Calibration
What we will not claim about our own product
Below is the list of sentences we have decided not to write, and the ones we write instead. It is normally an internal document. We publish it because our product reports its own measured accuracy inside the console, so overstated marketing would be contradicted by our own software within a week — and because the failure mode we are most afraid of is not a wrong answer. It is a customer who quietly learns to ignore us.
A small company cannot run enterprise-grade security without a specialised team. Our target is 60–75% of what a competent analyst would catch on known attack patterns — with a full audit trail, at a fraction of the cost — because the product does the setup and bounds its own autonomy.
The list
Never, and instead
Left column: what a competitor in this category will say, and what we have decided not to. Right column: what we say instead, which is narrower and checkable.
We do not say
Autonomous SOC. AI SOC. Replaces your security team.
We say
Autonomous Tier 1. You keep Tier 2 and Tier 3, and we tell you when you have one.
We do not say
24/7 SOC, or anything implying a human is watching your estate.
We say
Sentinel-Q investigates every alert within minutes, around the clock. There is no human on call unless you have your own.
We do not say
Detects zero-days, novel threats, or sophisticated targeted attacks.
We say
Strong on known attack patterns, with published coverage against MITRE ATT&CK. Novel attacks are what we escalate to you.
We do not say
Zero, or near-zero, false positives.
We say
The measured false-positive rate, published. Then it is a fact rather than a claim.
We do not say
You will not be breached. Any promise of a security outcome.
We say
Every alert investigated within minutes, with cited evidence.
We do not say
Fully automated remediation. Self-healing.
We say
Reversible actions inside an envelope you set, per action class, in writing, with an undo control on every one.
We do not say
NIS2 compliant. GDPR compliant. DORA compliant.
We say
Produces the evidence your NIS2 audit asks for: Article 21 measures, the 24-hour early warning drafted from the incident record, and access certification records. Software cannot make an organisation compliant.
We do not say
EU data residency, unqualified.
We say
Your organisation is bound to a home region and a request from another region is refused rather than served. AI inference is a separate question and we report it as unverified where our provider offers no regional guarantee.
We do not say
700+ integrations.
We say
A catalogue with available, beta or planned marked on every row. A short true list beats a long false one.
We do not say
A bare confidence percentage — “87% malicious”.
We say
A qualitative band with its reason: “Low — could not corroborate sign-in geography, source offline.” Practitioners read a bare percentage as theatre.
We do not say
No configuration required.
We say
We propose the configuration in plain language, with reasons. You approve it.
We do not say
AI role mining finds your roles.
We say
We run access certification campaigns over the roles you already have, and record who approved what. Deriving a role structure from observed access now exists — as proposals: Identity → Role mining analyses memberships and application assignments and drafts bundles, with the support figures shown so you can argue with the thresholds. It never applies one; acceptance is a human decision through the ordinary access-request flow, and a competitor's demo is not evidence that theirs applies roles without one either.
The limits
Eight things it cannot do
Not roadmap items we are being coy about. Most of these are properties of the problem, and the two that are not are deliberate architectural refusals.
The accuracy band, with its source
- Frontier language models
- 61–67%
- Purpose-built security agent
- 72%
- Where we aim to sit — our target, not a measured result
- 60–75%
- Strong human analysts
- 73–85%
Task completion on 100 full kill-chain scenarios in the Simbian AI SOC benchmark. A frontier model today is a mediocre-to-decent junior analyst: clearly better than nobody, clearly worse than a good human.
- A skilled attacker targeting you specifically
- Detection works on known patterns. A patient adversary who studies your business and uses no recognisable technique is outside what any product at this price detects. We escalate; we do not pretend.
- Anything your Microsoft licence does not emit
- Sign-in log analysis and Conditional Access evaluation need Entra ID P1. Where the data is unavailable the affected checks are reported as skipped, never as passed — a licence limitation must not become a false all-clear.
- The first fortnight of behavioural detection
- Learning what is normal for your business takes five to fourteen days of your real traffic. That is a property of the data, not the software, and nobody can compress it.
- Your business trade-offs
- Turning off a legacy sign-in protocol is a large security win that will break a printer and one laptop. We will tell you which ones by name. We will not decide for you.
- Deploying software to your endpoints
- The sensor has to be installed. If you cannot roll out software to your machines, the endpoint half of the platform cannot help you.
- Being a human on the phone at 3am
- Some vendors sell people; we sell software. If you want a specialist to call, that is a real thing to want, and we will tell you where to buy it rather than implying we are it.
- Granting privileged access
- Never automated, at any tier, at any confidence. The capability is not exposed to the model at all, so it cannot be talked into it.
- Rewriting its own rules
- The rung above acting — software that authors detections and adjusts its own limits — is not built. Rule-writing autonomy without versioning and rollback quietly degrades coverage, and you would find out from an intrusion.
Open questions
What we do not yet know
Things we have not resolved, listed here rather than left for you to discover during procurement.
EU-resident AI inference
Our model provider does not offer a contractual guarantee that inference runs inside the EU. Until it does, the AI system card reports the inference region as unverified and we do not sell the residency add-on. If the guarantee never arrives, we will remove the sovereignty positioning rather than soften it — a hedged residency claim in a NIS2-positioned product reads worse to a compliance buyer than no claim at all.
Who handles the escalated 40%
Sentinel-Q does roughly the Tier 1 half of the job. The rest needs a human specialist. Today that human is you or your MSP. We have priced a partner-delivered escalation service and will not pre-sell it before it exists.
How cyber insurers treat automated response
Whether an insurer's position changes when software is permitted to isolate a machine is a question we have asked and not yet had answered in writing. We would rather tell you that than let you assume it has been settled.
If you find a claim anywhere on this site that does not survive you checking it, tell us: [email protected]. We will correct the page and say what changed.
Our own security posture, datedFind out what an attacker would see
Connect Microsoft 365 and read your own posture report. 82 checks against your real tenant, naming your real administrators, applications and devices. Free, complete, and yours to keep — not a sample with the findings blurred.
No card. No sales call before you have read the report.