Signature-based antivirus and firewalls catch what they already know about. AI-driven detection works differently: it learns what normal looks like for your users, devices, and traffic, then flags deviations as they happen. The stakes for small firms are real. The Verizon 2025 Data Breach Investigations Report found ransomware in 44% of the breaches it analyzed, and in 88% of breaches at small and medium businesses, and IBM's 2025 Cost of a Data Breach Report put the average time to identify and contain a breach at 241 days. For a business with 10 to 100 employees and no dedicated security analyst, behavioral detection means a better chance of catching a compromised Microsoft 365 account, a new strain of ransomware, or an attacker moving between machines before data is encrypted or stolen. This guide walks through a practical rollout using a fictional 40-person accounting firm, Brightline Accounting, as the worked example. It follows the Detect function of the NIST Cybersecurity Framework 2.0 (categories DE.CM, continuous monitoring, and DE.AE, adverse event analysis).
Audit Your Current Detection Gaps and Missed Threats
Before buying anything, find out what you are currently missing. Review the last 12 months of incidents, near misses, and user reports. Brightline found three things: two employees had clicked phishing links that nobody detected until the employees reported them, one shared mailbox had a forwarding rule to an external Gmail address that sat unnoticed for five weeks (the technique MITRE ATT&CK catalogs as T1114.003, Email Forwarding Rule), and their antivirus console had 600 unread alerts.
Write each gap as a question the new tool must answer. Use a simple table or list like this:
- Can we detect a login from a new country within 5 minutes? Currently no.
- Can we detect a mailbox forwarding rule created by an attacker? Currently no.
- Can we see a workstation connecting to an unfamiliar server on port 445 (SMB)? Currently no.
- Can we triage alerts in under 15 minutes a day? Currently no, 600 unread.
For the forwarding rule example, you can check Microsoft 365 right now with PowerShell. This is a quick way to confirm whether you already have a gap. Note that it only finds mailbox-level forwarding; inbox rules that forward mail are a separate check (Get-InboxRule per mailbox):
Connect-ExchangeOnline -UserPrincipalName admin@brightlineacct.com
Get-Mailbox -ResultSize Unlimited | Where-Object {
$_.ForwardingSmtpAddress -ne $null -or $_.ForwardingAddress -ne $null
} | Select-Object DisplayName, ForwardingSmtpAddress, DeliverToMailboxAndForward
# Output
# DisplayName ForwardingSmtpAddress DeliverToMailboxAndForward
# Billing Team smtp:brightline.bill@gmail.com True
Inventory Assets and Log Sources to Feed the AI
AI models are only as good as the data they see. A tool that sees only laptops will miss a compromised cloud account. List every device, identity system, and service, and note whether it produces logs and where they go. NIST SP 800-92, the Guide to Computer Security Log Management, treats this inventory of log sources as the first step of any log program. Brightline's inventory looked like this:
- 42 Windows 11 laptops and 3 Macs (endpoint telemetry)
- Microsoft 365 with Entra ID sign-ins and Exchange (email and identity logs)
- 1 Fortinet FortiGate 60F firewall (network logs, syslog)
- 1 Synology NAS holding client tax files (file access logs)
- QuickBooks Online and a client portal (application logs)
- 2 Ubiquiti access points (Wi-Fi client logs)
Confirm that the logs you need are actually turned on. In Microsoft 365, unified audit logging is the foundation for most identity and email detections. It is on by default for most newer tenants, but older tenants and some configurations may differ, so verify it, and enable it if needed, with:
Get-AdminAuditLogConfig | Format-List UnifiedAuditLogIngestionEnabled
# If False:
Enable-OrganizationCustomization # only if prompted
Set-AdminAuditLogConfig -UnifiedAuditLogIngestionEnabled $true
Also check log retention. Microsoft Purview Audit (Standard) keeps records for 180 days, but many devices keep only 7 to 30 days locally, which is too short to investigate an intrusion that began six weeks ago. Aim for at least 90 days searchable and 12 months archived. That mirrors PCI DSS requirement 10.5.1 (12 months of retention, with the most recent 3 months immediately available) and is a sensible benchmark even if you do not process card data.
