Two narrow agents for the work between a promising conversation and cash.

01 / Market

Market conclusion

Our inference from the cited adoption, workflow, and trust signals is that the strongest near-term product opportunity is not maximum autonomy. It is a narrow outcome package that joins an existing system of record, handles repetitive preparation and follow-up, pauses at explicit checkpoints, and leaves enough evidence for a buyer to trust the result.

What buyers are buying

Our recommended purchase shape is an outcome attached to existing work, not a replacement for the buyer's CRM, accounting system, document stack, or judgment. This is a product hypothesis, not an observed Armalo purchase pattern. Upwork reports that 49% of full-time workers rely on freelancers to fill critical skill gaps and that 48% of CEOs planned to increase freelance hiring. Its skill-gap data is not direct purchase proof for AI workflow products. A separate Upwork analysis describes rising interest in AI-agent work while also finding demand for human verification. We interpret those signals as support for testing a named workflow, accountable review, and legible evidence; real willingness to pay for the Armalo products remains unproven.

The package therefore needs an explicit start state, a finite task graph, declared approval points, stop conditions, a limited authority scope, and deterministic receipts. “Automate sales” is too broad. “Prepare an approved proposal from this deal record and stop before sending” is buyable, testable, and reversible.

Who buys

02 / Opportunities

Ranked opportunities

Ranking favors painful recurring work, a visible budget owner, short time to value, integrations over replacement, and outcomes that can be measured without pretending correlation is causation.

QuickBooks' 2025 U.S. small-business late-payments report provides the clearest urgency signal for the first wedge: 56% of surveyed small businesses reported being owed money from unpaid invoices, with an average reported amount of $17,500, and 47% reported that some invoices were more than 30 days overdue. These are survey findings, not a promise that an agent will recover a particular amount.

Vendor pages position products in these categories. Qwilr positions proposal creation, signature and payment steps, buyer analytics, and CRM or accounting integrations as one proposal workflow. Kolleno positions centralized AR work, automated follow-ups, payment tracking, and configurable tone, escalation, and automation. These pages show how vendors describe their offers; they do not establish buyer demand for Armalo. Any outcome figures on those sites are vendor claims, not independent validation of these planned products. PandaDoc's proposal-software overview is also vendor-authored comparative evidence; it helps map category language but is not neutral market research.

03 / Products

Invoice Chaser product

A policy-aware receivables workflow that prioritizes approved overdue state, prepares the next permitted follow-up, and refuses to treat a payment, dispute, or promise as resolved without sufficient evidence.

Proposal Generator product

A proposal workflow that turns approved deal context, offers, proof, and terms into a reviewable draft without inventing commercial facts.

Quote-to-cash connection

The two wedges meet at an explicit state transition. Proposal Generator can prepare a versioned commercial offer and, after accountable approval, produce an accepted-scope handoff. Invoice Chaser can later consume approved invoice and customer state created by the buyer's existing systems. Neither agent should infer that acceptance created an invoice, that an invoice was sent, or that a payment cleared. Each transition requires an authoritative state or a human confirmation.

1. 01
   Discover

   Capture approved buyer context and unresolved questions.

2. 02
   Propose

   Prepare scope, proof, terms, and next steps for review.

3. 03
   Approve and send

   A human approves commercial content and the recipient.

4. 04
   Accept and hand off

   Record authoritative acceptance before fulfillment or billing state changes.

5. 05
   Invoice and follow up

   Use approved invoice state and bounded reminder policy.

6. 06
   Reconcile

   Close only against sufficient payment or resolution evidence.

Buyer, package, and metrics

Metrics need definitions before a pilot. Proposal cycle time should have a fixed start and end event. Acceptance rate needs a comparable cohort. DSO and recovery rate need an agreed baseline and must account for mix, seasonality, and unrelated collection activity. Manual review minutes, disputes, and false closes are guardrail metrics, not secondary details.

04 / Operating contract

Hermes autonomy contract

The conceptual design maps each product onto a bounded TaskPlan, an external Goal Gate, explicit goal materialization, tenant-scoped dispatch, capability scopes, leased execution with fencing, and evidence receipts. This is a capability-level review, not a deployment claim.

