◆Agents that turn product signals into revenue.northstar geocities
Better work · Safer agents

Better Work - Safer Agents

Our value report measures dollars: time back, revenue, agent savings, and faster product building. This page covers what it does not measure: how the work feels for the people using Northstar, how their customers are affected, and how to make AI agents feel safe enough to use. The two are connected. People keep using tools that make their week easier and that they trust, and use is what drives the numbers.

Status: the dollar figures here come from a simulation, not measured customer results. The quality-of-life effects are hypotheses we are testing with design partners. Some of the design principles below are in Northstar today; others are being built and tested.
Why the experience moves the numbers

Adoption is the assumption most likely to break any value estimate. So we ran the same 1,000-person company three ways. Year 1 is identical. After that, a good experience brings more teams in and helps people use the time they get back, and a rough one does the opposite. Each year builds on the last, so the gap grows.

1,000-person companyYear 1Year 2Year 33 years
Rough experience
It feels like extra work or feels risky. Use drifts down 15% a year, and less of the freed time gets used.
$370K$1.49M$1.26M$3.13M
Value Report, conservative
The report as published, held flat after year 1.
$370K$1.59M$1.59M$3.55M
Good experience
It makes the week easier and people trust it. More teams opt in (up 15% a year, capped), and more freed time gets used.
$370K$1.70M$1.95M$4.02M

Net value after Northstar cost: savings plus extra revenue. Simulated, built on our value report's conservative case. Growth rates are placeholders until design partners measure them.

The difference between a good and a rough experience is about $893K over three years at this size, and the good case lands about 13% above the report. That is why we treat quality of life as the strongest protection for the dollar results, not a side benefit. We report it next to the dollars, never added to them, so nothing is counted twice.

What the research says
28%

of a sales rep’s week is spent actually selling. The rest goes to deal admin, data entry, and other work.

Salesforce, State of Sales research (2023)
21%

of employees worldwide were engaged at work in 2024. Managers account for about 70% of the difference in team engagement.

Gallup, State of the Global Workplace 2025
46%

of people are willing to trust AI, even though 66% use it regularly.

KPMG and University of Melbourne, Trust, Attitudes and Use of AI (2025), via Euronews
40%+

of agentic AI projects are predicted to be canceled by the end of 2027, citing costs, unclear value, or weak risk controls.

Gartner press release, June 25, 2025
80%

of features in the average software product are rarely or never used.

Pendo, Feature Adoption Report (2019)
How the work gets better, by role

Sales

  • Fewer evenings updating the CRM and preparing for forecast calls.
  • A ranked answer to "what should I focus on this week?", with the math shown.
  • A real answer when a deal waits on a feature: the request is logged, tied to revenue, and maybe already prototyped.

Customer success

  • Early warning on renewal risk, so there are weeks to fix a problem instead of days.
  • The chance to tell a customer "you asked, we built it."

Product

  • Research arrives done: 1 to 4 hours back a week per product person in the simulation.
  • Priorities set by revenue-weighted evidence, not by who is loudest, so saying no costs less socially.
  • A review that stays small: at most 5 proposals, about 15 to 20 minutes.

Engineering

  • Fewer wasted builds: a cheap sandbox prototype stops features nobody would use before a full build.
  • More of the work actually gets used.

Security and leadership

  • One view of the AI agents at work, with approvals, budgets, and an audit trail.
  • A record of what each agent intended, did, and cost that a finance team can check.
How their customers benefit

These are the people our users sell to and support.

  • Their requests are heard, tied to their account, and tracked to shipping instead of vanishing into a backlog.
  • They get what they asked for sooner: 2 to 6 weeks earlier in the simulation.
  • Fewer surprises at renewal, because problems surface early.
  • Better conversations with sellers who arrive prepared.
  • No unreviewed AI messages, surprise price changes, or mistaken credits: a person approves those.
  • Prototypes go only to customers who asked and agreed to try them.
Agents that feel safe to use

Security controls lower the real risk. Feeling safe takes its own design work. People worry an agent will embarrass them, do something they cannot undo, get them blamed, or replace them. These principles answer those worries.

Preview before action

Anything that leaves the building shows what will change, who will see it, and what it costs before anyone approves it.

Undo where possible

When something cannot be undone, we say so plainly before the approval.

A short "never alone" list

Sending to customers, moving money, changing prices, refunds, deleting, and changing permissions always need a person.

Every answer explained

Claims link to their source, how strong the evidence is, and how sure the agent is. "I am not sure" is an allowed answer.

Autonomy is earned

Agents start by suggesting. A person approves each step up for a specific kind of task, and it steps back down after a problem.

Blameless by default

If an approved action goes wrong, the record shows what the person was shown. Reviews fix the system, not the person.

Teammates, not replacements

People see what agents did for them each week, and they stay the author of the work.

No approval fatigue

Low-risk approvals are grouped, only high-risk ones interrupt, and rubber-stamping is made visible so it can be fixed.

The line we will not cross

Quality-of-life features must never turn into employee monitoring. Any workload signal is opt-in, shown only in groups, and visible to the employee first. We do not build anything whose main use is to deceive, watch, or manipulate people.

How we will measure it

With each design partner we run a 14-week pilot. Most measures come from Northstar's own run records, so people only answer a short weekly survey, under a minute.

The weekly pulse

  • Northstar made my week easier. (1 to 5)
  • I trusted what Northstar told me this week. (1 to 5)
  • I am comfortable letting Northstar act on low-risk tasks. (1 to 5)
  • This week Northstar saved me about: none, under 1 hour, 1 to 3 hours, more than 3 hours.

Privacy rules for every pilot

  • Surveys and time tracking are opt-in. Using Northstar does not require either.
  • Results are shown only for groups of 5 or more. Managers never see individual answers.
  • Each person sees their own data first and can delete it at any time.
  • Pilot data is deleted or returned at the end unless the partner agrees in writing to keep it.

At the end, we keep going only if ease and trust are rising, there are no serious incidents, and the economics land in the report's conservative range or better. Every number we publish afterward is labeled measured, with its sample size.

See it play out

Follow a seller, a customer success manager, a product manager, an engineer, and a security lead through one week at a 1,000-person company, with the full value numbers and what is live today.

Read the article

Help us measure it

Design partners get our lowest prices and a real say in what we build. We get honest feedback and real numbers to replace our placeholders.

Become a design partnerSee pricing