Stack Innovations / Services / AI & Automation / Product Copilots
Product Copilots · AI & Automation

A copilot, in your product.
Answers and actions.

Not a chatbot bolted on the side — a copilot that lives inside your app, answers from your own data with citations, and fires real, structured actions: create the record, update the status, jump the user to the right screen. Grounded, guard-railed, and shipped to production.

/01Drag the sessions · watch the copilot earn its keep
Live · in-product copilot, embedded
Deflection 61%
"Create an invoice for the Northwind account"
Intents handled0
Actions taken0
Deflection0%
CSAT0
Users · monthly sessions 8,000 sessions
Intent-resolution rate89%
Support tickets deflected−61%
Actions taken in-app12M
Median time-to-value1.4s
Trusted by teams shipping copilots inside their products at
02 — Outcomes

Copilots that did the work.

A ledger of named copilots where the line that moved was a question answered or an action taken — deflection, time-to-value, tasks completed without a human in the loop. 6 of 30 shown · ledger updates as copilots scale.

Northwind SaaS
In-app copilot · B2B platform
Answers from the user's own workspace and fires structured actions — creates records, updates status, navigates — so users stop hunting through menus
−54%Support tickets
Cobalt Analytics
Analytics copilot · Dashboards
Plain-English asks compile to queries, the copilot returns the chart and cites the exact metric definition it used — grounded, not guessed
3.4×Self-serve reports
Vera Health
Clinical copilot · EHR workflows
Drafts notes and order sets grounded in chart data, gating every side-effectful action behind explicit clinician approval — never auto-submits
−43%Documentation time
Lumen Support
Support copilot · Help + product
Reads the user's account state, answers with cited docs, and completes the fix in-product instead of writing a how-to the user must follow
67%First-contact resolution
Drift Commerce
Merchant copilot · Storefront admin
Bulk edits, campaign setup, and catalog updates expressed in one sentence — structured tool calls preview the change before it commits
−71%Time-to-task
Forge Onboarding
Activation copilot · New-user flow
Walks new users through setup by doing it with them — connects sources, configures defaults, and only asks when an action is irreversible
+58%Activation rate
03 — The copilot, live

It doesn't just answer.
It does.

Pick what a user might ask inside the product and watch the copilot respond — streaming a grounded, cited answer and firing a structured action that executes in the mock app on the left. Flip to actions-only to see the work happen without the chatter; turn grounding off and the answer drifts to generic, uncited guesswork.

Northwind · Workspace
copilot · live
Invoices
4 open · sorted by due date
INV-2048 · Cobalt Analytics
Due in 6 days
Sent
INV-2047 · Lumen Docs
Due in 2 days
Overdue soon
Northwind account
No open invoice
Idle
INV-2045 · Drift Finance
Paid
Closed
Copilot · grounded answer
+ create_invoice · Northwind account
Try an in-product ask
Mode · chat vs. actionsChat + actions
Intent
Action
Grounded
A copilot earns trust two ways at once: it answers from your app's real data with citations, and it turns intent into a structured tool call the product can execute and preview — not free text a human still has to act on.
04 — Anatomy of the copilot

A capability map,
not a magic box.

Every surface where the copilot can help, every action it's allowed to take, the tool that backs it, and the guard that keeps it safe. This is the room we work in: each capability scoped, grounded, and gated — so the copilot is useful without ever being a liability.

Copilot capability map · Northwind Workspace
Intents 24 mapped · Deflection 61% · Action success 97%
SurfaceAction / toolWhat it doesGuard
AskStreaming · ReactCapture the user's intent in-context, with the current screen and selectionscoped
Retrievepgvector · PostgresPull the user's own records and docs they're allowed to see — tenant-filteredRLS
ReasonClaude Opus 4.8 · 1M ctxDecide: answer, call a tool, or ask a clarifying question — grounded in retrieved contextlive
Tool callStrict tool use · MCPEmit a structured, schema-valid action (create / update / navigate) — never free textstrict
PreviewStructured outputsRender the proposed change in the UI before anything commits to the databasedry-run
ExecuteFastAPI · app APIRun the action against the real product API once confirmed — idempotent, auditedgated
CiteAnthropic CitationsAttach the exact app records and docs the answer was grounded in, clickable in-appcited
EvaluatePostHog · LLM-as-judgeScore intent resolution, action success & grounding on a frozen set each release97%
blue measured & in target
live the stage running in the demo above
amber gated · human approves side-effects
01
05 — Ship to production

Map the intents.

