GoodGraph

FOR TEAMS ROLLING OUT OR BUILDING AI

Is your AI improving work, or just being used?

GoodGraph tracks what happens after the AI answers: whether people accept it, fix it, or throw it out, why, and what happens to the work next. You see where AI is saving effort, where it keeps getting things wrong, and what you can honestly tell leadership.

For companies, public bodies, and foundations, with AI they buy, like Copilot or ChatGPT, or AI they build, like their own agents.

Book a free working sessionSee how it works

GoodGraphAI evidence

Example

Customer replies · Support · in-house agent

Uses per week1,240 replies
Good enough?41% sent as is
Faster?15 → 10 min

Meeting notes · All teams · Copilot

Uses per week860 notes
Good enough?74% usable as is
Faster?20 → 5 min

Contract summaries · Legal · ChatGPT

Uses per week95 summaries
Good enough?31% usable as is
Faster?Not yet known
MEASURED STAFF-REPORTED NOT YET KNOWN

WHY NOW

Usage tells you AI was used. It doesn't tell you whether the work improved.

Leaders are being asked for the ROI of AI. Usage numbers cannot answer that question. Operational evidence can. It shows whether the work got better, faster, more accurate, more consistent, or less time-consuming.

ADOPTION COUNTS

  • Licenses or seats
  • Active users
  • Usage frequency
  • Prompts or agent runs
  • AI tools deployed
  • Self-reported time saved

OPERATIONAL EVIDENCE ASKS

  • Where did AI require human correction?
  • Which tasks is AI actually good enough for?
  • Are the same interventions recurring?
  • Where is human judgment still essential?
  • Did the work get faster or better than before?
  • What happened to the outcome?

Knowing whether AI improved the work means seeing the whole loop, not just the AI's output.

The most useful evidence is the moment a person decides what to do with the AI's work. Today it is rarely recorded.

WHERE THE LOOP BREAKS

What happens when someone corrects the AI?

CORRECTIONS GET LOST

A support agent checks a customer reply written by AI

Example · illustrative figures

AI WROTE

“You can request a refund within 30 days.”

HUMAN CORRECTION

“That’s the old policy. It changed to 14 days in March.”

TODAY, WITHOUT GOODGRAPH

The fix improves one reply and then disappears. Nobody sees that the same mistake was fixed 41 times, in which teams, or that it still slips through elsewhere. Leadership still only sees usage numbers.

WITH GOODGRAPH

Every review decision becomes evidence, rolled up for the whole workflow.

CUSTOMER REPLIES · WHAT REVIEWERS DID

  • Accepted as is 41%
  • Edited 47%
  • Rewritten or discarded 12%

THE FIXES PEOPLE MAKE MOST

A generic apologyThe specific next step for this customer64×
“Refunds within 30 days”“Refunds within 14 days”41×
A promise to call back todayThe real response time23×

WHAT YOU CAN TELL LEADERSHIP

  • Replies sent in 1.4 days, down from 2Measured
  • 41% sent without edits, up from 28%Measured
  • 5 minutes less per reply, 15 down to 10Measured
  • Staff rate the tool 4.1 out of 5 for repliesStaff-reported
  • Effect on customer satisfactionNot yet known

TWO STARTING POINTS

Whether you bought the AI or built it, the question is the same.

Did people trust it, what did they fix, and did the work get better? GoodGraph works across tools, so the evidence stays yours when you change vendors or models.

FOR TEAMS ROLLING OUT AI TOOLS

Microsoft 365 Copilot, ChatGPT Enterprise, Claude, Gemini

Rolling out an AI assistant?

Your vendor's dashboard shows who is using it. It doesn't show which tasks it's good enough for, what staff keep fixing, or whether the work changed.

GOODGRAPH DELIVERS

  • An AI evidence audit: what your tools already record, what you can claim today, and what to start capturing
  • Your AI usage data cleaned, kept, and ready in the reporting tool you already use, like Power BI
  • An evidence dashboard, task by task, labeled measured, staff-reported, or not yet known
  • Training, so your team can run it and leadership can trust it

FOR TEAMS BUILDING AI AGENTS

Your own agents and assistants, built on LangGraph or similar

Built your own AI agent?

Your traces show what the agent did. They don't show what the person reviewing it did: approved it, edited it, or overrode it, and why.

GOODGRAPH DELIVERS

  • An audit of what you already log in LangSmith, Langfuse, or OpenTelemetry
  • An instrumentation plan for human review: which decisions, reasons, and outcomes to capture
  • A telemetry guide your engineers can implement
  • A review funnel per workflow: approved, edited, overridden

HOW IT WORKS

Start with one workflow. Capture the evidence while the work is happening.

Some of the evidence already exists in your systems. The rest is a few deliberate signals, not a new platform.

  1. 01

    Pick one workflow

    Choose one recurring task where AI already does real work, and agree what it is meant to improve: time, quality, consistency, or turnaround.

  2. 02

    Capture the review

    Use the evidence you already have, then add what's missing: what the AI produced, what the person did with it, why, and what happened next.

  3. 03

    Report what changed

    A regular evidence report: what is working, what people fix, where errors get through, and what you can honestly claim, compared with before.

BUILT FOR TRUST

  • Your data stays in your own environment. We work inside your systems, like Microsoft 365, with your IT team's approval.
  • Results are reported by team or role, never by individual.
  • Conversation text is optional. When it is used, it stays in your systems.
  • Every claim is labeled measured, staff-reported, or not yet known.

START WITH ONE AI WORKFLOW

You don't need to measure everything at once.

Bring one recurring workflow where AI already does real work: a rollout task your staff use Copilot or ChatGPT for, or an agent your team has built. In a free working session we'll agree what it's meant to improve, find the evidence you already have, and name the signals that are missing.

Book a free working session

Or email hello@goodgraph.ai