Your read
Your organization, sorted
The bolt-on simulator
Everyone's first move is the same: drop an agent in, watch the compute bill shrink, call it a win. This runs that move on one outcome and shows the part the invoice hides.
The compute bill always drops. Whether the outcome gets cheaper is a different question, and the answer is set by the organization, not the agent.
A simulator, not a quote. One outcome, a fixed model rate and a fixed labor rate; every figure is directional. The point isn’t the number, it’s where the cost moves when the organization underneath can’t carry the change. The readiness dial is seeded from your Organization read and stays draggable.
Where does your work wait?
Pick one workflow and list its steps. For each step, put the hands-on work time and how long the case sits before that step starts. Most processes spend far more time waiting than being worked, and making the work faster doesn’t touch the wait.
Add at least two steps with their times, and the read builds as you type.
A first read, not a time study. Rough numbers are enough: the gap between work and wait is usually so wide that precision doesn’t change the answer. Wait time is calendar time, so a case that sits over a weekend counts the weekend. The pattern comes from Michael Hammer’s process work, where an application worked on for 17 minutes spent 22 days in process. Beta.
Agent ROI & capacity
—
Full-time equivalents: how many full-time people do the work you’re thinking of handing to agents. Two people who each spend half their time on it count as one.
Their average yearly salary, before benefits and overhead. The fully-burdened multiplier under Assumptions adds those.
The share of that manual work agents could take over. This is the one number we start from your scan: 70% if your Technology read is production-ready, 50% if foundational, 30% if not yet safe. Move it to what you believe is realistic.
Assumptions sensible defaults · all editable
What a person really costs beyond salary: benefits, payroll taxes, equipment and space. 1.35× means every $100 of salary costs about $135 in total. Most firms land between 1.25× and 1.4×.
How much of their week goes to repetitive manual or upkeep work, rather than judgment, relationships or customer work. Only this share of their time is available for agents to take.
The time a person still spends checking what agents do, as a share of the work agents take over. 20% means about one hour of checking for every five hours of work handed off. Checking is where saved time most often leaks back, so this comes off before anything counts as a saving.
How many agent steps run before a person checks. 1 means a check after every step. Chaining steps so a person checks the finished result once divides the review time by the length of the chain. Only chain steps whose result can be verified, and undone, at the end.
The share of agent output that’s usable without being redone. The rest comes back to a person as rework. In one large study of AI-written code changes, more than 15% introduced a problem, so 85% is a sensible start for work you haven’t measured yet.
How much of the full benefit shows up in year one while the team and the agents settle in. 60% means year one delivers 60% of a steady year; years two and three count in full.
What the freed hours are worth once people move to higher-value work. 1.0× counts them only as saved cost; 1.3× assumes each freed hour produces 30% more value in its new use. Only the full case uses this.
The one-time cost to design, build and launch the agent team. It drives the payback and the 3-year ROI.
The yearly cost to keep it running: models and compute, hosting, monitoring and support.
Counts only work agents take off people’s plates, never growth. Until you name the budget line that changes and who changes it, the figure is capacity, not cash: time freed that the business may or may not turn into money. Agent-handled / yr is the full cost of the hours agents take over (people × salary × multiplier × manual share × automated share). Review and rework come off that to give Freed / yr: review time divided by the steps per check, plus the share that has to be redone. Year one applies the ramp and subtracts the run cost. Growth is left out, so treat this as the low end. Payback is the months until yearly hard savings cover the upfront build.
Hours saved only become a saving when a budget line changes: fewer contractors, a hire you don’t make, overtime that stops. Name the line and the person who’ll change it. If nobody can be named, treat the floor as capacity, not cash.
Adds the value of the freed hours in their new work, using the value multiplier. Year one applies the ramp; years two and three count in full. 3-year ROI is three years of this return, minus the upfront build, as a percentage of the upfront build.
Show the math
Model, not a forecast. The automation target is the only figure seeded from the diagnostic — the Technology lane's score and its band (Production-ready ≥56 / Foundational 33–55 / Not yet safe <33). Review and rework come off the agent-handled work before anything counts as freed. The floor counts freed capacity once and excludes growth; the full case ramps year-1 and holds years 2–3 steady. Validate every other input.
Map your workflow
Pick one workflow you run today — claims, collections, renewals, onboarding, anything with many hands. Five questions show where agents can take the chasing, which decisions stay with people, and whether your record and audit trail can carry agents. See a mortgage mapped at Silver Lake →
The follow-up and hand-off work that eats the days. Pick all that apply.
Decisions that bind someone to money, a contract or a regulatory outcome — and the person who owns each.
The short list you would enforce, not just hope for. Pick all that apply.
Where the truth about one case sits while the work moves.
If an auditor asked who did this step, could your trail answer?
A first map, not a design. The read comes from your answers alone: agents take the chasing, people keep the gates, and the record and the trail decide whether agents can be governed at all. It’s the same pattern Silver Lake runs on a mortgage. Beta.
Ask the Work
Put a question to the whole body of Pegasus writing — Reggie’s essays, the Pegasus4i practice material, and the Pegasus Collective material on ExO 3.0. It answers from what’s written, names the pieces it drew on, and says so when a question falls outside the work.
Reads, doesn’t opine. Answers are grounded in the published writing and cite the pieces used. Beta — a fair daily question limit applies per profile.
Model Watch
The open and closed model landscape, one line each — lab, country, license, size, capability, and price — so the scoping and placement work can cite one source instead of restating figures. Every row carries the date it was last checked.
| Model | Country | Open? | License | Sizes | Context | Score | $/M in·out | Verified |
|---|
—
Vendor-reported. Capability scores and prices are list or vendor figures unless a row says otherwise; the Verified date is the honesty stamp. Score = Artificial Analysis Intelligence Index where shown. Beta.