---
title: "How to Select an Outsourced Cybersecurity Provider (MDR & SOC Framework)"
description: "A CTO's framework for selecting MDR and outsourced security providers. Verify SLAs, coverage scope, AI detection capabilities, and governance before signing."
url: https://onesight.global/blog/how-to-select-an-outsourced-cybersecurity-provider-mdr-soc-framework.html
source: https://onesight.global/blog/how-to-select-an-outsourced-cybersecurity-provider-mdr-soc-framework.html
generated_by: aeo.js
---

# How to Select an Outsourced Cybersecurity Provider (MDR & SOC Framework)

A CTO's framework for selecting MDR and outsourced security providers. Verify SLAs, coverage scope, AI detection capabilities, and governance before signing.

You've already decided to outsource security monitoring, so the hard question of whether isn't yours anymore. The deal stops breaking on the decision and starts breaking when you sign with a vendor whose coverage boundaries, SLAs, and headline time-to-detect figures evaporate once real traffic arrives. That is the actual failure mode here: an outsourced SOC that overpromises under production load. Time-to-response claims measured on quiet synthetic traffic, SLAs so vague they can't be enforced, coverage that leaves identity or cloud workloads outside what's actually watched, and a MSSP-vs-MDR-vs-ERD taxonomy that means nothing until scope and accountability are pinned down in writing.

The consequence isn't theoretical. When a confirmed compromise lands at 2am, you need to know exactly who closes the loop — and most vendor sales decks go silent on that question. The labels themselves are unreliable: MDR vendors sell "detection," MSSP vendors resell monitoring, and ERD is routinely confused with EDR/XDR tooling categories that have nothing to do with a service model. Buy any of them at face value and you can end up paying an MDR price for alerting forwarded to an unmonitored mailbox after hours.

This article gives you the framework for how to select MDR (and MSSP, ERD, or managed SOC) vendors before you sign. You'll get a four-model taxonomy keyed to accountability rather than marketing, a checklist of measurable criteria tied to verifiable evidence — enumerated in-scope telemetry, SLAs checked against your own pre-adoption baseline, and real tests for whether AI-augmented detection actually detects — plus the red flags that should knock a vendor out early. The goal is a ranked shortlist you can defend at the board level, where every KPI was measured against production reality instead of a slide.

## When an outsourced SOC is the pragmatic choice

### Outsourcing monitoring is the sensible default

For most enterprises, outsourcing security monitoring isn't a fallback — it's the expected move. Building genuine in-house SOC depth means staffing analysts across overlapping shifts to cover attack windows that don't close at 5 pm; few organizations can justify that bench on their current threat profile. The right triggers are clear: a security posture assessment, pen testing, or a genuine 24/7 SOC rather than business-hours monitoring dressed up as continuous coverage.

Several common situations call for this move:

- Consider an organization expanding into a new market that needs GDPR-aligned monitoring but has no analyst bench for continuous coverage.
- Consider a company whose attack surface — cloud workloads, distributed terminals, and a remote workforce — outgrew what two generalist IT staff can watch.
- Consider an org that runs a one-off pen test and posture assessment, then realizes ongoing detection must run continuously without hiring an internal SOC.

### The failure is vendor performance, not outsourcing

Picking the wrong partner is the genuine risk; the act of outsourcing usually isn't. Deals break when you sign with a vendor whose coverage boundaries, SLAs, and headline detection numbers evaporate once real traffic hits production — unverified claims are the actual trap here. When you vet vendors, treat governance and credentials as a competitive advantage of a well-vetted partner rather than compliance burden. If your make-or-buy question remains open, start with the dedicated [build-vs-partner guide](https://onesight.global/blog/cybersecurity-outsourcing-vs-in-house-soc-a-cto-s-build-vs-partner-guide.html) rather than this selection framework. This one assumes you've already committed to outsourcing, so the first job becomes defining exactly what category you're buying — MSSP, MDR, ERD, or managed SOC.

## How to Select MDR Vendors (and MSSP, ERD, Managed SOC)

Section 1 established that outsourcing monitoring is the default and selection is where deals break. Before you can hold a vendor to accountability in later sections, you need to know which of four models you are actually purchasing. Vendors conflate them constantly — which is exactly why the [best practices when outsourcing cybersecurity](https://onesight.global/blog/best-practices-when-outsourcing-cybersecurity.html) exist to separate genuine models from rebranded ones; the worst deal you can sign is alerting sold under an MDR name.

