Your support team already has the answer. SquirrelOps finds it.

It works inside the support conversation

SquirrelOps appears in the Intercom conversation sidebar. Nobody opens a separate AI app or pastes the customer's message into a chatbot.

• The initial customer report and every later message
• Support replies and the latest unanswered question
• Customer screenshots, page and referral context
• Company or environment clues
• Previous Squirrel findings and the current state of the conversation

A gatekeeper decides whether investigation is worthwhile

Not every incoming message should start an expensive investigation. Intent is evaluated first, and simple cases can be handled directly.

• Thank-you messages and acknowledgements
• Automatic emails, delivery and postmaster notices, system boilerplate
• Internal coordination and informational messages with no request
• Repeated facts and updates that do not change the existing finding
• Messages from explicitly excluded system accounts

Support history and knowledge

It searches previous support conversations and reads public and internal documentation.

• Similar symptoms, previous questions, known resolutions
• Workarounds and recurring problems
• Cases in the same product area
• Internal instructions inform Support without being exposed to the customer

Jira

It searches your tracker and reads the real state of the work.

• Known bugs, related development work, duplicates, enhancement requests
• Status, sprint, fix versions, resolution details, release context
• Whether work is queued, active, rejected, duplicated, completed, or deployed
• An open issue with no fix version still proves the company knows about the problem

Source code

It can inspect a private repository so the diagnosis is checked against real product behavior.

• Search exact error messages and locate the affected feature
• Trace validation and error-handling paths
• Inspect relevant tests and recent changes
• Code is evidence only — implementation detail never reaches customer-facing text

Logs and observability

It can query runtime sources such as Hikari and Datadog using imperfect clues: an email domain, a page URL, a company, an environment, an approximate time.

• Whether the reported event appears in the logs at all
• Whether the problem is isolated or recurring
• Whether multiple customers are affected
• Whether a service or integration failed
• Conclusions and safe references, never raw log dumps

Deployments and maintenance

It can check deployment systems such as Octopus Deploy before an outage is called a product defect.

• Whether a deployment is active or maintenance is scheduled
• Whether the environment recently changed
• Whether a Jira fix actually reached the customer's environment

Screenshots and attachments

Customer images are treated as evidence, not decoration. Email-signature graphics are excluded.

• Incorrect values and visible error messages
• Unexpected formatting and the relevant product page
• Environmental clues
• A mismatch between the written report and what the screenshot shows

Evidence-backed findings with calculated confidence

Results are classified, explained, and scored by the application — not by the model asserting a percentage.

• Known solution, existing or duplicate issue, related issue, possible regression
• New issue candidate, more information required, no support action required
• What is happening, why it is supported, what Support should do next, what is missing
• Which evidence sources were used and which Jira issues are relevant
• Plain ASD-STE100 style language: short sentences, consistent terms, minimal jargon

Pointed questions when evidence is incomplete

Low confidence produces the smallest set of questions that would move the case forward — not a vague reply.

• Which page were you on? What time did the error occur?
• Does this happen for one record or all records?
• What result did you expect? Can you send a screenshot?
• If asking will not help, it prepares an internal escalation instead

Guarded internal notes

For strong enough results, SquirrelOps adds a teammate-only note through an approved support identity.

• The complete finding in concise support language, with safe Jira links
• Posted once — no repetitive progress updates
• Held when the evidence does not meet the configured policy
• A separate note carries the suggested customer reply
• Automatic public replies are not currently enabled

Structured Jira tickets, on a teammate's request

A teammate chooses Create bug or Create feature request in the Squirrel Ticket Canvas.

• Uses the whole conversation, attachments, the finished investigation, and existing Jira matches
• Includes customer impact, Intercom priority, and the teammate's identity
• Creates a standardized Bug or Story, links it to the conversation both ways
• Prevents duplicate creation and posts a note recording what was created

Multi-tenant SaaS control plane

An isolated workspace per organization holding users and roles, connected sources, agent configuration, safety policies, scoring configuration, automation rules, usage and billing, audit records, and retention settings. Tenant data, credentials, memory, embeddings, and AI usage are designed to be isolated by default, with no company-specific assumptions hard-coded.

Self-service installation and onboarding

Create an organization, install the help-desk app, authorize your tracker, connect a code host and documentation, optionally connect logs, observability, and deployments, choose exactly which repositories, projects, environments, and indexes may be read, run a permissions test, complete a historical evaluation, and enable shadow mode. Onboarding is designed to show what each connector can read and what it can modify.

