversation is logged, request volume shows which agents are busy, and any run opens into its trace: every model call, every tool called, what each returned, how long each step took, and what the run cost. Only the organisation’s owners and admins can see it.

The Dashboard chat is the test bench, where an agent meets the questions it will face before it meets real users. Root-cause analysis shows why a bad answer went wrong.

7. Projects, roles and where it runs

A factory is laid out in rooms, each with its own rules about who may enter. Chocolate Factory is laid out the same way: an organisation, projects for each team or line of work, and sub-projects up to three levels deep that inherit their parent’s access. Each project has its own access, limits and allowed models, and its models, MCP servers, agents, API keys and data stay inside it.

It runs where the business’s data has to stay: in its own virtual private cloud, on-premises, air-gapped, hybrid, or in any public cloud. Because the AI Gateway can route to any host, a business can prefer models served in-country. Sovereignty becomes a routing rule.

Observe, correct, improve

A deployed agent is the start of the work, not the end of it. The businesses that get real returns from AI treat every run as evidence: they watch what the agent did, correct the part that led it astray, and let the fix apply from the next run onward.

Observe
Correct
Improve
…repeat

When an answer is wrong, the trace shows where: the wrong source, the right tool with the wrong values, a skipped rule, something out of date. The team changes only that part and tries the same question again on the test bench. Nothing has to be rebuilt. At the edges, the loop runs on its own: run limits end loops, circuit breakers cut off connectors, and budgets keep one mistake from becoming a large bill. Problems are contained first and diagnosed second.

It is how a head chocolatier runs a line: read the batch record, find the step that drifted, adjust the recipe card, and run the next batch.

Now multiply it across every team

One agent, observed and corrected, is a project. The point of a factory is that the same discipline holds on every line. Because every agent runs on the same platform, each lesson learned in one place is available to every team, and each improvement compounds across the business instead of staying inside one project.

The same is true of control. Each team works in its own project, with its own access, budgets and allowed models, from day one. Governance is not reinvented for the next project; it is already there. That is what lets a business say yes to AI team by team, without giving every team the keys to everything.

What changes for each person

Business leaders can take AI out of the pilot stage, because the three questions that kept it there now have answers.

The CIO and IT get one place to say yes: provider keys in the gateway, every model in one catalog, every connector in one registry.

Finance sees AI spend as it happens, not when the invoice arrives, and can bill it back to the team that used it.

Business teams write the instructions, choose the knowledge, decide which steps need a person, and correct the agent when it goes wrong.

Developers stop rebuilding plumbing: one endpoint for every model, one registry for every tool, and the Run API for their own apps.

Risk, security and compliance get control they can point to: access scoped to single actions, circuit breakers, guardrails, named approvers, audit logs and a full trace of every run.

Model providers and AI factories gain a place in every customer’s catalog, at their own prices.

In chocolate terms: the buyer knows every supplier and every price, the accountant sees the cost of every batch, the chocolatiers write their own recipes, and the food-safety inspector decides which machines each worker may run.

In short

Access to powerful models is no longer what sets a business apart. Every competitor can buy the same intelligence, at the same falling price. What sets a business apart is the ability to operate that intelligence reliably: to move from demonstrations to production, from black-box answers to observable work, from trial and error to targeted improvement, and from scattered projects to one platform.

The model underneath will change, many times. What keeps its value is everything around it: the business’s context, its rules, its approvals, and its record of what was done. That is the harness, and it belongs to the business.

Models make the cocoa. Chocolate Factory makes the chocolate.

Use any AI. Stay in control.

Chocolate Factory, the control plane for all your AI. See the AI Gateway, MCP Gateway and Agent Orchestrator, or book a demo.