Retail-Store Frontline Copilot AI Agent

What this agent does

Prompts from the terms, not from a summary

Price cards, promotion terms and broadband labels come from the systems that publish them.

Field labels and formats are normalised, and each term keeps the published document it came from.

Reason

Counter prompts are assembled — the offer, its material restrictions and the comparison the advisor reads out.

Account and usage data enter a prompt only on the consent basis recorded for that customer.

Quoted terms bind to the line on the published card they came from; anything absent is marked unquotable.

Decide

Unlimited, free-device, lease and tier-comparison wording is flagged with the restrictions that must travel with it.

Anything billable stops for the customer's own recorded authorisation, captured apart from the pitch.

Out

The prompt, the terms it quoted, the restrictions shown and the advisor who used it are retained against the store record.

Write actions run only inside the approval boundaries agreed during implementation.

→ Product statement

The copilot proposes the words; the advisor stays the seller of record, and the carrier answers for the sale.

Example workflow

One counter conversation, end to end

AgentHuman

1Customer at the counterWalk-in enquiry, upgrade, activation or device trade-in

2Terms retrievedPublished price card, promotion terms, broadband label and eligibility conditions

3Prompt assembledThe offer, its restrictions and confidence

4Controls appliedCard-binding checks, restriction pairing, enjoined-claim flags and confidence threshold

No human action required

Stages 1 to 4 run unaided, and nothing is sold at any of them — the copilot is prompting, and the advisor's lane opens at the confidence gate.

5DecisionBranches at the confidence threshold

High confidence

Goes to the advisor to read out.

Low confidence

Adds a store manager read first.

Advisor decision

The prompt is held with the terms it quoted, its restrictions and the confidence.

Use · Amend · Send to store manager

Read out — customer authorises▼

6Store systems updatedOnly where write access and approval policy allow it

7Outcome evaluatedTerm accuracy, restrictions shown, authorisation captured and post-sale complaints

Amendments

Every advisor amendment is counted in the evaluation.

What should not run autonomously

Human approval stays in control

Outside the boundary — human approval required8 items

Recording the customer's authorisation to be billed.

Attaching a charge, add-on or line to an account.

Quoting a term that is not on the published card.

Standing in as the independent verifier of a sale.

Automation boundaryAgent acts unaided

✓Retrieve the published terms that apply to the offer in hand.

✓Surface the broadband consumer label in its prescribed format.

✓Assemble a comparison the advisor reads out, restrictions attached.

✓Hold at the point of authorisation, and keep the prompt.

Any write happens inside the boundaries agreed at implementation, never ahead of the advisor.

Deciding what a promotion's conditions actually are.

Judging whether a trade-in condition has been met.

Setting how prompts rank, and on what signal.

Changes to store rules, terms or approval thresholds.

Example output

One counter prompt, annotated

Everything the copilot puts in front of an advisor is attached to the term it came from.

Copilot output · single counter promptIllustrative example

Offer

Prompted line

Stated price

Source of term

Confidence

Restrictions

Handset upgrade

Data allowance doubles for the promotional period on this plan

$45.00 / month

Published price card

93%

Speed threshold and end date

As receivedTaken from the published price card and the promotion terms — nothing on this side is worded by the copilot.

Published terms used Price-card line Promotion terms Broadband label

Why this wordingIt is the card's own sentence, carrying the threshold that qualifies it.

ActionUseAmendSend to store manager

What the score decidesBelow the configured threshold the prompt picks up a manager read before it reaches the advisor.

Value

Where AI adds value

The same four claims, placed at the point in the workflow where each one applies.

Where the value landsValue 01 – 04

Every counter conversationFrom the published terms

03Prompting

Prompt from the card

Draw on the published price card, the promotion terms and the store rules configured for the retailer.

01Approved path

Prompt the advisor, not the sale

Routine offers arrive already worded, with their restrictions attached.

02Human review

Send the read to the risky prompts

Enjoined-claim flags and low-confidence prompts are marked, so the manager's read starts where risk concentrates.

