Nearly every managed service provider now describes itself as AI-powered. The label spans an enormous range — from a genuinely automated remediation pipeline that resolves incidents without a human, to a ticketing system with a summarization feature, to a marketing page. Buyers need a test that distinguishes them, because the difference determines whether you are paying for capability or for vocabulary.
Gartner supplied the useful frame with "agent washing" — rebranding chatbots, robotic process automation, and AI assistants as agentic without the underlying agentic capability. Their June 2025 assessment was that of the thousands of vendors claiming agentic AI, roughly 130 were the real thing. Adoption data tells the same story from the buyer side: around 62% of organizations are experimenting with agentic systems while only about 23% have scaled them in even one function. Production deployment is rare. If a provider tells you autonomous AI is running their operations, the base rate says they are most likely describing a pilot.
This does not mean AI in managed services is vapor. It means the claim needs the same diligence you would apply to any other capability assertion — and unlike most, this one has questions that produce unusually clean signal.
First, Separate Three Different Claims
"AI-powered MSP" conflates three propositions with different value to you. Force the provider to say which they mean.
1. We use AI internally to run our operations more efficiently. Ticket classification and routing, alert correlation and deduplication, automated triage, drafted responses, predictive maintenance. This is the most common and most credible claim. Its value to you is indirect — potentially faster response and lower cost — and it should show up in SLA commitments rather than in a feature list. If a provider claims AI-driven efficiency but their contractual response times match everyone else's, ask why.
2. We deploy AI-driven security tooling in your environment. Behavioral detection, anomaly identification, automated containment. Also largely credible — most modern EDR and XDR platforms use machine learning for detection, and have for years. The thing to establish here is whether the provider is describing their capability or their vendor's. "We use an AI-powered EDR platform" means they bought a product. That is fine, and worth knowing, but it is a procurement decision rather than a differentiator.
3. We will help you evaluate, deploy, and govern AI tools in your business. A genuinely emerging service line and often the most valuable of the three, especially the governance part. Also the one where capability varies most wildly, because it requires expertise the traditional MSP skill set does not include.
A provider who cannot cleanly separate these three when asked is a provider whose AI story is marketing rather than architecture.
Six Questions and What Real Answers Sound Like
1. "Which decisions does your AI make without a human reviewing them?"
This is the question that cuts through agent washing fastest, because autonomy is the actual dividing line. A real answer is narrow and specific: "Ticket classification and routing are fully automated. Password resets for standard users after MFA verification are automated. Everything touching production configuration goes to an engineer." That is a provider who has thought about where autonomy is safe.
A weak answer is expansive and vague — AI "handles" support or "manages" your environment. Follow up with: what is the most consequential action it can take unilaterally, and what is the rollback? Providers running real automation know exactly where that boundary sits because they had to decide where to put it.
2. "What percentage of tickets are resolved with no human involvement, and how do you measure that?"
A specific number with a stated measurement definition indicates instrumentation. Note that the definition matters enormously — "AI-assisted" resolution where a human approved a suggested fix is a different and much lower bar than autonomous resolution, and providers routinely blur them. A provider who quotes a high automation rate and cannot explain the measurement is quoting a vendor's marketing figure.
3. "Is this in production for clients like us, or is it a pilot?"
Ask directly, and treat an honest "we're piloting it with three clients" as a good sign rather than a disqualification. Given that only about 23% of organizations have scaled agentic systems in even one function, a provider claiming mature production deployment across their base is making an unusual claim that warrants unusual evidence. Ask how many clients, for how long, and what broke early on. Providers with real deployments have failure stories; providers with slide decks do not.
4. "What happens to our data?"
The question with the most concrete risk attached. Establish specifically: does our data — tickets, logs, configurations, documents — get sent to a third-party model provider? Is it used for training? Is it retained, and for how long? Is it segregated from other clients' data? Which subprocessors are involved?
If you have regulatory obligations, this escalates from diligence to a compliance requirement. Under HIPAA, a provider routing ePHI-adjacent ticket content through an AI service has created a subprocessor relationship that needs to appear in your business associate arrangements — see our guide on HIPAA and IT services. For defense contractors, CUI transiting an AI service is a scoping question with the same weight as any other data flow, discussed in our coverage of CMMC and MSP scope. A provider who has not thought this through for their own AI tooling will not manage it well for yours.