A connector framework

Expansion beyond the initial integrations, with read and write permissions enforced separately.

• Support: Intercom, Zendesk, Salesforce Service Cloud, HubSpot Service Hub, Freshdesk
• Engineering: Jira, Linear, GitHub Issues, Azure DevOps
• Code: GitHub, GitLab, Bitbucket, Azure Repos
• Knowledge: Intercom Knowledge, Confluence, Notion, SharePoint, GitBook, Google Drive, internal doc APIs
• Runtime: Datadog, Sentry, Splunk, Elastic, Grafana/Loki, CloudWatch, New Relic, customer MCP or API endpoints
• Deployment and incident: Octopus Deploy, GitHub Actions, Azure DevOps Pipelines, Argo CD, Kubernetes, Statuspage, PagerDuty

Configurable investigation planner

Administrators map product areas to repositories, choose relevant Jira projects, allow specific log indexes and services, bind environments to customers, mark authoritative sources, decide when code inspection or a deployment check is required, tune gatekeeper exclusions, and flag which investigations need human review — so only relevant sources are queried.

Continuous evaluation and quality management

Squirrel-versus-support resolution comparisons, gatekeeper false positives and negatives, missed known-issue matches, incorrect Jira associations, draft acceptance and edit distance, follow-up question effectiveness, citation accuracy, confidence calibration, safety-rewrite frequency, and cases with no useful answer — plus testing a new model, prompt, scoring version, or memory policy against historical cases before deployment.

Administrative review console

Inspect investigations and the evidence chain, review tool calls and failures, see why the gatekeeper allowed or blocked a case, compare Squirrel's answer with the final resolution, correct or retire a memory, approve automation policies, review safety rewrites, re-run an investigation, check connector health, monitor queues and retries, and export audit records.

Customer-request lifecycle management

Connect the conversation, a customer-facing ticket, an optional internal tracker, the canonical Jira issue, deployment verification, and the final customer message — translating engineering states into simple customer states such as Request received, Under review, Scheduled, In progress, Awaiting your reply, Already tracked, Update available, Resolved, or Not scheduled. Only meaningful milestones would propose an update.

Incident and pattern detection

Cluster similar symptoms across conversations, compare environments, identify shared recent deployments, watch log volume, find a common Jira or code path, recommend a shared incident or tracker, show how many customers may be affected, and coordinate safe updates without exposing one customer to another.

Governed customer-facing automation

Safe after-hours support runs through graduated authorization levels, not a single switch: research only, internal notes, suggested replies, automatic follow-up questions, automatic replies for verified known solutions, automatic incident acknowledgements, and human escalation when uncertainty remains. Each policy can depend on evidence score, classification, time of day, customer tier, product area, required sources, active deployments, and whether privileged or regulated information is involved.

Enterprise security and compliance

Encryption in transit and at rest, tenant-isolated credential vaults, least-privilege OAuth scopes, read-only connector modes, role-based access control, SSO and SCIM, complete audit trails, configurable retention and deletion, regional data residency, customer-controlled model providers, subprocessor documentation, secret detection, prompt-injection defenses, PII controls, and SOC 2 readiness followed by further compliance programs.

Enterprise reliability

Durable queues, idempotent operations, session recovery after deployments, bounded retries with backoff and jitter, dead-letter handling, connector-health monitoring, status reporting, safe partial results, duplicate-write prevention, service-level objectives, and replayable audit events. A GitHub or Datadog outage should not make the support workspace unusable — SquirrelOps is designed to state which evidence was unavailable and continue.

Cost and value analytics

AI usage is already attributed to individual investigations and conversations. Reporting covers cost per ticket and per resolved ticket, cost by model and connector, cost avoided by gatekeeper decisions, investigation duration, handling time saved, escalations and engineering interruptions avoided, draft acceptance, reopened-ticket rate, known-solution reuse, and cost-versus-accuracy comparisons between models.

Model and agent independence

The durable product is the control plane: context gathering, evidence routing, source verification, confidence calculation, safety review, memory, workflow actions, auditing, evaluation, and cost control. Different models handle investigation, gatekeeping, safety review, summarization, and drafting, and customers can select approved providers or bring their own endpoint.