Choose an AI-Driven Detection Tool That Fits Your Size and Budget
Small businesses generally choose among three paths. Match the option to how many hours per week someone can realistically watch alerts.
- EDR/XDR platform (for example Microsoft Defender for Business, included with Microsoft 365 Business Premium at roughly $22 per user per month and also sold standalone, or CrowdStrike Falcon Go, SentinelOne Singularity). Best if you have an IT person or MSP who can review alerts daily.
- Managed detection and response (MDR) service (for example Huntress, Arctic Wolf, Blackpoint). A vendor's analysts watch AI-prioritized alerts 24/7 and act on them. Best if nobody on staff can respond at 2 a.m. Expect roughly $8 to $20 per endpoint per month, though pricing varies by vendor and contract.
- AI-enhanced email security and SIEM features (for example Defender for Office 365, Abnormal Security, or Microsoft Sentinel). Best as an add-on when phishing and account takeover are your top risks.
Brightline already paid for Microsoft 365 Business Premium, so Defender for Business and Entra ID P1 Conditional Access were available at no extra cost. Risk-based policies driven by Entra ID Protection need Entra ID P2, which they did not license. Because no one at the firm could watch alerts overnight, they added an MDR service on top for about $9 per endpoint per month (45 endpoints, about $405 monthly). The timing matters: Sophos's Active Adversary reports have found that the large majority of ransomware payloads are launched outside local business hours. Ask any vendor four questions: what data sources it ingests, whether it works without a dedicated analyst, how long the learning period lasts, and whether you can export your data if you leave. Independent MITRE ATT&CK Evaluations results are a useful, public way to compare how products detect specific techniques.
Deploy Sensors and Connect Endpoint, Email, Cloud, and Network Data
Start with a pilot group of 5 devices, including one executive laptop and one machine that handles sensitive files, then roll out to everyone within two weeks. For Windows machines managed through Intune, onboarding Defender for Business is automatic once the policy is assigned. For a vendor agent installed manually, a silent install looks like this (substitute your own tenant key):
# Example silent install of an MDR agent on Windows
msiexec /i "C:\Deploy\agent.msi" /qn /norestart ACCOUNT_KEY="a1b2c3d4e5f6" ORG_KEY="brightline-acct"
# Verify the service is running
Get-Service -Name "HuntressAgent" | Select-Object Name, Status
# Name Status
# HuntressAgent Running
Next, connect the non-endpoint sources. Point the firewall's syslog output to your collector or SIEM. On a FortiGate, this takes a short block of FortiOS CLI commands:
config log syslogd setting
set status enable
set server 10.0.10.25
set port 514
set format default
end
Then connect Microsoft 365 and Entra ID through the vendor's cloud connector (usually an OAuth consent by a Global Administrator). Use a dedicated service account with read-only permissions where the vendor allows it, and record which permissions you granted. After deployment, confirm each source is reporting. Brightline's dashboard check showed 45 of 45 endpoints, Microsoft 365, and the firewall reporting, but the NAS was silent, so they enabled its syslog forwarding the same day.
Run a Learning Period to Baseline Normal Behavior
Behavioral models need time to learn patterns such as when each user logs in, which servers each laptop talks to, how much data leaves the network on a typical Tuesday, and which applications normally run. Plan for 14 to 30 days, and check your vendor's documentation, since learning periods differ by product. Monthly cycles matter: Brightline's accountants upload far more data at quarter-end, so a two-week baseline in mid-month would later flag normal behavior as suspicious.
During this period, keep the tool in detect-only or audit mode so it does not block legitimate work. Document known unusual activity so it does not poison the baseline: a penetration test, a bulk data migration, or an employee traveling to Spain for three weeks. Do not let a known-compromised machine into the baseline either. Scan and clean everything before you start, or the model may treat attacker behavior as normal.
