RSI AGENT / OTHER APPLICATIONS

Different roles.
One shared workflow.

The enquiry, the job and the next person’s decision belong together. Explore how agents can prepare the handoff while people stay responsible for the work.

Explore three cases The product

Real-world situations, illustrated with fictional records. The product focus is the public-service CRM and case agent.

HOW THE TEAM WORKSEXAMPLE WORKFLOW

Context agentFind what matters

Planning agentConnect the next move

RSI COORDINATOROne shared mission.

Review agentKeep the human decision

CONTEXTPLANREVIEWHUMAN APPROVAL

MISSION STATEAWAITING A MISSION

THREE WORKPLACES / ONE APPROACH

Make the handoff
the strongest link.

Pick a case. Change a constraint, inspect what the agent prepares and see exactly what the next person receives.

CASE 01 / STORE OPENING

A Friday opening. A changed crew. A plan that holds together.

A Brunswick café needs its signage ready. The studio coordinator, designer, production lead and installer share the same scope, deadline and reviewed proposal.

Store opening

A café opening. One coordinated delivery.

The coordinator, designer, production team and installer work from the same opening-day brief.

Loading the example…

Detailed planning workbench The same case and current revisions

Preparing the shared business case…

CASE 03 / PUBLIC SERVICE HANDOVER

The caller explains it once. The next officer gets the full context.

Follow the prepared call, reconcile conflicting dates and review only permitted records. The receiving officer gets a source-linked handover before any proposed field update.

Service handover

One call. A case the next officer can use.

Follow a prepared service call into evidence review, a handover, a permitted update and a weekly workload view.

Loading the example…

Native workflow studio The same case, access policy and allowance

RSI AGENT / WORKFLOW STUDIO

A case in motion. Every intervention in view.

Explore the call, inspect the evidence, and review each proposed action.

ONE CASE · CONNECTED ASSISTANCE

Capture the call
Keep essential facts
Review permitted evidence
Prepare the handover
Officer approval
Reviewed processing

The interactive flow opens alongside the officer workspace.

PUBLIC SERVICE / LOCAL SIMULATION

One case. The right agents, at the right time.

A Victorian service officer moves from a call to a reviewed handover, with permissioned context and a visible work allowance.

Department-controlled environmentSample records only
Officer review at every write

Choose a case condition

A contact update with permitted evidence and enough illustrative budget for a reviewed skill draft.Depth: departmental policy profile

Estimated model tokens · simulated

80/ 2,400 illustrative allowance

850 reserved for essentials1,470 available

Hear the detail. Keep the context.

Follow a prepared caller transcript. Points worth checking appear beside their source.

Turn 1 of 6

Service enquiry · sample caseMasked caller · text playback only

CallerCALL-01 · 00:00

I am following up on service request [CASE-1042]. My callback preference has changed.

Sample text playback. At wrap-up, selected permitted records and a handover draft are prepared automatically.

Agents join when the case needs themPlanned, rule-based activation in this sample

Call capture

Active

Prepared transcript only. No audio is captured; speech costs are outside this estimate.

Essential facts

Active

Required extraction is reserved before optional checks.

Consistency check

Waiting

Only triggered by a timing question or selected evidence. Uncertainty stays visible.

Permitted documents

Waiting

Permission check → fixture masking → selected summaries with references.

Handover

Waiting

The template binds to the current source revision and officer note.

Reviewed processing

Waiting

Allowlisted sample field changes need permission, acknowledgement and the exact revision.

Skill draft

Not drafted

Derived from the reviewed sample. Human-reviewed local drafts are never auto-published.

Workload summary

Available

Local aggregation of synthetic work items; no model call or employee ranking.

Prepared playbooks supply the checklist. Optional analysis can wait; essential case facts and the handover keep a reserve.

2 prepared playbooks in use

Contact-preference checklist sample 0.1

Prepared, reviewable checklist: callback window, safe contact and written-update channel. Reused locally; no automatic learning.