04Build an evidence trail

The prompt, the terms it quoted and the advisor who used it stay on the store record.

Integrations

Typical integrations

Five system groups connect to the same agent. Which of them are in scope is decided in discovery.

Product and pricingProduct catalogue · price cards
Promotion and offer engines

Point of saleRetail POS · order capture
Activation and provisioning

Customer systemsCRM · account lookup
Billing and entitlements

Agent

Retail-store frontline copilot

Reads the terms
Prompts the advisor
Holds for authorisation

Store operationsWorkforce · queue management
Trade-in and device grading

Observability & evaluationOpenTelemetry · Langfuse
Supported monitoring/evaluation sources

Integration availability depends on the client's existing systems and API access.

Agent controls

Six layers between the model and the counter

The controls are nested. What survives all of them appears in the map underneath.

L6 · Outermost — last line of defenceInward → L1 · closest to the model

L6Rollback / safe modePull the copilot back to term lookup when evaluation or production signals degrade.Roll back

L5Version monitoringTrack model, prompt, term-set and store-rule configuration changes.Track

L4TraceabilityRecord the terms retrieved, the prompt, the restrictions shown and the advisor.Record

L3Advisor decisionHold prompts for a named advisor; it governs what is said, not whether the sale was right.Gate

L2Term guardrailsTest prompts against the published card and the store rules; a failure returns the prompt.Restrict

L1Confidence thresholdsRoute low-confidence prompts to a manager read before the advisor sees them.Require review

Model corePrompt produced — the offer, its restrictions, the comparison and confidence

L1 – L2Test whether a prompt may stand

L3Puts the decision in an advisor's hands

L4 – L5Show what the prompt was drawn from

L6Holds offers for the manager when signals degrade

How Nestack evaluates it

Evaluate the counter workflow — not only the sentence the advisor read.

Coverage runs the whole depth of the workflow, and every layer is cut by slice.

Surface — the words the customer hears

Depth of coverage ▼

E1Final-output evaluationDid every quoted term match the published card it came from?

E2Step-level evaluationDid the copilot use the right card, promotion terms and store rules?

E3Tool evaluationDid it read the correct account and write to the correct order?

E4Confidence calibrationDo low-confidence prompts actually attract more advisor amendments?

E5Slice evaluationHow does performance change across specific offer types?

E6Business outcomeHow many prompts needed an amendment or a correction after the sale?

Floor — the outcome the carrier answers for

Failure modes

Where each failure originates in the agent

Seven failure modes, placed at the stage each one begins.

Agent lifecycleDirection of processing →

01 · Retrieval1 mode

NX-03

Superseded price card

The prompt quotes terms the business has since replaced.

Stage gathersPrice cards, promotion terms, labels and store rules

02 · Reasoning2 modes

NX-04

Restriction dropped

A claim is prompted without the threshold that qualifies it.

NX-06

Paraphrase for a term

A published condition is restated in the copilot's own words.

Stage proposesThe wording the advisor is prompted with

03 · Tool / write2 modes

NX-02

Charge before authorisation

An add-on is prompted where no authorisation was recorded.

NX-05

Authorisation inside the pitch

Consent is captured on the same screen as the inducement.

Stage writesOnly where write access and approval policy allow it

04 · Output1 mode

NX-01

Comparison across unlike tiers

Two dissimilar service tiers are set side by side.

Stage returnsThe words the advisor reads to the customer

05 · Change / Version1 mode

NX-07

Silent ranking regression

A model or rule change lets commercial signals order the prompts.

Stage tracksModel, prompt, term set and ranking config

Sev-1 · a charge attached without authorisation Sev-2 · a wrong term reaches the customer Sev-3 · terms degrade, prompt routes to review

See the complete Telecom failure-mode catalogue →

Affected slices

Overall prompt quality can hide one offer type

Aggregate term accuracy reads acceptable across the estate, while free-device promotions carry most of the amendments on a thin share of conversations. Nestack reports the amendment rate by slice, not only in total.

