◆Agents that turn product signals into revenue.northstar geocities
Simulated projection

Summary of cost savings, revenue projections and quality of life (simulated)

A simulated look at cost savings, revenue, and quality of work at a 1,000-person company. By Charles Keatts, Samadhi. November 2026. Simulated projection, not measured results.

Most value reports for software stop at two numbers: money saved and money made. We built one of those for Northstar, and it’s summarized below. But the more useful question for anyone thinking about letting AI agents into their work is a different one. Does the work get better for the people doing it, and for the customers they serve?

This article describes what we expect Northstar to do at a 1,000-person software company: for sellers, customer success managers, product managers, engineers, and security teams, and for their customers. Every figure here is a simulation built from public benchmarks and our own assumptions.1 Parts of the product described are available today. Others are in preview or planned, and we mark which is which.2 We’re looking for design partners to replace our assumptions with measured results.

Three kinds of stress, three modules

Northstar has three parts. Each one removes a different kind of stress from the work.

ModuleWhat it decidesStress it removes
RevenueWhat is worth doing, ranked by revenue"Am I working on the right thing?" and arguments about priority
Innovation EngineWhat to build next, tested in a sandbox firstBuilding features nobody uses, and doing research by hand
Agent ControlWhat each AI agent is allowed to do"What if an agent does something wrong in my name?"

If people waste less work, argue less over priorities, and trust the agents working for them, two things follow. Their weeks get better, and the cost and revenue results become more believable, because people keep using tools that make their work easier.

A week at a 1,000-person company

The people below are illustrative. They show the intended product behavior at a company with 80 account executives across three regions, 40 customer success managers, 25 product managers, 250 engineers, and a small security team, running about 150 AI agents from several vendors.3

Monday morning, an account executive

Sales reps spend most of their time on work other than selling.4 Instead of updating the CRM on Sunday night, Dana gets one daily digest, in her own time zone, never before working hours. Three of her deals changed. One prospect’s security review has stalled on a data-residency question, and Northstar shows the same blocker on 11 deals across all three regions. That pattern is already in front of product. Forecast prep takes 20 minutes.

Monday, a revenue operations analyst

Riya used to spend her week chasing reps to update deal stages. A cleanup agent now updates internal fields, with a receipt for every change and the ability to undo it. It is never allowed to delete or merge records. Riya spends her time on territory design.

Monday, a customer success manager

A large account shows three warning signs at once: usage is down, high-priority tickets are open, and invoices are overdue. Any one of them would be noise; together they’re a signal. Northstar drafts a save plan as a task. Nothing is sent. Luis edits it and calls the customer, 14 weeks before renewal instead of two.

Tuesday, a product manager

Each product line gets a weekly review of at most five proposals, built to take 15 to 20 minutes. Each one shows risks and potential side by side, the revenue at stake, and an evidence grade. The data-residency request is grade A: real customers, in their own words, tied to real pipeline. An idea drawn mostly from competitor news is grade C. Northstar also notices that another team has a similar proposal, so the two meet once instead of building twice. Priority comes from evidence, not from who complained loudest.

Wednesday, an engineer

The approved idea is built first as a small prototype in a sandbox with sample data. Last quarter, four other ideas were stopped at this stage when the customers who tried them said they didn’t solve the problem. That shows up as “builds avoided: 4, about 13 engineer-weeks,” and the team counts it as a win, not wasted work.

Wednesday, the security lead

One view covers every agent the company runs, including agents from other vendors. Actions that need a person wait in an approval queue. Six agents a marketing team had connected on its own are brought into the registry with limits instead of being shut down. Hidden instructions on a web page the research agent read are flagged as untrusted and can’t influence any proposal. What used to be a day of searching through logs takes ten minutes.

Thursday, approval load

One region’s managers are approving 14 actions a day, too many to read carefully. Northstar suggests grouping low-risk approvals and adding a delegate. Rubber-stamping falls, and so does the workload.

Friday, the customer

The compliance lead who asked about data residency gets a note from her account executive, written by a person: “Here’s a prototype. Would your team test it?” She agrees, with consent recorded. Her request didn’t disappear into a backlog.

Friday evening, the pulse

A one-minute survey asks whether Northstar made the week easier. Results are shown only for groups of five or more. The most common comment becomes next week’s proposal.

What gets better

For employees

For their customers

Safer agents, and less anxiety about using them

Security controls lower the real risk. Our design follows the OWASP Top 10 for Agentic Applications, with a test planned for each risk.6 But people worry about more than security incidents. They worry an agent will embarrass them, do something that can’t be undone, get them blamed, or replace them. So the product is designed to:

  1. Show a preview of every outbound action before it happens.
  2. Make actions undoable wherever possible, and say so plainly when they aren’t.
  3. Publish a short “never alone” list: send external messages, move money, change prices, issue refunds, delete records, change permissions.
  4. Explain every answer, with links to the source record and an evidence grade.
  5. Let agents earn autonomy one task type at a time, after a clean record, with a person approving each step.
  6. Review incidents without blame, fixing the system rather than the person who approved.
  7. Keep people the authors of their work, with agents shown as helpers.
  8. Prevent approval fatigue by grouping low-risk approvals and watching for rubber-stamping.

One line we won’t cross: none of this becomes employee monitoring. Workload signals are opt-in, shown only for groups, and visible to the employee first.