Source and uncertainty checklist sample 0.1

Keep caller statements, document evidence and unresolved questions distinct. New skill drafts require separate human review.

Synthetic call started. Reveal each turn to see the required facts and estimated budget change.

Why an agent runs · allowance and boundaries

The sample scheduler applies a fixed departmental policy: retain required facts, reserve a handover, then admit optional checks when enough estimated model-work allowance remains. The ledger below shows each estimate and the amount used. Deferred work uses zero. These are illustrative model tokens, not a claim about Copilot’s billing or a configurable effort parameter.

Operation	Priority / state	Estimate / used	Reason
Turn 1 · essential facts	Required facts · Completed	80 / 80	Fixed estimate charged once for this event.

Policy-approved depth is set by the department. RSI would operate within approved tools, permissions and procurement terms.
Speech transcription, OCR, CRM licences and computer-use charges are outside this token estimate.
All call turns, masked records and weekly tasks are synthetic. No microphone, document upload, live Copilot or CRM connection is used.
Approved skills can supply checklists. New skills remain proposals until reviewed for a sandbox; they do not rewrite their own permissions.
Unresolved discrepancies are retained for an officer. This demo does not make eligibility, entitlement or other consequential decisions.
Redaction is not the same as anonymity. A real deployment needs access control, data classification, retention and an approved privacy process.

Preparing the public service workbench…

From prototype to an approved deployment

A proposed integration for an agency’s approved environment. This public demonstration runs on synthetic data in your browser. It has no connection to an agency, Copilot, telephony or CRM.

Work within the allowance.

The router reserves essential capture and handover capacity, then defers optional analysis when the sample budget is tight. Actual Copilot entitlements, credits, token meters and reasoning controls depend on the selected product and tenant. The demo’s estimated tokens are a separate illustration. Usage reference · Effort reference

Retrieve the permitted evidence.

Use case-scoped access checks, minimum-necessary text and source references. Salesforce search access and write access are separate integrations; field permissions and indexing freshness need explicit review. Search reference · Action reference

Keep the source and its limits.

Production transcription needs an approved audio path. Speaker labels do not verify identity. Masked case details still require access controls: de-identification does not guarantee anonymity. Speech reference · Privacy reference

Review actions and reusable skills.

Show the proposed field changes, recheck permissions and invalidate approval when evidence changes. Skills begin as reviewable drafts. Any permitted computer use needs separate configuration; government decisions and consequential assessments stay outside this prototype. An approval click alone does not authorise a prohibited use. Victorian guidance · Computer-use reference

Download the implementation blueprint

Technical source notes checked 8 October 2026. Capabilities and permitted uses must be confirmed for each deployment. RSI AI is an independent Melbourne venture; this scenario does not represent a government customer or endorsement.

Enable JavaScript to explore the cases. The whitepaper describes the workflows and their review boundaries..case-section[hidden]{display:block!important}

Prepared examples and local rules. No model request, customer message, real booking or external record update is made.

AI operating cost labKeep model usage and team subscriptions visible alongside the workflow.

BUSINESS CODING / SEPARATE BUDGETS

Seats for people.
API usage for agents.

Compare team seat fees alongside the selected token scenario. Plan limits and included coding tasks cannot be derived from API token prices.

Team membersClaude premium seatsSeat billingAPI workdays / month

Assume the eligible Claude Team API credit has been claimed.Off by default. Rollout, eligibility and linking apply; only Claude Platform API charges can be offset.

ChatGPT Business + Codex—USD / month · seats + selected API usage

Business seats: —
Gross API estimate: —
API credit assumed: $0.00

Official plan reference ↗

Claude Team + Claude Code—USD / month · seats + selected API usage

Standard + premium seats: —
Gross API estimate: —
Eligible credit applied: —

Official plan reference ↗

Loading the selected token scenario.

