Back to Blog

LangChain vs Semantic Kernel - Which Framework for Your AI App

May 24, 2026•10 min read•Michael Ridland

If you are picking between LangChain and Semantic Kernel right now, you are probably staring at two open browser tabs, two GitHub repos, and a vague feeling that "they look similar enough." They are not similar enough. The wrong choice will cost you four to six months of rework once your prototype hits production load, hits enterprise security review, or hits your real customers.

We have built and shipped systems on both frameworks across financial services, mortgage broking, healthcare and manufacturing clients. We have also been called in to migrate projects from one to the other after the original team made the wrong pick. This post is what I wish someone had handed those teams before they wrote line one of code.

The Short Answer

If your engineering team lives in Python, your data scientists are involved, and your AI app is the product itself - pick LangChain.

If your engineering team lives in .NET, your stack is Azure, and the AI app is being added into an existing enterprise system - pick Semantic Kernel.

That is the answer for about 80% of buyers. The rest of this post is for the other 20%, where the choice is genuinely close and the details matter.

Quick Comparison Table

Dimension LangChain Semantic Kernel
Primary language Python (TypeScript second-class) C# (Python and Java supported)
Stewardship LangChain Inc. (VC-backed startup) Microsoft (part of Azure ecosystem)
Best for AI-first products, data teams, RAG-heavy apps Enterprise apps, .NET shops, regulated industries
Azure integration Good via connectors Native, first-party
Observability LangSmith (paid, very good) Application Insights, Azure Monitor
Agent abstraction LangGraph (graph-based, explicit state) Agent Framework (plugin-based, declarative)
Hiring pool in Australia Larger, mostly Python Smaller, mostly C# senior devs
Typical project cost (12-week pilot) $90k-$180k AUD $120k-$220k AUD
Time to first working prototype 1-2 weeks 2-4 weeks
Long-term maintenance cost Moderate (lots of breaking changes) Lower (Microsoft cadence is calmer)

Where LangChain Wins

LangChain is the framework that gets out of your way fastest. If you give a competent Python developer a weekend, they will have a working RAG prototype connected to your data, with citations, in a notebook, by Sunday night. That speed is real and it matters.

The community is also a genuine asset. When you Google an error message at 11pm on a Tuesday, you will find an answer. When you need a connector for some obscure vector database your client signed up for two years ago, someone has already written it. The LangChain ecosystem is wide and active in a way Semantic Kernel is not, at least not yet.

LangSmith is the other quiet advantage. Once you start running real agents in production, you need to see what is happening inside them. LangSmith traces every call, shows you token usage, lets you replay failed runs, and gives you proper evaluation tooling. We pay for it on most LangChain projects and have never regretted it. The pricing starts around $39 USD per seat per month and scales with traces, so budget around $300-$800 AUD per month for a small team.

When LangChain is the right call

  • The AI app is the product. You are building a new SaaS, a customer-facing assistant, or an internal tool from scratch.
  • Your team is already strong in Python or Node.
  • You are doing heavy retrieval-augmented generation across messy data.
  • You need fast iteration. You expect to throw away v1 and v2.
  • You want best-of-breed components and you are willing to glue them together.
  • You are using non-Microsoft models heavily (Anthropic, open source via Bedrock, etc.).

Where LangChain hurts

The pace of breaking changes is real. Pin your versions, write tests around the framework boundary, and budget time every quarter for upgrades. We have had clients on LangChain 0.0.x who could not upgrade to 0.1 without rewriting a third of their code. The 0.2 to 0.3 jump was kinder. Still, treat LangChain like an early-stage startup product, because that is what it is.

The abstractions can be too clever. Chains, runnables, LCEL, agents, tools, expression language - it is a lot. Junior developers get lost. We usually strip the cleverness back and write more direct code that is easier to debug six months later.

Where Semantic Kernel Wins

Semantic Kernel is what you choose when the boring stuff matters more than the shiny stuff. Authentication via Entra ID, logging into Application Insights, secrets in Key Vault, deployment via Bicep or Terraform, code reviewed by enterprise security teams who have opinions - all of this is easier when your AI framework is from Microsoft.

The C# story is the real differentiator. If your client has fifteen years of .NET code, a team of senior C# engineers, and an existing API surface they want to add AI features into, Semantic Kernel slots in cleanly. You do not need to spin up a separate Python service, you do not need to argue with the security team about a new runtime, and you do not need to hire Python developers you cannot find.

The Python version of Semantic Kernel exists and works, but be honest with yourself - if you are going to write Python, just use LangChain. The reason to pick Semantic Kernel is the C# version.

When Semantic Kernel is the right call

  • You are adding AI into an existing .NET application.
  • Your client or employer is deeply on Azure already.
  • You have enterprise security and compliance constraints (banking, insurance, government).
  • Stability and slower change cadence matters more than newest features.
  • Your team is C# strong but Python weak.
  • You are using Azure OpenAI as your primary model provider.

Where Semantic Kernel hurts

The plugin ecosystem is thinner. If you need a specific vector database or a niche SaaS connector, you will probably write it yourself. Budget for this.

Documentation has gotten better, but the conceptual gap between "I read the docs" and "I built something useful" is still wider than it should be. Plan to spend the first two weeks of any Semantic Kernel project genuinely learning the framework, not just shipping features.

