Digital Employees · Ledger
See exactly what you invested and what it returned, with evidence
The Investment & ROI Ledger records every cost of a rented agent, separates what happened from what the agent caused, and stops spend when the numbers say so.
Three questions, answered with evidence
For every rented agent, squad, campaign, client and agency, the ledger answers:
- How much did I invest? All-in, including the costs that are easy to hide.
- What happened, and how much can be credited to the agent? Observed outcomes, agent participation and — only where an experiment supports it — incremental effect.
- Should I continue, change or stop? Guardrails and recommendations.
Waves: W1 v0 (spend meter and cost capture) → W2 v1 (observational attribution and monthly statements) → W3 v2 (experiments, incremental value, outcome contracts).
Four things that must stay separate
| Concept | Unit of record | What it proves | Can it be summed as business value? |
|---|---|---|---|
| Observed outcome | A deal paid, meeting held, ticket resolved, collection received | The event happened | Once per canonical outcome, net of refunds and corrections |
| Agent touch | An Action or message linked to that outcome in the attribution window | The agent participated | No. Touch counts and influenced revenue are descriptive only |
| Incremental effect | An experiment cohort, metric, population and window | The treatment changed results on average | Yes — as an estimate with an interval and a named method, once per final result |
| Customer investment | A reconciled cost entry | What you spent, including supervision time | Yes, in one base currency |
Stated plainly
An experiment can show that a treated population produced more than its control. It can never identify which individual sale was caused. Individual rows are therefore observed or assisted — never caused.
What gets counted as investment
| Cost line | How it is captured |
|---|---|
| Agent rent | Daily accrual, prorated on hire and return |
| Usage credits and AI inference | Per trace span, linked to agent instance and goal or campaign |
| Messaging pass-through | Per message and conversation category; reconciled monthly against provider invoices (verify provider pricing models) |
| Voice calls | Per minute or call, including failed attempts |
| Ad spend controlled by agents | Daily spend per campaign where the controller is an agent, separate from human-managed spend |
| Avatar, video, image and voice generation | Per render job, including re-renders |
| Data, enrichment and SERP services | Per lookup |
| Human supervision time | Minutes × your hourly rate. Always shown — leaving it out overstates automation return |
| Setup, migration and onboarding | One-time, amortised over months you choose |
| Human Expert Cloud escalations | Per task |
Metric vocabulary
| Metric id | Formula | Break-even |
|---|---|---|
| `observed_revenue` | Unique realised outcome amounts, net of refunds | — |
| `influenced_revenue` | `observed_revenue` restricted to outcomes with an eligible agent touch (descriptive) | — |
| `return_on_spend_ratio` | measured_return / investment | **1.0** |
| `net_roi_ratio` | (measured_return − investment) / investment | **0.0** |
| `incremental_contribution_estimate` | treatment_effect × exposed_population × contribution_per_outcome (final experiment only) | — |
| `incremental_net_roi_ratio` | (incremental contribution − incremental investment) / incremental investment | **0.0** — the only ratio that may be called financial ROI |
| `cost_per_accepted_outcome` | total_investment / unique accepted outcomes | — |
| `human_minutes_per_accepted_outcome` | (setup + review + correction minutes) / unique accepted outcomes | — |
Why "ROI" is never used alone
Two common ratios differ by exactly 1.0. Confusing them turns break-even into an apparent 100% return. So every guardrail, tile, API field and statement line uses a metric id with its unit and break-even value.
Arithmetic check (not a performance claim): 100 invested and 100 returned gives return_on_spend_ratio = 1.0 and net_roi_ratio = 0.0.
Financial return means contribution, not gross revenue. It is incremental gross contribution after the direct cost of delivering the product or service, or an evidenced avoided cash expense. For collections it counts the financing benefit or avoided bad debt, never the whole invoice. If you cannot supply margin data, the statement shows revenue lift and spend side by side and marks financial ROI unavailable. Time saved and weighted pipeline are never inside financial return.
Attribution honesty rules
Nine rules the ledger follows on every figure.
- 1
1. Show the method
Every figure shows its attribution model, evidence level and data-quality state.
- 2
2. No individual causation
Individual outcomes are `observed` or `assisted`. No individual outcome is ever labelled `caused`.
- 3
3. Final results only
Incremental value posts only from a result that is final, significant under the pre-registered plan, within the contamination rule and review-passed.
- 4
4. Negative results stay
A negative or inconclusive result stays visible. It is not re-run until it turns positive.