Seat references: ChatGPT Business and Claude Team Standard $20/user/month annually or $25 monthly; Claude Team Premium $100 annually or $125 monthly. Minimum two seats. Annual billing involves an annual seat commitment; displayed monthly equivalents are not monthly invoices.

Credit conditions & what is excluded

Claimed Team credits: $20 per Standard seat and $100 per Premium seat, pooled up to $500 per month. They expire each billing cycle and do not extend interactive coding limits. The estimator offsets only the Anthropic portion of the API workload, never OpenAI usage. Credit eligibility is an assumption, not your account balance.

This is a procurement planning illustration, not a bill or an equivalence between coding allowances. Seats and API may serve different activities. Taxes, tools, infrastructure, extra subscription usage and negotiated enterprise terms are excluded.

API credit eligibility and rollout ↗

MODEL ECONOMICS / 08 OCT 2026

Different agents.
Visible prices.

Official direct API list prices in USD per million tokens. Select a charge type and a prompt size to see the applicable tier.

Input token pricesUSD / 1M tokens · Standard

Full input per request includes instructions and cached texttokens

Claude Haiku 5.5Standard tier

$0.10

GPT-6 LunaStandard tier

$0.10

Claude Sonnet 5.5Standard tier

$2.00

GPT-6.1 SolStandard tier

$2.00

Same units and zero-based scale. Prices checked 8 October 2026.

Exact rates, context tiers & sources

Applicable Standard API rates, USD per million tokens
Model	Input	Cache write	Cache read	Output
Claude Haiku 5.5	$0.1	$0.125	$0.01	$0.5
GPT-6 Luna	$0.1	$0.125	$0.01	$0.5
Claude Sonnet 5.5	$2	$2.5	$0.1	$10
GPT-6.1 Sol	$2	$2.5	$0.1	$10

Haiku 5.5: higher rates above 100,000 input tokens. GPT-6 Luna and GPT-6.1 Sol: higher rates above 272,000 input tokens. The higher tier applies to the full request. Sonnet 5.5 retains its rates across its supported context window.

Anthropic cache writes here use the 5-minute tier. Cache eligibility, minimum prefix size and lifetime matter. These prices exclude tools, hosting, staff, tax, regional premiums and faster processing. Tokenisers differ; equal token counts do not establish equal task cost or quality.

Haiku 5.5 release & prices ↗Claude API prices ↗OpenAI API prices ↗

Full token simulator Editable workload and model assumptions

TOKEN LAB / LOCAL SIMULATION

Same workload.
Two execution paths.

Explore where an agent spends tokens, then change how it works. Every input, cache read and output stays visible.

Illustrative rates. USD only.
No API requests or measured results.

Start with an example

Quick controls

Tokens are pieces of text the model processes.

Standard API reference · checked 8 October 2026

Tasks per day

One task = one model request.

Context tokens / task

Text available before instructions. Long-context price tiers are applied.

Potential cache reuse%

Exact-prefix reuse, after cold writes.

Choose what changesSwitch a policy to see its effect.

Preparing the simulator…

Results update as you edit. Run the comparison to follow the four stages.

Baseline

Full context · general model

$9.48USD / day

01 · Workload240 tasksSame daily batch

02 · Context16.5K / taskFull prompt · no cache

03 · ModelGeneralAll tasks use general

04 · Output650 / taskExpected average

Four stages · swipe to follow the path →

Input3.96M

Cache write0

Cache read0

Output156K

Projected daily token account

Optimised

Your selected policies

$2.00USD / day

01 · Workload240 tasksSame daily batch

02 · Context10.1K / taskSelected context · cache reuse

03 · ModelCompact + generalIllustrative task policy

04 · Output450 / taskExpected average

Four stages · swipe to follow the path →

Input926.4K

Cache write455.52K

Cache read1.04M

Output108K

Projected daily token account

Modelled token-cost reduction— / dayBased on these assumptions

Token account— fewerSelection and expected output

Cache token account— read / — writtenCache tokens stay in the account.

Keep this scenarioAssumptions + exact totals. JSON, saved locally.

