If you build AI features for an AIOps product, a customer will ask "can we use our own LLM?" sooner or later. What they want is clear: keep the data private and under their control, and still get useful answers from the LLM.

The strongest models mostly sit behind public APIs, and that is also where vendors can test and evaluate their features properly. A model running inside the customer's own network keeps every alert and log inside their walls, and it even works at air-gapped sites. But it is hard for the vendor to manage, and it tends to struggle more with multi-step tool calls. Some vendors try a middle path: They host an open model in their own cloud instead of calling a public API. Data no longer goes to a third-party model provider, and the vendor tests the exact model it runs. Open models usually are weaker than the top API models, while the vendor now pays for the GPUs.

I went through public docs for 28 products to see who chose what. AIOps here means event correlation and incident intelligence platforms, plus the newer AI SRE agents that investigate incidents on the same data.

Short answer: most of them run the model for you. 13 give you no real say in it. 10 ship a default and let you plug in your own. 5 expect you to bring it. Three of those are open source, and the other two are commercial products that still leave the model to you.

The tools

Vendor-managed: the vendor picks and runs the model.

Hybrid: there is a vendor default and you can swap in your own.

Customer-provided: means the AI features stay off until you connect a model.

Product Strategy What the docs say
Dynatrace Davis CoPilot Vendor-managed Runs on Azure OpenAI. Terms say prompt logs are kept 35 days in Azure US East.
New Relic AI Vendor-managed Runs on Azure OpenAI. Queries are inspected and responses redacted before going to the LLM.
PagerDuty Advance Vendor-managed Powered by Amazon Bedrock models, with Anthropic Claude named by PagerDuty. Billed as AI Actions.
LogicMonitor Edwin AI Vendor-managed Addendum says customers can't choose the model. Azure, OpenAI and Bedrock are subprocessors.
Datadog Bits AI SRE Vendor-managed OpenAI is listed as an AI subprocessor, and Nebius was added in Oct 2026. No BYO option found.
Azure SRE Agent Vendor-managed Microsoft runs it. You choose Anthropic or Azure OpenAI, and the default depends on region.
AWS DevOps Agent Vendor-managed AWS FAQ says it runs on Amazon Bedrock foundation models.
incident.io AI SRE Vendor-managed OpenAI, Anthropic and Google Vertex. FAQ says no BYO model or key, because features mix models per task.
Resolve AI Vendor-managed Its own domain models plus frontier models. A dedicated single-tenant option comes with bring-your-own keys.
Cleric Vendor-managed Hosted. The CTO said they're open to customer-brought models, but nobody has asked for it yet.
Komodor Klaudia Vendor-managed Kubernetes AI SRE running on the AWS Bedrock API.
ScienceLogic Skylar AI Vendor-managed Version 2.5 (Aug 2026) runs the AI tier in sovereign cloud, on-prem or secure cloud, and tracks token usage of its own agents.
Selector Vendor-managed Its own network language model, built on an open base model. Fine-tuned per customer, on-prem at the customer.
BigPanda Biggy AI Hybrid Default is BigPanda's OpenAI service with zero data retention. Azure OpenAI can be BigPanda-hosted or customer-hosted.
ServiceNow Now Assist Hybrid ServiceNow-managed by default. BYOK through your Azure, Google AI Studio or Bedrock. BYOLLM for uncertified models.
BMC Helix AIOps Hybrid BMC's own fine-tuned Helix LLM on Vertex AI and Azure, or your own Azure, OpenAI or Google model.
IBM Cloud Pak for AIOps Hybrid Assistant runs on a mix of Llama and Granite models. Incident summary (tech preview) uses the customer's own watsonx.ai instance.
IBM Instana Hybrid SaaS uses default LLM gateways provided by Instana. Self-hosted installs connect your own watsonx.ai or vLLM.
Traversal Hybrid Runs in your cloud account (BYOC). BYOM can point it at your self-hosted model.
NeuBird Hybrid SaaS by default. VPC installs need specific Claude and Llama models enabled in the customer's own Bedrock account.
Metoro Guardian Hybrid SaaS by default. On-prem installs point it at your OpenAI-compatible endpoint, a LiteLLM gateway included.
Rootly Hybrid AI incident workflows can use your own Anthropic or Azure OpenAI key. For the AI SRE agent itself, BYOK is on the roadmap.
Robusta Hybrid Robusta-hosted models by default. You can add your own models and keys, and turn the hosted ones off completely.
HPE OpsRamp OpsPilot Customer-provided SaaS, but off until an admin adds LLM credentials. Supports Vertex AI, Bedrock, Azure Foundry and OpenAI, and recommends Claude Sonnet 4.6.
OpenObserve AI SRE Customer-provided Has a cloud SaaS, but the AI SRE runs on your own provider and key. Markets this as the difference from bundled assistants.
HolmesGPT Customer-provided Open source AI SRE agent (CNCF Sandbox). You configure the provider. Ollama is marked experimental with limited tool calling.
Aurora Customer-provided Open source AI SRE agent. Hosted providers or Ollama, with an air-gapped install bundle.
OpenSRE Customer-provided Open source AI SRE framework. Set the provider by env var, Anthropic by default, Ollama supported.