### MSSP: breadth-first monitoring that stops at alerting

An MSSP aggregates and monitors telemetry at scale but historically stops at notification. It is breadth-first coverage — it tells you something looks wrong without taking ownership of investigating why, containing what spread, or remediating the root cause. The accountability boundary ends where your internal team's begins. A typical failure mode is an organization that signs an MSSP, receives thousands of alerts, and finds no one owns what happens next. That breadth-first trap is what pushes most buyers toward MDR once real traffic arrives.

### MDR: outcome-owned detection and response

An MDR vendor owns outcomes rather than notifications — detection, investigation, containment, and remediation support all sit on its side of the accountability line, with responsibility for results. You are buying someone to act through an incident, not page you and step away. The test that separates a real MDR from a rebranded MSSP is blunt: when a confirmed compromise lands at 2am, who closes the loop?

### ERD: embedded analysts (not EDR or XDR)

An ERD embeds analysts within, or closely aligned to, your internal team — blending outsourced depth with internal control so you retain continuity and oversight while gaining capability. For example, an organization might place a vendor analyst directly on its small security desk to gain investigation depth while keeping incident decisions in-house.

Two traps. First, ERD is frequently confused with EDR (endpoint detection and response) and XDR (extended detection and response). Those are tooling categories — broader telemetry across endpoints, network, and cloud — not a service model. Confirm whether you are buying embedded analyst time or simply wider sensor coverage. Second, "embedded" must mean the analysts operate as an extension of your desk under your runbooks and access controls, not a vendor team running its own playbooks behind your perimeter.

### Managed SOC: an off-site in-house function

A managed SOC builds and runs a dedicated security operations center off site — SIEM, analysts, playbooks, 24/7 shift coverage — as the closest external analog to an in-house function. It suits organizations that want full SOC capability without twenty new hires. A common shape for this is an organization that wants the complete SOC — analysts, SIEM, playbooks — but not the headcount, so it contracts off-site delivery rather than building internally. That is precisely the shape of Onesight's partner-run 24/7 AI SOC, operated on your behalf rather than staffed internally.

### The labels lie — pin down accountability in writing

The four categories overlap heavily in practice: MDR vendors sell "detection," MSSP vendors resell monitoring, and category names carry no guarantee until you test them. Before trusting any label, require the scope, the accountability tiers, and the response obligations pinned down in writing — specifically which party owns investigation, containment, and remediation, and where each accountability tier ends. A vendor's category name is a starting hypothesis, not a contract term.

That vocabulary is what Section 3 puts to the test against real production traffic before any SLA signing.

## Coverage and scope — define the boundaries in writing

Section 2's labels carry no accountability guarantee until you pin down what is actually monitored. A category name says nothing about which systems feed detection or who responds at 2am. That must be written, enumerated, and testable before any SLA signature.

### Pin down the telemetry inventory in writing

A defensible scope enumerates endpoints, cloud workloads, identity systems, and network telemetry as a defined inventory — not one vector, not a blanket "we monitor everything." If the contract does not list which systems feed detection, coverage is whatever the vendor decides during onboarding.

Coverage gaps are where promises fail. A vendor may cover endpoints and cloud while leaving identity systems unmonitored — SSO, Okta, SAML sessions — precisely the vector account takeover exploits. The fix is an explicit asset/telemetry inventory listed as in scope before signing, not deferred to onboarding. The same drift applies at runtime: a contract promising "comprehensive monitoring" without an attached inventory lets the vendor exclude any system not explicitly enumerated, so a new Kubernetes cluster falls silently out of coverage until a change order is signed.

### Test what '24/7' actually means

'24/7' frequently means business-hours staffing with after-hours alerting only. The separating test is operational: who is on shift at 2am, at what tier, and how escalation works outside core hours — not the word "continuous" in a slide. A genuinely continuous operation keeps monitoring around the clock rather than forwarding alerts to an unmonitored mailbox; a partner-run 24/7 AI SOC is one model built for that, but any provider must prove its off-hours coverage with specifics, not a label.

![A dark office building at night with just one desk window glowing warm, showing someone genuinely on shift.](https://lm3.onesight.global/blog/2026/11/how-to-select-an-outsourced-cybersecurity-provider-mdr-soc-inline-2-80c40a.png)

### Make off-hours response a contract term

Require the contract to state monitoring hours per telemetry source and the exact response expectation outside core hours. 'After-hours' should name a defined responder on call and a target time-to-triage; otherwise it means first triage at the next business day. An "after-hours" SLA that lands alerts in a queue at 11pm with first triage at 8am by a tier-1 analyst gives a compromise hours of unmonitored runway before anyone looks. That inventory is also the denominator against which every detection-latency and response-SLA KPI is later verified — fix these boundaries now.

