Togetherwork Ops AI

Operational work spans customer support, implementations, knowledge management, managed-services automation and reporting. The prototype explores how AI could reduce repetitive analysis and surface next actions across those workflows.

• Customer support
• Implementations
• Knowledge management
• Managed services
• Reporting

1. 1

   Work is fragmented

   Support, implementations, knowledge, managed services and reporting each live in a different workflow.

2. 2

   Analysis is repetitive

   Reading a ticket, drafting a reply, building an onboarding checklist or spotting a documentation gap is manual and repeated.

3. 3

   One AI-assisted prototype

   Five pages, one per workstream, share a single demo dataset and a single AI layer.

4. 4

   Next actions surfaced

   Each page turns its input into a structured result a person can act on.

An independent prototype built for demonstration. All data is synthetic, it makes no claim about real usage or efficiency gains, and it is not affiliated with or endorsed by Togetherwork.

Five operational roles, grouped by the workstream they own. Each card lists only what the prototype gives that role.

• Support operationsTriage tickets by category, sentiment, urgency and churn risk, review a routing decision with a confidence score, and edit an AI-drafted reply before approving or escalating.
• Implementation and professional servicesTrack client onboardings by stage and risk, and expand a row for a tailored checklist, blockers, configuration notes and a next action.
• Knowledge and documentation ownersBrowse published articles and see documentation gaps ranked by how many tickets point at them.
• Managed services and operationsWatch a queue of recurring billing, payroll and reconciliation tasks, see exceptions called out, and run a simulated workflow.
• Product and operations leadersRead cross-page metrics: ticket volume, deflection, sentiment and category mix, and a modeled hours-saved figure.

Screenshots of each appear in the walkthrough below.

• SupportImplemented353 seeded tickets with search and status, brand, vertical and priority filters. Selecting a ticket runs AI triage and drafts a reply with knowledge-base suggestions; Approve & Send and Escalate update the ticket in the session.
• ImplementationsImplementedEight client onboardings (seven active) with metric cards that act as filters, a five-stage progress track and full-row expand. One AI call per implementation produces a checklist, blockers, a configuration document and a next action.
• KnowledgePartialPublished-article metrics, search and category filters, working links, and eleven documentation gaps with draft badges. The Claude article-generation flow and coverage scores are not built yet.
• Managed servicesPartialTask metrics, status filters, exception callouts and a scheduled to running to complete workflow simulation. It is a timer-driven simulation, not an integration. Health bars and the Monday pulse are not built yet.
• AnalyticsImplementedLive counts from the same session state: AI actions, ticket deflection, sentiment and category mix. Hours saved is modeled from assumed minutes per automated unit, not measured.

The deployable architecture, not the local-only prototype. The original app called Anthropic straight from the browser with a key in the Vite bundle, which is safe only on a laptop. The groups below separate what was originally built from the changes made to host it publicly.

How one AI action runs

Every AI action runs on Claude Sonnet 4.6.

1. 1

   Select an input

   A ticket in Support, or an implementation row.

   Browser
2. 2

   Send an operation

   Only the operation name and demo data leave the browser.

   POST /api/claude
3. 3

   Build the prompt

   A server-held system prompt and a template filled from validated fields.

   Pages Function
4. 4

   Ask Claude for JSON only

   The prompt requires a single JSON object and no prose.

   Anthropic API
5. 5

   Parse resiliently

   parseModelJson() handles fenced or wrapped JSON before parsing.

   Browser
6. 6

   Render the result

   Triage fields, a drafted reply, or a checklist with blockers and a next action.

   Browser
7. 7

   A person decides

   Approve & Send or Escalate, edit the draft, tick checklist items. State changes stay in the session.

   Browser

Decisions and trade-offs

Choices visible in the code, each with what it bought and what it gave up.

Challenges taken from the code, its comments and the audit rather than recalled after the fact.

• Browser-side AI keyThe prototype called Anthropic directly from the browser with a VITE_ key and the direct-browser-access header. That is fine on a laptop and unsafe on a public site, because Vite compiles the key into the page. Public deployment required moving inference behind a server-side boundary first.
• JSON from a modelReplies are asked to be JSON only, but can arrive fenced or wrapped. A resilient parser was added, and failures surface as an error with retry that clears the cache.
• Duplicate in-flight callsSelecting a ticket or expanding a row could start the same AI call twice. In-flight guards and result caches in App state prevent repeat requests.
• A dense support inboxThe inbox is a fixed-height layout with scrolling confined to the list and the detail pane, so filters, list and detail stay visible together. It is a desktop-first layout.
• Cross-page stateAnalytics reads tickets, analyses, checklists and tasks from the same session, and Reset has to clear every AI cache, error and simulation together.
• Simulated versus generatedThe managed-services workflow is a timer, the hours-saved figure comes from assumed rates, and offline samples are rules. Each is labelled so none passes as measured or model-generated output.

Security lesson

A Vite-exposed Anthropic key suits only a local prototype. Putting the project on a public domain required moving inference behind a server-side boundary first, with validation, size and rate limits, and an AI kill switch.

Finish the pages that are only partly built first, then the production foundations the prototype skips. Items are marked Partial when some of the work exists and Planned when none does.

• Knowledge: article generation and coveragePartialGenerate a draft article for each documentation gap with Claude, store it against the gap, and add coverage scores.
• Managed services: health and pulsePartialAdd the 30-day health bars and a Monday pulse view that the build plan calls for.
• Analytics refinementsPartialFinish the analytics refinements from the plan and replace assumed savings rates with measured inputs.
• Real persistencePlannedStore tickets, drafts and decisions instead of keeping them in the browser session.
• Authentication and authorizationPlannedAdd accounts and role-based access before any real data is involved.
• Audit logging and telemetryPlannedRecord who approved or escalated what, and measure AI quality and latency.
• Real integrationsPlannedReplace the mock datasets with real support-desk, PSA, billing and payroll systems.