Quick answer
To choose an AI agent platform without getting locked in, judge it on how well it connects, not just on how clever its agents are. Before you sign, check that it works with agents and models from other vendors, that you can swap one part without rebuilding the rest, and that your instructions, workflows and data can be exported in a usable form. Run a short trial on one real task, score every option against the same list, and get the exit terms in writing. This follows the advice Nate B Jones gives in his video on the agentic control plane: make interoperability a requirement, not a nice extra.
Key takeaways
- Lock-in is rarely about raw data. It is about instructions, workflows and connections that only work in one place.
- Decide your must-haves first: the job, the systems involved and who approves the agent’s work.
- Ask every vendor the same questions about other agents, model choice, exports and exit terms.
- Score options in a simple table so the slickest demo doesn’t win by default.
- Test with one real task and try removing or swapping a part during the trial.
- Building your own is possible, but only if you can resource and maintain it.
📘 Main article: This guide supports Agentic Control Plane: Why Microsoft, Salesforce and OpenAI Want to Run Your AI Agents. It turns Nate B Jones’s buying checklist into questions, a scoring table and red flags you can use straight away.
In this guide
- What does lock-in look like with an AI agent platform?
- What should you decide before you talk to vendors?
- Which questions should you ask every vendor?
- How do you compare platforms fairly?
- What are the red flags for lock-in?
- Should you build your own agent platform instead?
- Frequently asked questions
- The bottom line
- Related posts
More software now comes with AI agents built in, and most vendors want to be the place where all your agents run. That makes the choice of platform more important than it first looks.
This guide answers one question: how do you choose an AI agent platform without getting locked in? It builds on our main article, Agentic Control Plane: Why Microsoft, Salesforce and OpenAI Want to Run Your AI Agents, and on Nate B Jones’s video about the trillion-dollar fight to run your AI agents.
What does lock-in look like with an AI agent platform?
Lock-in is when leaving a platform would cost so much work that you stay even when something better appears. With AI agents, it usually looks like this:
- your agents can only use that vendor’s models,
- your workflows and skills are stored in a format nothing else can read,
- agents from other tools can’t take part in the work,
- you could export your records, but not the instructions and corrections that make the agent useful.
Nate’s point is that customers who demand interoperability before they sign get to set the terms. Once the contract is signed, you have far less leverage.
What should you decide before you talk to vendors?
Write down three things before any demo: the job, the systems and the approvals.
- Pick one job. For example, “prepare a weekly sales summary from our CRM and email”. Our guide on choosing the next workflow to automate with AI helps you pick well.
- List the systems involved. Which apps hold the information the agent needs? Where do your staff already spend their day?
- Name the approvals. Which actions must a person check before they happen, such as sending, paying or deleting?
- Set your must-haves. Usually: works with your main apps, lets you choose or change models, and exports your setup.
💡 Tip: Keep this to one page. A short, clear brief makes vendor answers easy to compare, and it stops a long feature list distracting you from the job you actually need done.
Which questions should you ask every vendor?
Ask the same questions of every vendor, and ask for answers in writing.
| Question | Why it matters |
|---|---|
| Can agents built elsewhere take part in workflows here? | Shows whether the platform plays well with others, like the MuleSoft bridge for Copilot Studio agents |
| Can we choose or change the AI model? | Lets you move to a better or cheaper model later |
| How do we export instructions, skills and workflows? | Your setup is often worth more than your data |
| Which data can it read in place, without migration? | Avoids a costly move before you know it works |
| How are permissions and approvals set? | Keeps people in charge of risky actions |
| What happens to our data and setup if we leave? | Tells you the real exit cost |
| How is it priced as usage grows? | Avoids surprises; see our piece on per-seat vs per-token pricing |
If an answer is vague, ask for a demo of that exact point. “Yes, it integrates” means little until you see it work with one of your own tools.
How do you compare platforms fairly?
Use a simple score from 0 to 2 for each must-have, then trial the top two on one real task. 0 means no, 1 means partly or with extra work, 2 means yes and shown in a demo.
| Must-have | Platform A | Platform B |
|---|---|---|
| Works with our main apps | 2 | 1 |
| Accepts agents from other vendors | 1 | 2 |
| Lets us change models | 2 | 2 |
| Exports instructions and workflows | 0 | 2 |
| Clear approvals for risky actions | 2 | 1 |
| Total | 7 | 8 |
These numbers are only an example layout. During the trial, try one deliberate swap: change the model, or remove one connected tool. How much breaks tells you more about lock-in than any sales deck. Our guide on choosing an AI tool for your business covers testing on real work in more depth.
Want the bigger picture? Read Agentic Control Plane: Why Microsoft, Salesforce and OpenAI Want to Run Your AI Agents to see why vendors are fighting over this layer, then use our guide on keeping your AI agent setup when you switch tools to protect the work you put in.
What are the red flags for lock-in?
Watch for these warning signs during sales calls and trials:
- No export for instructions or workflows, only for raw data.
- One model only, with no plan to support others.
- “Integrations” that need the vendor’s own consultants for every change.
- Long minimum terms with no clear exit process.
- Pricing that jumps once agents run regularly.
- Features promised for later that your plan depends on today.
⚠️ Warning: Don’t build your plans around announced or preview features. In Nate’s video, Oracle’s ChatGPT workspace is still in development and Salesforce’s Agent Context Engine is in preview. Buy for what works now, and treat the rest as a bonus.
Should you build your own agent platform instead?
Only if you can pay for the build and keep people on it. Nate warns that building your own control plane is not a small job. You need the skills to build it, the budget to keep improving it and someone responsible for it over time.
For most small businesses, a sensible middle path is to buy a platform that scores well on interoperability, and keep your own written copy of instructions and workflows outside it. That gives you most of the flexibility with far less work.
🔎 Unconfirmed / rumour: How open each platform will be in a year’s time is unknown. Vendors can change export options, pricing and partner access. Recheck your top answers at each renewal rather than assuming today’s terms will last.
Frequently asked questions
What is AI vendor lock-in?
It is when moving away from an AI platform would take so much work, or cost so much, that you feel stuck with it. With agents, the hardest part to move is usually your setup, not your data.
What does interoperability mean for AI agents?
It means agents, models and tools from different vendors can work together, share information and hand work to each other without rebuilding.
How long should an AI agent platform trial be?
Long enough to run one real task several times, typically a few weeks. Include at least one deliberate swap, such as changing the model.
Is it safer to pick the platform we already use?
Not automatically. Nate B Jones warns against choosing by default. It may be the right choice, but score it against the same list as everything else.
Can a small business ignore all this?
You can keep it light, but the core idea still applies: keep your instructions and workflows somewhere you own so changing tools later is easy.
The bottom line
Choosing an AI agent platform is really choosing how easily you can change your mind later. Write a one-page brief, ask every vendor the same questions in writing, score them fairly and test one deliberate swap during the trial. Your next step: write down the one job you want an agent to do and the three apps it would need to reach.
Related posts
- Agentic Control Plane: Why Microsoft, Salesforce and OpenAI Want to Run Your AI Agents
- How to Keep Your AI Agent Setup When You Switch Tools
- Per Seat vs Per Token AI Pricing: Why It’s the Wrong Question (and What to Ask Instead)