Token charges only. Routing and shorter output are strategy assumptions, not quality scores. Evaluate them on representative tasks. Flow lines show the calculation stages, not execution speed. Changing a value animates its previous and new totals.

Advanced assumptions Token shape, routing and optional prices

These are editable assumptions, not observed usage. Selection and routing use ordinary code in this model, with no additional model calls. Infrastructure and staff time are excluded.

Token shape & routing

Instruction tokens / taskContext retained (%)Cacheable context (%)Compact-model tasks (%)Expected output, no budgetExpected output, with budgetOutput ceiling, with budget

Edit illustrative rates USD per million tokens · editing switches to custom rates

General inputGeneral outputCompact inputCompact outputCache write × input priceCache read × input price

See the arithmetic Every category, rate and unrounded total

Each token category is priced separately. Every active model starts cold. Reuse requires exact prefixes within a valid cache window; all other prefixes are written again.

USD = Σ (tokens × category rate) / 1,000,000

Model / category	Tokens × USD rate	USD

Totals use unrounded values. The repeat setting is a potential hit share; one cold write per active model can reduce the realised share. Cache writes may make a cold run more expensive.

DESIGNED AROUND THE HANDOFF

Less reconstruction.
More shared context.

Each role receives a defined input and leaves a result the next person can inspect. Agents prepare the work; consequential decisions retain their own human gate.

Inside the product thesis ↗

One case, carried forward.

Changes to the brief, sources or draft flow through the shared case. Old approvals do not survive a material edit.

Intervene at the friction.

Gather missing context, compare constraints, name the evidence gap and prepare a clear handover. Keep the next responsible role visible.

Review the whole outcome.

Check quality, corrections, operating cost and the effort required to act on the result. A polished draft alone is not a completed job.

BUILT FROM MELBOURNE

Bring a workflow.
Define the next step.

RSI AI is developing practical agent workflows with a focus on the requests, records and decisions that service businesses handle every day.

Talk to the founder ↗The company ↗

PRODUCT & ARCHITECTURE

Different roles.
One shared workflow.

Make the handoff
the strongest link.

A Friday opening. A changed crew. A plan that holds together.

A café opening. One coordinated delivery.

The site visit is finished. The evidence handover isn’t.

Finish the job. Complete the handover.

The caller explains it once. The next officer gets the full context.

One call. A case the next officer can use.

A case in motion. Every intervention in view.

Contact and address update

Work within the allowance.

Retrieve the permitted evidence.

Keep the source and its limits.

Review actions and reusable skills.

Seats for people.
API usage for agents.

Different agents.
Visible prices.

Same workload.
Two execution paths.

Quick controls

Baseline

Optimised

Less reconstruction.
More shared context.

One case, carried forward.

Intervene at the friction.

Review the whole outcome.

Bring a workflow.
Define the next step.

The public whitepaper.

A roadmap with evidence.

Clear working boundaries.

Different roles.One shared workflow.

A Friday opening. A changed crew. A plan that holds together.

A café opening. One coordinated delivery.

The site visit is finished. The evidence handover isn’t.

Finish the job. Complete the handover.

The caller explains it once. The next officer gets the full context.

One call. A case the next officer can use.

A case in motion. Every intervention in view.

Work within the allowance.

Retrieve the permitted evidence.

Keep the source and its limits.

Review actions and reusable skills.

Seats for people.API usage for agents.

Different agents.Visible prices.

Same workload. Two execution paths.

Quick controls

Baseline

Optimised

Less reconstruction.More shared context.

One case, carried forward.

Intervene at the friction.

Review the whole outcome.

Bring a workflow.Define the next step.

The public whitepaper.

A roadmap with evidence.

Clear working boundaries.

Different roles.
One shared workflow.

Seats for people.
API usage for agents.

Different agents.
Visible prices.

Same workload.
Two execution paths.

Less reconstruction.
More shared context.

Bring a workflow.
Define the next step.