Goal contract: Invoice Chaser

Goal

Move each approved overdue case to the next permitted, evidence-backed state without exceeding contact, tone, escalation, or closure authority.

Required inputs

Approved invoice state, customer identity reference, aging state, prior communication summary, reminder policy version, channel permission, and any active exception.

Success

A permitted reminder is prepared or delivered through an approved checkpoint, a reply or exception is routed, or a case is reconciled against sufficient evidence; every transition has a receipt.

Guardrails

No invented balance or due date, no message beyond approved cadence, no escalation of a sensitive case without review, and no paid or closed state without authoritative evidence.

Metrics

Time to first reminder, DSO, overdue aging, recovery rate, manual review minutes, dispute rate, and false closes.

Goal contract: Proposal Generator

Goal

Produce a buyer-ready proposal review packet from approved sources while preserving uncertainty and withholding send authority until accountable approval.

Required inputs

Approved deal context, buyer problem, offer and scope sources, exclusions, proof library, commercial terms source, recipient state, and document policy version.

Success

A versioned proposal packet exposes source coverage, unresolved decisions, approvals, and a deterministic content digest; accepted state and fulfillment handoff remain separate authoritative events.

Guardrails

No invented price, scope, claim, term, testimonial, recipient, acceptance, signature, payment, or fulfillment commitment.

Metrics

Proposal cycle time, acceptance rate, revision count, manual review minutes, source-coverage failures, and corrections after approval.

Task graphs

Approvals and stop conditions

A stop is a successful safety outcome, not an agent failure. When the system cannot prove that a state transition is permitted, it should preserve the current state, emit a bounded exception, and request the smallest decision needed to resume.

05 / Proof

Deterministic proof and evidence receipts

A fluent summary is not proof. The completion packet must let a separate reviewer reconstruct which approved state entered the task, which plan and policy applied, who or what approved the action, and what observable result followed.

For Proposal Generator, deterministic proof includes source coverage, unresolved markers, the approved content digest, and the version sent. For Invoice Chaser, it includes the invoice-state digest, allowed-action decision, message digest, delivery result, reply classification, and the evidence used for reconciliation. Sensitive payloads belong in the authorized customer system, not in this public contract.

Retry, lease, fencing, and idempotency rules

• Retry only retryable work. Transient transport failure may retry; a policy rejection, conflicting source, or denied approval may not.
• Keep the action identity stable. A retry uses the same idempotency key for the same logical action, preventing duplicate sends or duplicate state transitions.
• Expire ownership. Every worker acts under a bounded lease. Expiration requires reacquisition and revalidation of current state.
• Fence stale workers. A newer fence token invalidates commits from a previous lease holder, even if the older worker finishes later.
• Re-read before consequential writes. Pricing, send, escalation, acceptance, payment, and closure steps must confirm that the approved version and source state have not changed.
• Bound attempts and time. Exhaustion creates a receipt and review item; it does not silently loop or broaden authority.
• Make observation idempotent. Duplicate webhooks, replies, or payment observations must converge on the same normalized evidence record.
• Never convert uncertainty into success. An unknown delivery, acceptance, or payment result remains unknown until authoritative evidence arrives.

Capability and verification boundaries

Rollout sequence

1. 1

   Offline fixtures

   Define states, policies, task plans, goal gates, deterministic receipts, and adversarial cases without contacting buyers or changing external state.

2. 2

   Read-only shadow

   Ingest approved exports, propose actions, and compare against human decisions. No send, escalation, acceptance, payment, or closure authority.

3. 3

   Draft-assist pilot

   Generate proposals or reminders for human review. Measure cycle time, review minutes, correction types, disputes, and false-close risk.

4. 4

   Approved bounded action

   Permit a narrow action only after explicit approval, with idempotency, leases, fencing, receipts, and an immediate stop control.

5. 5

   Policy-bounded routine work

   Consider limited pre-approval only for low-risk repeated actions whose false-positive rate, exception routing, and outcome evidence meet an agreed threshold.

6. 6

   Expand or stop

   Expand one capability at a time when evidence supports it. Stop or narrow when disputes, corrections, false closes, or unattributed outcomes exceed the contract.