The free tier now teaches orchestration. It still cannot run your workflow.

What is genuinely available for free now?

More than most vendors will admit, and the honest thing is to tell you the free tier has got broader. It is not just prompting any more. Microsoft ships free, open-source curricula that teach the hard parts: ai-agents-for-beginners is 18 lessons on building agents, covering multi-agent orchestration, tool use and evaluation, and mcp-for-beginners teaches the Model Context Protocol across six languages, from session setup to service orchestration, under an MIT licence. Add AWS, IBM and nasscom FutureSkills for foundational literacy, and the concepts an enterprise needs are largely teachable at zero cost.

We say this plainly because it is true, because a transformation owner can verify it in one search, and because a site whose whole brand is honesty cannot afford a claim that falls apart on inspection. Anyone still telling you "nobody free covers orchestration" is selling against a course you can open in a browser this afternoon.

What this does disqualify is paid concept-teaching. Prompt Engineering 101 as a paid cohort (and now, agent-orchestration-101 as a paid cohort) competes with free national-scale curricula and hopes you have not noticed. We refuse to sell either. The honest question is no longer "what does free not cover." It is "what can a course never do, however good," and the answer is specific.

So what is actually left to pay for?

Application, not concepts. A free curriculum can teach agent orchestration brilliantly in the abstract, and it still cannot orchestrate the exception queue your team runs on your systems, against your data, under your accountability. That gap between understanding a concept and applying it to one specific, messy, real workflow is the entire distance a pilot has to travel, and no population-scale course can travel it for you, because your workflow exists only inside your organisation.

Concretely, the paid-worthy layer is: one named workflow; a working system left behind that your team owns; an evaluation harness scored against your cases, not a tutorial's; retrieval tuned on your documents rather than a sample corpus; and governance built for your risk appetite and your regulators. A course covers MCP the standard; it cannot decide which of your systems an agent may touch, or what happens when it fails on your data.

This is a stronger position than the old one, not a weaker one. "Free does not cover the topic" was a claim about catalogue coverage, and catalogues changed. "Free cannot apply the topic to your workflow" is a claim about the nature of the work, and that cannot change.

Why does a course not change a workflow?

Because a workflow is not a person who has watched the right lessons. A workflow is a sequence of steps with owners, exceptions, handoffs and systems of record. Teaching an individual (even teaching them orchestration and MCP properly, for free) improves what that individual understands; it does not touch the process the organisation actually runs.

This is the same gap MIT located in the 95% pilot-failure finding: high individual capability, near-zero workflow change. You can have an entire team who have completed Microsoft's agent course and a process that runs exactly as it did before, because nothing connected the capability to the work. The knowledge lives in a learner's head; the work lives in a queue, a database and a set of exceptions nobody has written down.

Closing that gap is a build, not a course. It requires someone to sit inside your workflow, orchestrate against your real systems, ground the model in your documents, evaluate it on your cases, and govern its access to your data. The concepts for all of that are now free to learn. Doing it on your workflow is the work.

Only about 130 of thousands of self-described agentic AI vendors are genuine: the rest is 'agent washing'. >Gartner press release (via MarTech) (2025)

Two markups to watch for: charging for concepts the free curricula now teach, and 'agent washing' a chatbot as an agent. The defence against both is the same: ask the vendor to apply orchestration to your workflow, on your data, and show the evaluation harness they leave behind.

Why is governance the sharpest example?

Because it is irreducibly organisation-specific: the one thing no course can ever generalise, because it is defined by your data, your regulators and your risk appetite. A free curriculum can explain what an AI governance framework is; it cannot tell you which human is accountable for your deployed agent, what your DPDP consent provenance looks like, or where your escalation path runs.

The numbers show the gap is real and unclosed. IBM found 83% of Indian executives call effective governance key to successful AI, while only 4% have robust frameworks for managing AI-related risks (IBM Institute for Business Value, November 2025). No free course closes a 79-point gap that is specific to each organisation, and Gartner names inadequate risk controls among the causes of the over-40% agentic cancellations it expects by 2027 (Gartner, June 2025).

This is why our ladder leaves concepts to the free curricula and prices only application: a workflow sprint that ships an orchestrated, evaluated system on one of your workflows, and a governance and CoE retainer that maintains the controls no course could specify for you. An engagement scenario for RFQ response drafting shows the shape of an orchestrated build, as method rather than a client result.

What does this mean for a transformation owner?

Send your people to the free curricula, and send them to the good ones. Microsoft's agent and MCP courses, AWS, IBM and FutureSkills will teach the concepts, including orchestration and tool use, better and cheaper than any paid concept-cohort. Do not pay a rupee for teaching that a free national-scale curriculum already delivers.

Then spend transformation budget only on application: one named workflow, built on your systems, with a working system and an evaluation harness left behind, and governance specific to your risk. That is the line the free tier cannot cross, and it is the only line worth paying across.

Apply Gartner's base rate to whoever sells you the paid layer. Ask them to orchestrate against your workflow, show an evaluation harness scored on your cases, and describe how they would govern the agent's access to your data. A vendor who answers in course-syllabus terms is charging you for something Microsoft gives away; the roughly 130 genuine agentic vendors will answer in terms of your systems.

Disclosure: Chokmah is a new practice with no completed client engagements. We name the free curricula generously and by URL because the argument is about application, not coverage, and because being the one who tells you the free tier got broader is itself the honesty proof. Before you buy the paid layer from anyone, confirm it applies to your workflow and is not concept-teaching you can get free.

Learn the concepts free. Pay only to apply them to your workflow.

A workflow sprint applies orchestration, retrieval, evaluation and governance to one of your workflows, on your systems. Your team owns the code.

Scope a Workflow Sprint · Talk about a Capability Cohort

Frequently asked questions

Does free AI training cover agent orchestration and MCP?

Yes, increasingly. Microsoft ships free, open-source curricula (ai-agents-for-beginners (18 lessons on orchestration, tool use and evaluation) and mcp-for-beginners (MCP across six languages)) alongside AWS, IBM and nasscom FutureSkills. The concepts are largely free to learn. What no course covers is applying them to your specific workflow, data and governance.

What should enterprises pay for in AI, if the concepts are free?

Application, not concepts. Pay for one named workflow built on your own systems, with a working system left behind, an evaluation harness scored against your cases, retrieval tuned on your documents, and governance specific to your risk and regulators. A course teaches orchestration in the abstract; it cannot orchestrate the exception queue your team actually runs.

Is paid prompt-engineering training still a mistake?

Yes, and so is paid agent-orchestration-101. Both compete with free national-scale curricula that teach the same concepts, so paying a vendor to teach them is paying for something you can open in a browser. Pay for a vendor to apply those concepts to your workflow, not to lecture on them.

How long does it take to train an engineer on agent orchestration?

The concepts can be learned quickly from free curricula. Becoming productive is a build-alongside problem: engineers learn to orchestrate your systems by shipping a real orchestrated workflow with someone who has done it, then maintaining it. That is why our sprints and cohorts build on the client's own workflows rather than teaching orchestration in the abstract.

Does AI training reduce pilot failure?

Concept training on its own does not, because pilot failure is a workflow and governance problem, not a knowledge one. A team can complete Microsoft's agent course and run the workflow exactly as before. What reduces failure is applying the concepts to a named workflow with a baseline and an evaluation harness: the build, not the course.