## SLAs and performance — verify against your own baseline

### The five criteria, measured against your inventory

Every SLA KPI must be expressed against the asset/telemetry inventory you enumerated as in scope before signing. Without that denominator, a number is marketing. The five you pin down are: detection latency (time from telemetry to alert), response/containment time, mean time to restore (MTTR), escalation paths by tier, and reporting cadence. A detection-latency figure means nothing for a system never confirmed feeding detection — so if your identity provider isn't in scope, the vendor's time-to-alert metric silently excludes the account-takeover vector you care about most.

### Why headline numbers evaporate

Vendor time-to-detect and time-to-response figures frequently collapse under production load. They are often measured on low-volume synthetic test traffic where a single analyst handles every alert, so a quoted "30 minutes" reflects a quiet quarter, not real alert volume. During an actual multi-vector incident the same team faces thousands of alerts, and the median time-to-detect stretches past two hours before anyone isolates the true threat.

### Establish your own baseline

Before signing, log your current change failure rate and MTTR over 90 days from Jira, GitLab, or Splunk so you measure the vendor's actual delta instead of accepting an invented starting point. When a vendor later reports "we cut MTTR by 40%," you confirm it against your real baseline rather than trusting their fabricated start. Without a buyer-owned number, performance claims cannot be defended at the board level.

### Demand the right KPIs

Demand metrics measured against that baseline: change failure rate, MTTR, detection precision/recall, and escape rate (confirmed incidents never alerted) — not raw alert counts or "99% coverage." A provider firing 200 alerts a day with a 3% true-positive rate is drowning your analysts in noise; the honest signal is precision (true positives divided by total alerts) and escape rate.

### Reporting that reveals where it failed

Reporting must show trended KPIs over time and incident-level detail, not a summary percentage. A monthly "98% SLA compliance" green figure can hide that both misses were after-hours identity-provider compromises; the incident breakdown shows which telemetry source missed and why. Note that even an AI-driven SOC is one delivery model — its detection latency and containment figures require the same verification as any other vendor's headline number.

## Detection capability — AI-augmented detection is a claim to verify

Section 4 gave you what to measure and your own pre-adoption baseline; this section tests whether the detection engine actually delivers against that baseline before any SLA is signed, then Section 6 moves to trust, credentials, and governance as the assurance layer that makes those verified KPIs defensible at board level.

### Treat "AI-powered" as a capability to verify, not a label

The phrase "AI-powered detection" carries no technical content until you force the vendor to show the mechanism. A genuine ML detector must name its training data, feature sets, retraining cadence, and how it is validated against real incidents. A vendor that cannot say which telemetry its model ingests is likely selling correlation rules dressed as artificial intelligence — a SIEM search on a static threshold (more than five failed logins) rebranded as machine learning.

Probe with four questions: How are the models trained, on which data sources, how often are they retrained, and at what false-positive rate in production? Vendor headline numbers evaporate once real traffic arrives — that is exactly the pattern where AI gains disappear in production while sales slides stay intact. If the answer stops at "our model detects anomalies," walk away; anomaly detection with no named feature space or validation set is unfalsifiable marketing.

### Threat hunting must sit beside alerting

Automated alerting is reactive: it fires only on configured patterns, so telemetry flags exactly what it has been told to look for and nothing beyond it. A documented hunting cadence plus red-team validation proves the partner finds what automated systems miss. Both must be in scope for MDR or ERD. Consider an attacker who exfiltrated data in small daily transfers that stayed under every alert threshold; automated logging recorded nothing suspicious over a week until proactive hunting correlated volume across days surfaced the slow bleed. Alerting alone would have missed it entirely. A vendor that only alerts has not bought detection — it rented a dashboard.

![A lone analyst scanning shadowed corners with a narrow beam of light while a fixed sensor glows dimly in the distance.](https://lm3.onesight.global/blog/2026/11/how-to-select-an-outsourced-cybersecurity-provider-mdr-soc-inline-1-31ce47.png)

### Telemetry quality bounds everything downstream

Detection can be no better than the data feeding it, and correlation lives in clean SOAR/tooling integration rather than algorithm prestige. Logs arriving without AWS CloudTrail field-level detail leave an IAM role assumption to read secrets invisible at the source; no model catches what telemetry never carries. Ask how the vendor ingests, retains, and correlates your identity, cloud, and endpoint data against the inventory you defined in Section 3 — not a generic tool list.

