Ask five things before anything else. Do you sell directly to end clients, ever. Which surfaces do you correlate, and which do you only receive alerts from. Who is my named contact, and does that person change. What does the contract say about exit, and how long does exit take. What is the onboarding window and what do I have to supply. Everything below expands those five, including the SLA terms worth reading twice.

What to check on coverage

Ask which surfaces are correlated rather than which are supported. A provider receiving an alert from your firewall is not the same as one reading firewall traffic against identity and endpoint activity together. Ask for a worked example of a detection crossing two surfaces. If they cannot give one, the correlation is a data feed.

What to check on the channel model

Ask whether an end-client price list exists anywhere. Ask what happens to an inbound lead from one of your clients. Ask whether the provider has ever sold direct in your market. A channel-only provider answers all three in one sentence.

What to check on the named contact

Ask whether the security contact is assigned or pooled, whether they change on holiday or escalation, whether they join your client calls, and what tier they sit behind. A named contact available only on the top tier is a sales feature, not a service model.

What to check in the contract and SLA

Six terms. Notice period and any auto-renewal. Whether an early exit exists and what it costs. Who owns your log and detection data on exit and in what format it leaves. What the response commitment actually measures, acknowledgment or containment. Whether the SLA carries a remedy or is a statement of intent. And whether any published SLA figure has a source behind it.

What to check on onboarding

Ask for the window in days and what drives it. Ask what you supply and what they do. Ask whether hardware is needed and in which cases. A provider quoting a fixed number for every environment has not scoped yours.

What to check on exit

Ask how long a full offboarding takes, whether you can run a replacement in parallel, and what the provider does with your historic data. The answer tells you how confident they are you will stay.

The phrase that's going to follow you into every deal

Heimdal launched AI Wingman in April 2026 with a positioning statement worth paying attention to. The company argued that competitor AI is "bolted onto the edge" of existing products, and that genuine AI integration requires something different.

They're right. And they're also exactly describing a problem that applies to a lot of vendors in this space, including some that have been making agentic SOC claims for much longer than Heimdal has.

The phrase "bolted onto the edge" will land in procurement decks and vendor pitches for the next 12 months. You need to know what it means before a vendor uses it to describe their competitor and not themselves.

What bolt-on AI actually is

Bolt-on AI is a capability layer added to a product that was not originally built with AI in the detection loop. The AI handles some tasks, typically alert enrichment or ticket categorization, but the core detection engine and the core response workflow run independently of it.

The tell is what happens when the AI component is switched off or unavailable. If the SOC keeps running with minimal visible degradation, the AI was bolted on. If the detection and response quality drops significantly, the AI was genuinely integrated.

Bolt-on AI is not useless. It reduces noise. It categorizes faster. It handles the straightforward, single-signal alerts that would otherwise take an analyst three minutes to close. The problem is that the attacks worth worrying about in 2026 are not straightforward single-signal events. They span identity, endpoint, email and network layers simultaneously. Correlated detection across those layers requires AI that is part of the detection logic, not sitting on top of it.

What integrated AI SOC looks like

Integrated AI is built into the correlation engine. It does not process alerts after they are generated. It informs what gets flagged as an alert in the first place.

The practical difference is that integrated AI reduces false positives at the detection layer rather than at the triage layer. That is where the real workload reduction happens. An analyst reviewing 50 high-confidence alerts is doing different work from an analyst reviewing 200 low-confidence alerts that an AI layer has sorted into priority order. The latter is still reviewing 200 alerts. The former is reviewing 50.

This is also where the alert fatigue problem either gets solved or gets shuffled downstream. Bolt-on AI shuffles it downstream. Integrated AI addresses it at the source.

Seven checks before you believe an agentic SOC claim

1. Where does the AI sit in the detection workflow?

Ask the vendor to walk you through the detection pipeline step by step and identify where the AI model operates. If it operates after alert generation, it is enrichment. If it operates within the correlation engine, it is integrated. These are different things and the vendor should be able to tell you which one they have built.

2. What happens to detection quality if the AI is unavailable?

This is the fastest way to identify bolt-on AI. If the answer is "the alerts keep coming but they take longer to categorize," the AI is processing output. If the answer is "detection coverage and false positive rates change significantly," the AI is part of the input layer.

3. Can the AI take containment actions, or does it only make recommendations?

Agentic AI should be able to initiate response actions, not just surface suggestions for a human to action. If the AI component generates a recommended action and then waits, it is an AI-assisted workflow, not an agentic SOC. Ask specifically: what response actions can the AI initiate autonomously, and what is the escalation path when it does?

4. How does the AI perform on multi-vector attacks?

Single-vector threats are the easy case. The harder case is a credential stuffing campaign that pivots to lateral movement and then exfiltrates data over three days. Ask the vendor for examples of how their AI handled multi-stage attacks in real client environments. Vendors with genuine integrated AI can provide these. Vendors with bolt-on AI typically cannot.

5. Is the AI trained on your clients' environment or on generic data?

An AI model trained on generic threat data performs differently from one that has learned the baseline behavior of a specific environment. The question is whether the vendor's AI models are tuned per-tenant or whether they apply a one-size model across all clients. Per-tenant tuning is operationally harder to build and maintain, and vendors that have done it will tell you exactly how it works.

6. What does the AI hand off to a human analyst, and when?

The best AI SOC implementations have a clear and documented escalation model. The AI handles the high-volume, high-confidence routine work. Human analysts handle the ambiguous, complex and high-stakes events. If a vendor cannot describe that handoff model precisely, the AI and the human workflows are not genuinely integrated.

7. How is AI performance measured and reported to the MSP?

If the vendor does not provide MSP-level reporting on AI detection accuracy, false positive rates and escalation rates, you cannot evaluate whether the AI is performing as claimed. Ask for a sample report. If the reporting does not exist or covers only high-level metrics, that is relevant information about how seriously the vendor treats the AI as an operational component rather than a marketing claim. enhanced.io's reporting covers these metrics at the client environment level.

The bigger picture

The agentic SOC category is going to generate a lot of vendor claims in 2026. CrowdStrike, ConnectWise, Kaseya, Heimdal and SentinelOne all shipped AI SOC updates in the same window. That's not coincidence. That's a market responding to demand from MSPs who are being asked by their clients to deliver more security with the same team.

The seven checks above are not designed to make every vendor fail. They're designed to separate genuine capability from repositioned marketing. Most vendors will pass some of them. The ones worth talking to will pass all seven.

Check your current providers against this list before the next renewal conversation. The gap between integrated AI and bolt-on AI is not a theoretical one. It shows up in alert volumes, escalation rates and, eventually, in breach outcomes.