The numbers: a 1,000-person company, year 1

These are simulated results from our value model, conservative to optimistic, built on public benchmarks and our own assumptions.1,9

Value lineConservativeOptimistic
Time and cost savings$162,984$1,379,813
Revenue attainment (better prioritization, winning with the right features, keeping customers)$122,326$1,093,944
Agent Control (management time, wasted run cost, incidents avoided)$227,858$791,505
Innovation Engine (research done, bad builds caught early, revenue sooner)$40,623$360,356
Total benefit$553,791$3,625,618
Northstar cost (conservative case)$184,180
Benefit for each $1 of cost3.0x12.0x

At this size Agent Control is the largest single line in the conservative case, about 41% of total benefit, because more agents mean more to manage and more that can go wrong. Avoided incidents are counted as expected losses avoided; a company that never has a serious incident saves nothing on that line.

Over three years, revenue attainment becomes the largest line, at $2.53M to $17.3M, because extra revenue keeps building while savings hold steady. Total net value over three years is $3.55M to $24.9M.1 Features adopted within 90 days rise from 60.0 of 240 built to 61.4 to 69.9.7

Quality of life, alongside the numbers

We don’t convert quality of life into dollars. Doing so would count time savings twice, and it’s the kind of figure a careful buyer should question. Instead we track it next to the dollar figures, because it shows whether those figures are believable. Adoption is the assumption most likely to break the model, and people keep using tools that make their week easier.

Here is what we expect by the end of year 1 at a 1,000-person company. These are placeholders until design partners measure them.8

MeasureEnd of year 1 (placeholder)
Weekly active users, share of licensed50% to 70%
"Northstar makes my week easier" (1 to 5)3.5 to 4.1
"I'm comfortable letting Northstar act on low-risk tasks" (1 to 5)3.4 to 4.0
Hours back put to use, per seller per week1.5 to 3
After-hours admin work, vs. beforeDown 15% to 30%
Approvals per manager per day3 to 5
Requesting customers told when their request shipped50% to 85%
Serious agent incidentsTarget zero

How we’ll measure it with design partners

Each design partner runs a 14-week pilot:

  1. Baseline (2 weeks). Read-only connection and a short survey. No agent actions.
  2. Suggest only (4 weeks). Agents draft; people do everything.
  3. Lighter approval load (4 weeks). Grouped approvals and quiet hours, rolled out one region at a time so the others serve as a comparison.
  4. Earned autonomy (4 weeks). Low-risk tasks can step up after a clean record, with a person approving each step.
  5. Joint readout. Every figure labeled as measured, with sample sizes.

Most measures come from agent run logs, so the only extra work is a five-question weekly survey that takes under a minute. Survey and time tracking are opt-in, results are shown only for groups of five or more, and people can see and delete their own data. Any serious agent incident pauses the move to greater autonomy.

Become a design partner

We're looking for a small number of companies willing to test these assumptions with us and publish what's real.

Write to usSee Agent Control

Appendix: results by company size

Year 1, conservative to optimistic. Net value is after Northstar cost and includes all four value lines.1

Company sizeTotal benefit, year 1Northstar costBenefit per $1Net value, year 1Net value, 3 years
5 people$6,606 to $34,766$3,3672.0x to 8.3x$3,239 to $30,597$22,751 to $145,030
100 people$81,911 to $515,939$34,4512.4x to 10.7x$47,460 to $467,845$524,723 to $3,290,545
200 people$148,783 to $948,171$62,9402.4x to 10.6x$85,843 to $858,415$971,225 to $6,247,765
1,000 people$553,791 to $3,625,618$184,1803.0x to 12.0x$369,611 to $3,324,700$3,549,049 to $24,925,275
10,000 people$4,599,062 to $25,463,964$1,007,0504.6x to 14.7x$3,592,012 to $23,730,115$25,214,652 to $168,974,519

Agent Control’s share of benefit grows with company size: about 24% at 5 people and 72% in the conservative 10,000-person case.1

Notes

  1. Northstar Value Report 2, Samadhi Consulting LLC, November 2026. An illustrative simulation of five company sizes over 6, 12, and 36 months, with conservative and optimistic cases. It is not measured customer results. Three-year figures exclude compounding. Benefit per $1 uses the conservative-case cost. Prices in the model are placeholders.
  2. Status as of November 2026. Available today: revenue-weighted scoring, CRM and billing sync, and Agent Control’s registry, approval queue, review policies, and run history. In preview on sample data: the Innovation Engine. Planned: approve-before-act checks, budgets, alerts, audit replay, shadow-agent discovery, cascade limits, and a kill switch. Proposed in this article: earned autonomy, quiet hours, grouped approvals, and the time-back survey. See the Argus page.
  3. The company, people, deal sizes, and agent counts in this story are illustrative placeholders, not a customer.
  4. Salesforce reports that sales reps spend 60% of their time on non-selling tasks. Salesforce, 40 Sales Statistics to Watch for in 2026.
  5. Design principle for the Northstar operator: send external communications, move money, change prices, issue refunds or credits, delete or merge records, change permissions, invite users, accept terms, or write to an ERP only with a person's approval.
  6. OWASP Top 10 for Agentic Applications for 2026, OWASP Gen AI Security Project.
  7. "Adopted" means used regularly by at least 20% of the people it was built for within 90 days. The number of features built stays the same; what changes is how many get used. These counts describe the same effect as the engineering savings and are not added to the dollar totals.
  8. Quality-of-life ranges are our placeholders, assuming slower adoption and lower use of freed time at large companies, consistent with the value model's caution about large organizations. They will be replaced only with design-partner measurements.
  9. Benchmarks the value model draws on: SaaS Capital, 2026 Revenue per Employee Benchmarks; Salesforce State of Sales; Landbase, 2026 win rate benchmarks; Benchmarkit and Aleph, 2026 SaaS and AI Metrics Benchmarks; and SaaStr headcount data.