Observations

Vendor-managed wins, and it follows the billing. PagerDuty sells AI Actions and Datadog draws Bits from a credit pool, so running the model is part of the product they sell. HolmesGPT, Aurora and OpenSRE are open source with no inference bill to charge, so you must bring the model. Customers may also ask for their own model less often than you'd think. Cleric's CTO said they'd accept customer-brought models, but no customer has asked yet.

AIOps vendors with on-prem customers ship their own model instead of asking for yours. Selector fine-tunes its network model on-prem at each customer. ScienceLogic runs its AI tier in sovereign cloud or on-prem. BMC has its fine-tuned Helix LLM. This makes sense for AIOps, where telcos, banks and government still run a lot on-prem and refuse to send events to a public API.

"Bring your own" here mostly means bring your own cloud contract. BigPanda's customer-hosted Azure, ServiceNow's BYOK, NeuBird's VPC install on your Bedrock, and every OpsRamp option are cloud providers. So the model still runs in a public cloud, just under the customer's account. Running the model on your own hardware is the exception, and it shows up mostly in on-prem installs like Metoro and Traversal and in open source agents that support Ollama.

Even required BYO comes with a model recommendation. OpsRamp won't turn OpsPilot on without your credentials, but it still recommends Claude Sonnet 4.6. NeuBird lists exact model IDs to enable. OpenObserve is the one that turns bring-your-own into a selling point. Vendors let you hold the contract. They still want a say in which model sits behind their agent.

Bring your own agent, not your own model

A growing number of vendors skip the model question. They open their data to the customer's own agent over MCP, and the customer brings the agent and the model together. Most big names in the table now ship an MCP server. What differs is what they put behind it.

Opening the raw data

The simplest version gives the agent query access and lets it complete its tasks. Datadog's MCP server is now generally available, and AWS DevOps Agent already plugs it in as a telemetry source. In that setup, the reasoning is done by AWS agent, and Datadog only supplies the data. It gives agents logs, metrics, traces, monitors and incidents. But a few tools still call models hosted by Datadog's own providers, and you have to ask the support to turn those off.

Splunk went further on control. MCP server 2.0 comes preinstalled in Splunk Cloud 10.6, admins get guardrails like an SPL allow list, and every agent search carries a provenance label into the audit index. Dynatrace offers a remote MCP server too. Its open source local version warns that Grail queries are billed by GB scanned, and ships a per-session budget that defaults to 1,000 GB.

Opening the vendor's intelligence as a tool

Others expose their AI results, not just data. LogicMonitor MCP for Edwin AI puts monitoring tools and Edwin's investigation tools behind one endpoint, with inherited RBAC, rate limits, audit logs, bounded time windows and capped payloads. LogicMonitor describes MCP as the interoperability layer, while Edwin stays the intelligence layer. So your agent gets an investigation that LogicMonitor's model already produced.