5. "Who is accountable when the AI is wrong?"
Automation fails differently from people: less often, and then all at once across every client simultaneously. Establish what happens when an automated action causes an outage — is it covered under the same SLA and remedy structure as human error, or is there an exclusion for automated actions? An exclusion carving out AI-driven actions from the provider's liability transfers the risk of their efficiency gain onto you, and it is worth reading the agreement specifically for that clause.
6. "What does your AI do when it's uncertain?"
A well-engineered system has an explicit confidence threshold and escalates below it. A provider who can describe that mechanism — "below a confidence threshold it routes to a human with the reasoning attached" — is describing a real system. A provider who has never considered the question is running something they do not understand well enough to be accountable for.
Where AI Genuinely Helps a Managed Service — and Where It Doesn't
The realistic value today concentrates in high-volume, low-consequence, pattern-heavy work: ticket classification and routing, alert correlation and noise reduction, documentation drafting and knowledge base maintenance, log analysis for known patterns, and first-draft responses to routine requests. These are real efficiencies, and a provider who has automated them well may genuinely deliver faster response at lower cost.
The value is thinner where consequences are high or context is deep: architecture decisions, incident response judgment under ambiguity, and anything requiring knowledge of your business rather than your systems. An AI agent that can route a ticket cannot decide whether taking your ERP offline for emergency patching is the right call on the last day of your quarter. That judgment is what you are paying a good provider for, and it is precisely what automation does not supply.
There is also a second-order effect worth watching in the pricing conversation. As automation absorbs routine labor, the link between seat count and provider effort weakens — which is what is driving the move away from per-user billing, covered in our piece on outcome-based MSP contracts. A provider making strong automation claims while holding per-user pricing constant is capturing the efficiency gain; that is a legitimate business decision, but you should recognize it and factor it into the negotiation.
The Governance Service Line — Worth More Than It Looks
Of the three claims, the third is the one most SMBs actually need and least often ask for. Your employees are already using AI tools, almost certainly including ones nobody approved. Someone has pasted a customer list into a chatbot. A provider who can help you inventory that shadow usage, write a workable acceptable-use policy, evaluate tools against your data-handling obligations, and monitor for unsanctioned services is solving a live problem.
Ask what they have actually done for other clients here. Cross-client due diligence frameworks are increasingly becoming the standard template clients send to their providers, so an MSP who can answer a rigorous AI questionnaire about themselves has a sales asset — and an MSP who cannot is unlikely to help you build one.
The practical takeaway: open with "which decisions does your AI make without a human reviewing them?" and follow with "what happens to our data?" The first separates real automation from repackaged tooling in about thirty seconds, because providers with production automation know exactly where their autonomy boundary sits. The second surfaces a compliance exposure that most providers have not examined in their own stack. Everything else in the AI conversation is negotiable; those two answers tell you whether there is a capability underneath the label.
Frequently Asked Questions
What is "agent washing"?
Gartner's term for rebranding chatbots, robotic process automation, and AI assistants as agentic AI without the underlying autonomous capability. Their June 2025 assessment found that of thousands of vendors claiming agentic AI, roughly 130 genuinely qualified.
Is an MSP using an AI-powered security platform the same as being an AI-powered MSP?
No. Most modern EDR and XDR products use machine learning for detection, so deploying one is a procurement decision rather than a capability of the provider. Ask whether they are describing their own automation or their vendor's product — both are worth knowing, but only the first is a differentiator.
Should we be suspicious of any MSP claiming AI capability?
Not suspicious — specific. Real efficiencies exist in ticket classification, alert correlation, and documentation. The concern is scope inflation: with only about 23% of organizations having scaled agentic systems in even one function, a provider claiming mature autonomous operations across their client base is making a claim that requires evidence. An honest "we're piloting this" is a better signal than a confident generality.
What's the biggest risk in an MSP's use of AI?
Data handling. Ticket content, logs, and configurations routed through third-party model providers create subprocessor relationships with retention, training, and segregation questions attached — and regulatory consequences if you handle ePHI, CUI, or cardholder data. The secondary risk is a liability exclusion carving automated actions out of the provider's SLA remedies.
← Vetting Ransomware-Resistant Backup Claims | The 2026 Shift to Outcome-Based MSP Contracts →