Agent support has caught up over the last year with the Agent Framework, but the patterns are different from LangGraph. If your team has LangChain muscle memory, expect a re-learning curve.

The Hiring Question Nobody Talks About

You can build an app on either framework. You then need to keep it running, fix bugs at 3am, and add features for the next three years. The single biggest predictor of whether that goes well is whether you can hire and retain people who know your stack.

In Sydney, Melbourne and Brisbane, the Python AI developer market is hot but available. You can hire a mid-level LangChain developer for around $130k-$170k AUD plus super. Senior developers with real production AI experience are $180k-$240k.

The C# AI developer market is harder. Most senior .NET developers in Australia have done some AI work but Semantic Kernel specifically is still niche. You will pay $160k-$220k for someone who has actually shipped a Semantic Kernel system, and you will wait longer to find them. The upside - turnover is lower because the role is rarer.

If you are a small business that does not want to build an in-house team, this is one of the reasons to work with our AI agent developers or our Microsoft AI consultants. The cost of hiring and onboarding the wrong person dwarfs the cost of bringing in a team that has shipped this work before.

What About LangGraph vs Microsoft Agent Framework

Both frameworks have a serious agent story now. LangGraph treats your agent as a graph with explicit state. You define nodes (the steps), edges (the transitions), and a state object that flows through. It is verbose, but you can see exactly what is happening and you can debug it.

Microsoft's Agent Framework, layered on top of Semantic Kernel, takes a more declarative approach. You describe the agent and its capabilities, and the framework handles a lot of the orchestration. It is faster to get going, but harder to debug when behaviour goes sideways.

In production, we have seen LangGraph win for complex multi-step workflows where you need to inspect and adjust the path. Microsoft's Agent Framework wins for simpler conversational agents where Azure-native deployment matters.

If agents are central to your build, read our notes on enterprise AI agents and how we approach the architecture.

Cost Comparison - Real Project Numbers

These are rough numbers based on actual projects we have delivered or quoted in 2025-2026.

Small RAG app, internal use, 6-week build

  • LangChain: $60k-$100k AUD
  • Semantic Kernel: $80k-$130k AUD

Customer-facing assistant, integrations to 3 enterprise systems, 12-week build

  • LangChain: $120k-$200k AUD
  • Semantic Kernel: $150k-$240k AUD

Multi-agent system, regulated industry, 6-month build

  • LangChain: $300k-$500k AUD
  • Semantic Kernel: $350k-$600k AUD

The Semantic Kernel premium comes from two things - the smaller talent pool and the typical client (more enterprise, more compliance, more meetings). It is not the framework itself.

Ongoing run costs are also worth thinking about. Azure OpenAI and OpenAI direct are within 10-15% of each other on price. The real cost driver is your prompt design and your retrieval architecture, not your framework choice.

Common Objections and Misconceptions

"LangChain is dying because every model provider has its own SDK now."

It is not dying. It is becoming less monopolistic, which is healthy. We still pick LangChain on most greenfield Python projects because LangGraph, LangSmith and the connector ecosystem still beat rolling your own.

"Semantic Kernel is just a Microsoft attempt to lock us into Azure."

It works fine with OpenAI direct, Anthropic via API, and open source models. The Azure integration is best-in-class, but you are not locked in any more than using any other Microsoft library locks you in.

"We should use both and pick the best tool for each job."

In theory, yes. In practice, the operational cost of running two frameworks (two sets of dependencies, two deployment pipelines, two skill sets to hire for) is high. Pick one for the next 18 months and revisit.

"Can we start with one and migrate later?"

You can. We have done it. It costs roughly 40-60% of the original build to migrate. Plan to pick correctly the first time.

A Decision Framework You Can Actually Use

Walk through these questions in order. Stop when you get a clear answer.

  1. Is your engineering team primarily .NET/C#? Yes - Semantic Kernel. No, continue.
  2. Is your AI app being added into an existing Azure-hosted enterprise system? Yes - Semantic Kernel. No, continue.
  3. Do you have hard compliance requirements (APRA, ISM, ISO 27001) and a security team that already trusts the Microsoft stack? Yes - Semantic Kernel. No, continue.
  4. Is your team primarily Python or willing to be? Yes - LangChain. No, continue.
  5. Do you need a wide ecosystem of connectors and rapid iteration? Yes - LangChain. No, continue.
  6. If you got here, the choice is genuinely close. Go with whatever your senior engineer is more excited about. Their motivation will matter more than the framework.

When to Bring in Help

If this is your team's first production AI build, the framework choice is honestly not the hardest part. The hardest parts are data quality, prompt design, evaluation, deployment, and the operational maturity to know when your model is failing in production.

We do this work for Australian businesses every week. If you would like a second opinion on your stack before you commit, the fastest way to get one is to book a session with our AI strategy consultants or talk to our Microsoft AI consultants about a paid technical review.

You can also contact us directly if you want to chat about whether your project is better suited to LangChain, Semantic Kernel, or one of the newer frameworks like CrewAI or Microsoft's Agent Framework. We will give you a straight answer, including telling you when you do not need us.

The best AI app is the one that ships, stays in production, and gets used. The framework is a means to that end. Pick the one your team can actually run.