ServiceNow's Action Fabric does the same with Now Assist skills, the knowledge graph, subflows and scripted REST APIs. The MCP server went GA at Knowledge 2026 and is included in every Now Assist SKU, but headless calls burn the same Assist credits as the built-in assistant. BigPanda ships a remote MCP server for Biggy's action plans, plus an A2A endpoint.

Going both ways

Some vendors also let their own agent call out. ServiceNow agents can act as MCP clients, and over A2A they can lead an external agent or take orders from one. Komodor lets customers plug their own tools and agents into Klaudia via MCP or OpenAPI, next to its 50+ built-in agents. LogicMonitor lists collaboration with ServiceNow ITOM, Microsoft Copilot and customer MCP servers.

What this changes

Bring-your-own-agent solves the model question only for raw data tools. Once your agent calls a vendor skill, like Edwin's investigation or a Now Assist skill, the vendor's model runs again, and your data goes with it. The meter doesn't go away either. It moves, to GB scanned at Dynatrace and to Assist credits at ServiceNow.

It may also explain the Cleric quote above. Maybe customers rarely ask for their own model because what they really want is their own agent.

Vendor-managed

Good for the customer who wants it to just work. Bad for the one with strict data rules.

Pros:

  • Works the day you turn it on. No API key, no endpoint, no GPU.
  • The vendor tests prompts and tool calls against models it knows, and can route each task to the model that does it best.
  • One contract, one DPA. Zero data retention with the LLM providers is the vendor's job (BigPanda states it for its default OpenAI service).
  • Pricing is a credit or an action count, not a token bill you have to forecast.

Cons:

  • Your data goes to the vendor's subprocessor on the vendor's terms. Dynatrace, for example, keeps prompt logs in Azure US East for 35 days.
  • No model choice. One LogicMonitor user put it bluntly: why pay for more tokens when we already have Bedrock or Copilot?
  • You can't decide who sees your data. Azure SRE Agent defaults to Anthropic outside the EU, and Anthropic is not covered by Microsoft's EU Data Boundary.
  • Air-gapped or fully on-prem is basically off the table.

Customer-provided

Bringing your own API key is not the same as self-hosting. With the key, your data still goes to AI companies, just under your contract instead of the vendor's. Only a model running inside your own network can keeps the data in. Alation's BYOM write-up makes the same point.

Pros:

  • The data path sits under your own contract, region and retention settings.
  • You reuse LLM spend you already committed, instead of paying for tokens twice.
  • Self-hosting makes air-gapped setups possible, which some regulated sites require.
  • You pick the model, and you decide when to upgrade it.

Cons:

  • Quality depends on your pick now. HolmesGPT says Ollama tool calling is limited and can give inconsistent results. OpsRamp names one recommended model for a reason.
  • Vendor support gets thin past the defaults. HolmesGPT tells you to try a hosted model first and make sure everything works before switching to Ollama.
  • The vendor tuned its prompts and evals on its own model. Your model might pass simple chat and still break multi-step agent features.
  • Capacity, latency and GPU cost become your problem.

That third con is the one I'd watch most. A summary feature degrades gently with a weaker model. An agent that has to pick the right tool five times in a row does not.

Sources

Everything comes from public docs, trust pages and vendor blogs, checked in early November 2026. "No BYO option found" means I didn't find one, not mean the vendor doesn't offer it. LogicMonitor and incident.io are the exceptions, since their docs say no outright. Some products have more than one AI feature, and I looked at the main assistant or agent only. I left out general observability and ITSM tools that don't sell event correlation or incident investigation. Docs in this space change every few months, so check the vendor page before you quote a row. If I got one wrong, tell me in the comments and I'll fix it.

Each row in the table links to the page it comes from. Other pages used, by section:

What the split tells you and the trade-offs

Bring your own agent