Before a single tool is wired, we pin down what users actually want to do inside your product — the top asks, the actions worth automating, and the ones that must stay behind a human. Then we build the eval set: the intents we'll grade every release against.

/ Week 00 · Intents & eval set
Intents80 real asks pulled from support, search, and session replays
ActionsWhich intents map to a safe, reversible structured tool call
Eval setGolden intent → expected answer / action pairs, frozen for grading
GuardsSide-effectful actions gated behind preview + explicit confirm

Ground in app data.

A copilot that doesn't read your data invents answers. We wire retrieval over the user's own records and docs — tenant-scoped and access-aware — so every answer is grounded in what this user is actually allowed to see, with citations back to the source.

/ Week 01 · Grounding
SourcesApp records · docs · knowledge base · settings
Retrievalpgvector hybrid · semantic + keyword over user data
AccessRow-level security · tenant + role filters at query time
CitationsAnthropic Citations · clickable back to the in-app record
CachingPrompt caching for stable context · cheaper, faster

Wire the actions.

Answers are half the job. We define the copilot's tools as strict, schema-validated functions — create, update, navigate — so the model emits a structured action your product can execute, not free text a human has to interpret and re-type.

/ Week 02 · Tool use
Tools · strict tool use · JSON-schema validated
Structured outputs · output_config.format enforced
MCP · product API exposed as governed tools
Preview · dry-run the change before it commits
Idempotency keys · safe retries on every action

Gate the irreversible.

The copilot can read freely. Writing is different. Anything side-effectful — sending, deleting, charging — pauses for an explicit human confirm with the proposed change shown in the UI. Reversible actions flow; irreversible ones wait.

/ Week 03 · Guards & gating
Look up an invoiceautoread
Draft a replyautopreview
Create a recordpreviewconfirm
Send an emailgatehuman
Charge a cardgatehuman

Evaluate honestly.

Run the frozen intent set every change and score it — intent resolution, action success, grounding, refusal correctness. No "looks good in the demo." A number that moves, or the change doesn't ship.

/ Week 04 · Evaluate
Intent res.89% — the right answer or action was chosen
Action success97% — tool calls executed without error
Grounding94% — answers supported by cited app data
Abstain rate5% — asked rather than guessed

Ship & monitor.

Live with caching for cost, streaming for feel, and logging on every intent and action. We watch resolution and action-success as usage grows, catch regressions before users do, and keep the copilot useful as your product changes underneath it.

/ Ongoing · Ship & monitor
Prompt caching · context
Streaming responses
Action audit log
PII + access guards
Citations on by default
Confirm on side-effects
Weekly eval run
Human-in-the-loop review
06 — Why it compounds

A used copilot gets sharper.

Every session is signal: failed intents sharpen the prompts, missed actions justify new tools, wrong answers tighten the grounding. Ship-and-forget copilots stall as your product moves on without them. Evaluated and tended, intent resolution compounds.

Eval-driven by Stack Innovations — resolution climbs as intents and tools widen
Ship-and-forget — plateaus, then drifts as the product changes around it
Representative of a typical 12-month engagement · intent-resolution rate on a frozen evaluation set.
07 — Tools · honest kit

The kit, shown.

The models, stores, and tools we actually wire together to ground a copilot in your data, let it answer with citations, and let it take structured, gated actions inside your product. No mystery framework — just the kit that makes a copilot trustworthy.

Reasoning
Claude Opus 4.8
Structured out
Output format
Actions
Tool use
Tool gateway
MCP
Citations
Anthropic Cite
Frontend
React
Streaming
SSE
Database
Postgres
Vector DB
pgvector
Serving
FastAPI
Analytics
PostHog
Comms
Slack
Start the build

Stop explaining.
Start doing. In your product.

A free copilot audit to start — bring your product, a list of real user intents, and we'll show you what an in-product copilot would answer and which actions it would take, grounded in your own data. A working prototype, not a pitch.

Get a copilot audit
Accent
Hero shader
Motion