Tune Alert Thresholds and Reduce False Positives
In the first weeks, expect noise. Brightline's tool generated 85 alerts in week one. Reviewing them, 61 came from three sources: a backup agent that scanned all files at 1 a.m., a remote support tool used by their IT contractor, and a tax software updater. Adding those to an allowlist cut week-two alerts to 22.
Tune carefully. Allowlist a specific signed binary or a specific IP, never an entire folder or a broad category. Microsoft's own guidance for Defender Antivirus warns that exclusions reduce protection and should be kept to a minimum. Here is how to review and add a process exclusion in PowerShell, and why a narrow one is better than a broad one:
# Review current exclusions
Get-MpPreference | Select-Object -ExpandProperty ExclusionPath
# Narrow exclusion for a specific backup agent executable
Add-MpPreference -ExclusionProcess "C:\Program Files\VeeamAgent\veeamagent.exe"
# Avoid broad exclusions like this one:
# Add-MpPreference -ExclusionPath "C:\"
Keep a tuning log with the date, alert name, reason for the change, and who approved it. Tuning decisions are often forgotten, and an old exclusion is a common place for attackers to hide.
Automate Response Actions for High-Confidence Threats
Once false positives are under control, let the tool act on its own for situations where speed matters and the confidence is high. Ransomware is typically launched when no one is watching (see the Sophos finding above), and the behaviors that precede encryption, such as shadow copy deletion (MITRE T1490) and mass file encryption (T1486), are exactly what EDR tools are built to recognize. Reasonable automation for a small business includes:
- Isolate an endpoint from the network when ransomware behavior is detected (mass file renaming, shadow copy deletion).
- Disable a user account and revoke sessions when sign-in risk is high, such as impossible travel combined with a new device and an MFA fatigue pattern.
- Quarantine a phishing email from all mailboxes after one user reports it and the tool confirms it is malicious.
- Block a file hash or domain across all devices after it is confirmed malicious on one.
With Entra ID P2, an Entra ID Protection risk-based Conditional Access policy can require a password change or block access automatically at high user risk. On Business Premium's P1 licensing, a Conditional Access policy that requires MFA or blocks sign-ins from countries where you have no staff is a practical substitute. Keep your break-glass admin accounts excluded from automated blocks and store their credentials offline, as Microsoft's emergency access account guidance recommends, so a false positive cannot lock you out of your own tenant. Start with automation for the narrowest, highest-confidence scenarios and expand only after reviewing a month of results.
Create a Human Review and Escalation Workflow
AI prioritizes and acts, but a person owns the outcome. Decide in advance who reviews alerts, how fast, and what happens when they are unavailable. NIST SP 800-61 Revision 3 on incident response and the IRS Publication 4557 guidance for tax professionals both expect a documented plan with named roles. Brightline wrote a one-page workflow:
- Critical (ransomware, confirmed account takeover): MDR calls the office manager and IT contractor immediately. Response target: 15 minutes, any hour.
- High (suspicious sign-in, malware blocked but spreading attempt): reviewed by the IT contractor within 4 business hours.
- Medium (unusual file access, new device sign-in): reviewed in a 15-minute daily check by the office manager.
- Low (informational): reviewed weekly, closed in bulk if no pattern appears.
Name a backup person for each role and keep a printed contact sheet with the MDR phone number, cyber insurance hotline, and your attorney. For each alert, a reviewer should record three things: what happened, whether it was a true or false positive, and what was done. Those records feed back into tuning and give you evidence for insurers and auditors. For firms like Brightline that prepare tax returns, the FTC Safeguards Rule (16 CFR 314.4) also requires continuous monitoring, or annual penetration testing plus vulnerability scans every six months, and a written incident response plan, so these records double as compliance evidence.
Test Detection with Simulated Attacks
Do not assume the tool works. Test it with safe, controlled simulations and confirm that an alert fires and the right person is notified. Always get written approval, tell your MDR provider beforehand so they are not surprised, and run tests on a pilot machine.