### Handoff that doesn't rebuild the silo

Containment is not the end of the contract. The handoff must transfer full incident context — investigation notes, the decision trail, containment actions — to your team without recreating the knowledge silos that plague augmented teams. A vendor that closes a ticket with "resolved" and no notes forces your people to re-triage from zero when the same indicator recurs. Require documented runbooks, escalation triggers, and a clear RACI defining who does what during an incident; absent these, the alert transfers but the context does not.

## Trust, credentials, and governance — an advantage, not a burden

Once you have verified the detection engine actually delivers, credentials and governance are the assurance layer that makes those KPIs defensible at board level and protects your data while a partner runs it. This is where most selection processes go quiet: buyers accept "we're compliant" as a finish line instead of reading the underlying report — the same gap our analysis on the [truth in outsourcing cybersecurity](https://onesight.global/blog/the-truth-in-outsourcing-cybersecurity.html) dissects, separating vendor labels from verifiable accountability.

### What SOC 2 actually is — and why the label lies

"SOC 2 certified" means nothing until you read the document behind it [[1]](https://www.aicpa-cima.com/news/article/aicpa-proposes-update-of-soc-2-r-trust-services-criteria-to-enhance-the-relevance-usability-and-value-of-cpas-reports). Formally, SOC 2 is an examination performed by a licensed, independent CPA under the AICPA Attestation Standards that evaluates controls relevant to security, availability, processing integrity, confidentiality, or privacy of an organization's systems. The phrase has degraded into a marketing checkbox precisely because anyone can print "SOC 2" without disclosing what was examined.

Two things must be in scope before you trust the report: which Trust Services criteria were actually audited, and whether the audit covers the cloud environments and subprocessors where your data really lives. A sanitized subset — controls over the vendor's corporate network while your data sits in an unlisted tenant — is not assurance, it is a narrower claim dressed as a blanket one.

### Require Type II, not Type I

Ask for a Type II report, not a Type I [[1]](https://www.aicpa-cima.com/news/article/aicpa-proposes-update-of-soc-2-r-trust-services-criteria-to-enhance-the-relevance-usability-and-value-of-cpas-reports). Type II demonstrates operational effectiveness of those controls over a period, typically six to twelve months; Type I only attests that controls were suitably designed at a single date. A snapshot proves nothing about how the partner behaved during an actual incident window or under production load — it shows the design was sound on the day the auditor walked through. The buyer who needs assurance the partner won't leak data in month eleven receives a snapshot, not a track record.

### GDPR Article 28 and the addendum that backs it up

Whenever a partner ingests, stores, or forwards your data, processing must be bound by contract under Article 28 [[2]](https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX:32016R0679#art_28). Those obligations are: act only on documented instructions from you as controller, keep authorised personnel committed to confidentiality, apply the technical and organisational security measures of Article 32, and respect the conditions for engaging another processor. Each is a contractual term, not a best-practice suggestion.

The addendum that enforces them must disclose data residency, name or scope subprocessors, and grant you a right-to-audit clause. Without it you cannot verify the partner's controls, and when a breach traces through a downstream subcontractor's CI/CD pipeline you have no contractual footing to demand that subcontractor's controls or force remediation under Article 28(1)(d). A standard DPA your legal team signs blind is exactly how this gap appears.

### Governance as a differentiator

The well-vetted partner folds auditability into its delivery model rather than bolting it on as compliance theater. A provider built around rigorous security and auditability — one that treats controls as part of delivery, not a checkbox — is structured to produce an independently audited report and a clean DPA when the board asks whether you can trust it with customer data, because that structure is what lets it defend its detection KPIs in the first place. Governance then converts an uncomfortable board question into a defensible answer instead of a tax your legal team forces you to pay.

These same criteria fold into the red-flag checklist and scoring rubric in Section 7, so you can rank shortlisted vendors on credentials rather than slides.

## Red flags, the vetting questionnaire, and scoring your shortlist

The criteria built across this article — taxonomy, scope, SLA verification, detection capability, governance — are only useful if you can turn them into a decision. This section distills them into three tools: a red-flag scan to disqualify vendors fast, a questionnaire that forces falsifiable answers, and a scoring rubric to rank the survivors against measurable criteria so the final pick survives a board question.

