Your team already works in Slack. We build the AI tools that live there too: ask a question in natural language, get a sourced answer with a deep link back to the record, and approve anything the system proposes before it acts. The shipped example is a read-only copilot over Airtable and Google Sheets, built in four to eight weeks.
How it works
The bot plans the query first, fetches only the records it needs, returns a structured answer with a deep link back to the source view, and writes the whole interaction to an audit log. Read-only by design: every query is the equivalent of a SELECT.
User question
#ops in Slack
Slack bot
receives + plans the query
Claude
picks base, fields, filters
Live read
Airtable + Google Sheets
Sourced answer
inline table + deep link
Audit log + admin page
Every query, every result, response time, daily LLM token spend in dollars.
Mark answers correct or wrong; flag unanswered questions for prompt review.
The problem
Sales asks “which deals are stuck?” Finance asks “what is our AR aging?” Project managers ask “what is due this week and who owns it?” Today, every one of those questions is a Slack thread, a screenshot, and a wait.
The ops person who knows the right view is the bottleneck. The team makes decisions on stale data, or no data. And when a number gets challenged, nobody can show how it was calculated, which is when finance meetings turn awkward.
What we ship
A private channel where any team member types a question in natural language and gets a structured answer; text, an inline table or chart when the data shape calls for one, and a direct link back to the source so every number is traceable.
Connects to one Airtable workspace (multiple bases and tables) and a Google Sheets folder. Auto-discovers Airtable schema on first run and infers Sheet structure from column headers, no training, no manual mapping.
Every response includes a deep link back to the exact Airtable view or Sheet range used to compute the answer. If a number looks wrong, you can click through and verify the underlying records in seconds.
Every query is logged: the original question, the data sources hit, the answer returned, and the response time. A lightweight admin page lets the ops lead review answers, mark correct or incorrect, flag unanswered questions, and watch daily LLM cost.
Human in the loop
Read-only answers ship first, because a wrong number is embarrassing and a wrong write is expensive. When the tool starts proposing actions, every one of them waits for a person in Slack. This is the same pattern we use in every internal tool we build.
01
Someone asks
A plain-English question in a Slack channel
02
The tool proposes
An answer, or an action with its reasoning
03
A person approves
Approve or reject, in the same thread
04
It acts, and logs
Approved actions run, and all of it is logged
Phase one of the shipped copilot is strictly read-only. Write actions, such as creating a follow-up record or marking an invoice as sent, arrive in phase two behind exactly this gate, with an extended audit trail.
Real questions it answers
The bot filters Airtable deals by stage-change timestamp and returns deal name, current stage, days stuck, and owner, with a direct link to the filtered view.
Groups outstanding invoices by client and aging bucket (0–30, 31–60, 60+), returns a clean inline table, and links to the source range in Sheets.
Joins time-tracking with the client table, filters by current calendar month, sums logged hours, and compares against the contracted retainer if that field exists.
Filters the project status table by due date inside the current ISO week, returns name + owner, and flags overdue items from the previous week separately.
Filters the time tracking table by project and date range, deduplicates by team member, and returns hours logged per person with a link to the underlying records.
Substring-matches calendar event titles against known client names, applies a confidence threshold, and notes which matches were exact vs fuzzy. Low-confidence rows are surfaced separately.
If a number is challenged, the audit log shows exactly which records were summed, the date range used, and the field interpretation, so disagreements (calendar vs fiscal quarter, missing record, wrong field) get resolved in minutes.
Reviews the past 24 hours of queries, marks answers correct or wrong, flags unanswered questions for prompt review, and watches daily token spend in dollars, no separate Anthropic dashboard needed.
How we ship it
Three milestones from kickoff to handover, typically four to eight weeks in total. The first lands inside two weeks with a live connection to your data.
Technical specification document and a working live connection to one Airtable base, with schema auto-discovery confirmed end-to-end.
The bot answering five canonical question types accurately inside Slack, with source links and per-query logging active.
Admin page live with full audit log, daily LLM cost display, and full project handoff with documentation your team owns.
Scope
What you get
The copilot is built on Claude and MCP, deployed under your own Anthropic or AWS Bedrock account, and handed over documented. These are the parts.
01
Installed in your workspace, answering in a private channel, with honest "I do not know" responses instead of guesses.
02
Live read connectors for Airtable and Google Sheets, with bases, tables and fields discovered automatically and Sheet structure inferred from headers.
03
The MCP tools and prompts that turn a question into the minimal set of records to fetch, with confidence flags on fuzzy joins.
04
Deep links to the exact Airtable view or Sheet range used, and a plain explanation of which fields were summed.
05
Every question, source, answer and response time, with correct or incorrect marks, flagged questions, and daily LLM spend in dollars.
06
The full repository, deployment under your account, and a runbook for adding a connector or a new question pattern without us.
The copilot is a chatbot pattern built on Claude and MCP. See how we ship conversational AI that finishes the job, how we ship Claude in production, or another shipped connector in the Meta and Google Ads MCP connector.