The EICAR test file is a harmless 68-byte string, published by the European Institute for Computer Antivirus Research, that every antivirus and EDR product should treat as malware. Atomic Red Team, an open-source library from Red Canary, provides small, documented tests mapped to the MITRE ATT&CK framework. For example, this test creates a scheduled task for persistence (technique T1053.005):
# 1. EICAR test (harmless; should be blocked immediately)
$eicar = 'X5O!P%@AP[4\PZX54(P^)7CC)7}$EICAR-STANDARD-ANTIVIRUS-TEST-FILE!$H+H*'
Set-Content -Path "$env:TEMP\eicar.txt" -Value $eicar
# 2. Atomic Red Team: scheduled task persistence (MITRE T1053.005)
Install-Module -Name invoke-atomicredteam -Scope CurrentUser
Import-Module invoke-atomicredteam
Invoke-AtomicTest T1053.005 -TestNumbers 1
# 3. Clean up afterwards
Invoke-AtomicTest T1053.005 -TestNumbers 1 -Cleanup
Also test the non-technical paths. Send a simulated phishing email using Microsoft Attack Simulation Training (included with Defender for Office 365 Plan 2) and see whether the AI-based email filter catches it. In Brightline's first test, the EICAR file was blocked in 3 seconds, but the scheduled task test generated no alert. The vendor traced it to a policy set to audit-only on the pilot group, which they corrected the same day. This is exactly the kind of gap a test is meant to find.
Monitor Performance Metrics and Retrain or Adjust Regularly
Review a small set of metrics monthly to confirm the tool is earning its cost. Compare your numbers with the 241-day industry average to identify and contain a breach reported by IBM; a small firm with MDR coverage should be measuring in minutes and hours. Useful measures include:
- Mean time to detect (MTTD) and mean time to respond (MTTR). Brightline's target: under 10 minutes to detect and under 60 minutes to contain.
- False positive rate: the percentage of alerts closed as benign. Brightline treats anything above roughly 50 percent as a signal to review tuning; this is an internal rule of thumb, not an industry standard.
- Coverage: percentage of endpoints and log sources actively reporting. Target 100 percent, and investigate any device silent for more than 24 hours.
- Alert volume per week and the number of alerts that required human action.
- Results of the last simulated attack test, ideally tracked as the share of tested ATT&CK techniques that produced an alert.
Models drift when your business changes. A new office, a switch to remote work, a new line-of-business app, or a seasonal busy period can all make old baselines inaccurate. Re-run the learning review after major changes, repeat your simulated attack tests every quarter, and review your allowlist twice a year to remove entries for software you no longer use. Finally, remember that AI detection is one layer. It works best alongside multi-factor authentication, prompt patching, tested offline backups, and a written incident response plan that your team has practiced.
Sources and Further Reading
- Verizon, 2025 Data Breach Investigations Report (ransomware prevalence, small business findings).
- IBM Security, Cost of a Data Breach Report 2025 (time to identify and contain).
- NIST Cybersecurity Framework 2.0, Detect function (DE.CM, DE.AE).
- NIST SP 800-92, Guide to Computer Security Log Management; NIST SP 800-61 Rev. 3, Incident Response Recommendations.
- MITRE ATT&CK techniques T1114.003, T1053.005, T1486, T1490, and MITRE ATT&CK Evaluations.
- Microsoft Learn: Purview Audit (Standard) retention, Defender Antivirus exclusion guidance, Entra ID licensing, and emergency access account guidance.
- PCI DSS v4.0, requirement 10.5.1 (audit log retention).
- FTC Safeguards Rule, 16 CFR Part 314; IRS Publication 4557, Safeguarding Taxpayer Data.
- Sophos, Active Adversary Report (timing of ransomware deployment).
- Red Canary, Atomic Red Team documentation; EICAR anti-malware test file specification.