### The disqualify-fast red-flag scan

A shortlist collapses quickly once you scan for failure modes rather than features. These are the same breakdowns from the earlier sections, consolidated into a kill list:

- Unverifiable KPIs. Every headline number cites "vendor-measured" with no methodology. Ask how detection latency was computed against what denominator, and the answer is usually "our internal tooling." A figure without an enumerated inventory and a stated method is marketing.
- Scope that refuses to enumerate telemetry. The vendor cannot list which endpoints, cloud workloads, identity systems, and network sources are in scope; coverage becomes whatever it decides during onboarding.
- '24/7' that is business-hours staffing with after-hours alerting only. No one is on shift at 2am; alerts forward to a queue for next-business-day triage.
- No pre-adoption baseline culture. The vendor quotes its own starting point instead of measuring your delta against numbers you logged before adoption. Without it, performance claims cannot be defended at the board level.
- Reporting that shows summary percentages while incident-level detail never surfaces — headline compliance figures that mask exactly where coverage failed.

Any one of these is a legitimate reason to remove a vendor before scoring begins.

### The vetting questionnaire

The purpose of the RFI is not to collect polished answers but to force concrete ones. Each question demands a falsifiable response:

- Telemetry scope. Which sources are in scope (endpoints, cloud workloads, identity systems such as SSO/SAML, network)? How is the inventory enumerated and updated when a new system goes live?
- True monitoring hours per source. State monitoring hours for each telemetry type rather than a blanket "24/7." Who is on shift at 2am, at what tier, and how does escalation work after core hours?
- Detection latency versus containment. Report detection latency (telemetry to alert) separately from response/containment time, with the measurement window and denominator.
- AI training and false-positive rate. How are models trained, on which telemetry, at what retraining cadence, and what is the production false-positive rate?
- Escalation paths by tier. What triggers escalation from tier 1 to tier 2, and who owns containment off-hours?
- SOC 2 as a Type II report. Deliver the current Type II report — an independent CPA examination over a period of operational effectiveness [[1]](https://www.aicpa-cima.com/news/article/aicpa-proposes-update-of-soc-2-r-trust-services-criteria-to-enhance-the-relevance-usability-and-value-of-cpas-reports) — not just the claim "we're certified." Confirm which Trust Services criteria were in scope and whether subprocessors and cloud environments are covered.
- GDPR Article 28 obligations. Provide a data-processing addendum binding the processor to act only on documented instructions, keep personnel confidential, apply Article 32 measures, and respect subprocessor conditions [[2]](https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX:32016R0679#art_28), plus named or scoped subprocessors and a right-to-audit clause.

### The scoring rubric

Score each surviving vendor against measurable criteria — not price, not how polished the demo was:

- Coverage breadth — which assets/telemetry are enumerated in the contract, and whether identity and cloud sources are included.
- Verified KPIs measured against your own pre-adoption baseline rather than the vendor's invented starting point.
- Off-hours response accountability — who closes the loop at 2am, with a named responder and target time-to-triage.
- Governance artifacts — a Type II report dated months back plus a DPA with right-to-audit and named subprocessors.
- Handoff quality — full incident context (notes, decision trail, containment actions) transferred without rebuilding the knowledge silos that plague augmented teams.

Tabulate results from an identical two-day RFI across every vendor. The winner is usually neither the cheapest nor the loudest on "AI," but the one that produced a Type II report dated eleven months, an incident-level escape-rate table, and a named after-hours responder — defensible because it was scored against your baseline rather than its marketing.

### Apply these criteria directly

If running this capability assessment is time you'd rather spend elsewhere, Onesight can evaluate your outsourced-security needs against the same rubric. It offers military-grade AI security including red teaming, zero trust, cloud hardening, and a 24/7 AI SOC operated on behalf of the client so enterprises can detect, neutralize, and harden before attacks reach their perimeter. Bring your shortlisted vendors in for a capability assessment or discuss your needs directly — https://onesight.global/contact.

## References

1. [AICPA Proposes Update of SOC 2® Trust Services Criteria to Enhance the Relevance, Usability and Value of CPAs’ Reports](https://www.aicpa-cima.com/news/article/aicpa-proposes-update-of-soc-2-r-trust-services-criteria-to-enhance-the-relevance-usability-and-value-of-cpas-reports) — AICPA & CIMA, accessed 2026-11-15.
2. [Art. 28 — Processor, Regulation (EU) 2016/679 (GDPR)](https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX:32016R0679#art_28) — EUR-Lex, accessed 2026-11-15.