- 5
5. Frozen baseline
The baseline (default: the 90 days before rental, adjustable) is frozen at contract start and cannot be edited later.
- 6
6. Confounders flagged
Seasonality, discounts, price changes and existing pipeline are flagged.
- 7
7. Say "not established"
Without a valid experiment, the statement shows observed outcomes, participation and full cost, and says incremental value is not established.
- 8
8. Insufficient evidence
If data quality for a metric falls below your threshold, it reads "insufficient evidence", not a number.
- 9
9. Overrides are visible
Humans may override attribution with a reason. The override is logged, versioned and shown on the statement.
The monthly Agent Investment Statement
| Section | Display rule |
|---|---|
| 1. Investment | Every cost line by agent; reconciled, estimated, pass-through and supervision shown separately |
| 2. Operations | Actions completed, customers reached, meetings held, tickets resolved — counts, no value claims |
| 3. Influenced outcomes | Unique observed outcomes with agent touches and evidence links; labelled observational |
| 4. Incremental effect | Estimate, interval, population, window and method — or "not established" |
| 5. Financial return | Contribution lift or evidenced avoided expense; `unavailable` without margin data |
| 6. Metrics | Each printed with its metric id, unit and break-even |
| 7. Guardrails & confounders | Pauses, cap breaches, approvals, seasonality, data gaps, overrides with reasons |
| 8. Exceptions & disputes | Refunds, prior-period adjustments, dispute window and deadline |
Delivery and three numbers that are never merged
The statement is delivered on day 1–3 of each month as PDF, in the portal and as a WhatsApp summary, signed with a hash and stored immutably. The summary always shows three numbers separately, in this order:
- Total customer investment, including supervision time.
- Observed and influenced outcomes, labelled observational.
- Experimentally estimated incremental contribution, with interval, population, window and method — or "not established".
For agency clients, the statement uses the agency's price as investment. One3's wholesale price appears only in the agency's margin view.
Budget guardrails and stop triggers
| Guardrail | Default behaviour (configurable) |
|---|---|
| Daily spend cap per agent or squad | Hard stop on R3 spend Actions; notify owner |
| Monthly cap per agent, squad or tenant | Hard stop; approval to raise |
| `min_net_roi_ratio` after minimum sample | Below threshold after minimum days and outcome opportunities: pause external-facing Actions, keep internal work, alert owner with a diagnosis |
| `max_cost_per_accepted_outcome` | Alert at 80%, pause at 100% |
| Max supervision ratio | Alert: supervision cost exceeds a set share of investment |
| Pass-through anomaly | Spike vs rolling baseline: pause the channel, alert |
| Provider quality signals | Pause outreach on that channel |
| Data-quality floor breached | Metrics read "insufficient evidence"; no automated scaling decisions |
| Evidence level required | Guards on `incremental_*` metrics fire only on a final, reviewed experiment |
Live controls
On every rented agent and squad.
Spend meter
Today's spend vs daily cap, month-to-date investment and return, `net_roi_ratio` with its evidence badge, supervision minutes.
Stream view
Each cost-incurring Action appears as a line, with who approved it.
Forecast bar
Projected month-end spend at the current run-rate vs cap.
One-click controls
Pause agent, lower cap, switch to draft mode (A2), open diagnosis.
Dispute evidence pack
Trace ids, Actions with approver, message logs, CRM history, payments, attribution rule version, baseline and holdout proof.
Roll-ups
Agent → squad → client → agency portfolio, as alternate views of the same facts, never added together.
Questions about the ledger
Does One3 guarantee a return?
No. One3 never promises guaranteed revenue, leads or ROI. It shows what you invested and what happened, with the evidence.
Why is supervision time a cost?
Because it is real. Hiding it would overstate the return of automation and erode trust.
What is the difference between influenced and incremental?
Influenced revenue is any outcome the agent touched — descriptive only. Incremental contribution comes from a pre-registered experiment and is the only figure that can be called financial return.
Can I disagree with a statement?
Yes. Raise a dispute in the window. An evidence pack is assembled, reviewed by One3 billing and you, and upheld disputes produce a credit note and a rule fix.
Will an agent be paused on noise?
No. Minimum-sample rules apply before any ROI-based pause. Defaults are labelled estimates until real data sets them per role.
When is the ledger planned?
v0 spend meter and caps in W1, monthly statements in W2, experiments and outcome contracts in W3.
See the ledger design
Finance teams are welcome to review the metric definitions and statement layout with us.
More in Digital Employees