Slice performance — reported separately, not only in aggregateIllustrative example

Slice	Failure rate	Lift	Status
Free-device promotions	4.4%	3.8×	Review
Unlimited-plan comparisons	3.3%	2.9×	Review
Device trade-in and lease offers	2.3%	2.0×	Watch
Single-line SIM-only sales	0.8%	0.7×	Normal

Bar: advisor-amendment lift vs. the SIM-only baseline · scale 0–4.0× · tick marks the 2.0× review threshold 2 of 4 slices over threshold

Evidence-linked improvement

A cycle ends in a case, not an apology

Explaining a bad sale does not close the cycle. A standing regression case does, and that suite is what the next counter conversation is measured against.

Improvement cycle · five stagesSwitchback — the path turns at Improve and returns at Learn

01Detect

Advisor amendments rise in one offer type.

02Diagnose

If the amendments cluster on one card, a store manager reads the prompts and the terms behind them until the cause narrows to one.

03Improve

The fix is versioned against the prompts that produced it.

04Verify

The release stops until every affected case is passing.

05Learn

The case stays in the suite, and the store playbook is amended.

Learn → DetectThe return edge. Whatever is detected next is detected against a longer suite.

Typical build scope

Twelve workstreams across six weeks

The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, terms, prompting workflow, evaluation, integration, then production validation and handover.

Workstream Week 1Week 2Week 3Week 4Week 5Week 6

01Counter workflow discovery and automation-boundary definition.

02Point-of-sale and catalogue assessment.

03Published-term, label and store-rule mapping per market.

04Term ingestion and normalisation.

05Prompt assembly and card binding.

06Confidence scoring and manager routing.

07Advisor decision workflow.

08Point-of-sale and billing integration.

09Term-accuracy regression cases.

10Guardrails and authorisation controls.

11Prompt-trail instrumentation.

12Deployment, documentation and Agent Care handover.

12 workstreams · 6 weeks · bar shows the weeks a workstream is active — several run in parallel Final scope and sequence confirmed in discovery

Engagement tiers

What each tier includes

Rows are the capabilities named in each tier's scope. Higher tiers include everything below them.

Capability✓ in scope · — not at this tier PilotOne store, one market ProductionProduction store estate AdvancedMultiple regions / brands

Introduced at Pilot

Prompting from your published terms✓✓✓

Advisor decision✓✓✓

Prompt-accuracy baseline✓✓✓

Introduced at Production

Reporting by store and line—✓✓

Review workflow in your systems—✓✓

Approved write-back—✓✓

Point-of-sale integration—✓✓

Introduced at Advanced

Multi-market term and label rules——✓

Multi-stage store approvals——✓

High counter volume——✓

Multi-region store controls——✓

Build price From $5,000 From $8,000 Custom quote

Final build priceConfirmed after discovery based on integrations, workflow complexity, counter volume, approval controls and deployment requirements.

Separate from buildBuild pricing is separate from recurring Agent Care, which covers managed monitoring, evaluations, incidents and verified improvements after launch.

What we need from you

What you bring, and what we build with it

Each input maps to a piece of build scope and a week in the delivery timeline.

You bringWe build with it

01Your published price cards and promotion terms → Term ingestion and card bindingWeek 1

02Representative counter conversations → Prompting baseline, comparison assembly and term bindingWeek 2

03The claims your compliance team already refuses → Enjoined-claim mapping and the automation boundaryWeek 1

04Access to relevant APIs, feeds or exports → Point-of-sale, catalogue and billing assessment, then integration setupWeek 2

05Prompts you would not want acted on → Term cases and the evaluation suiteWeek 4

06Where a prompt must stop and wait for the customer → Confidence scoring, manager routing, guardrails and authorisation controlsWeek 3

07Named store managers to release prompts → Advisor decision workflow, then pilot and production validationWeeks 5–6

Nothing else is required Deployment, documentation and Agent Care handover are ours.

Delivery timeline

Four phases across six weeks

Phases occupy the weeks the work really needs, so week 5 runs evaluation and launch side by side.

Phase W1W2W3W4W5W6

Discovery W1

Build W2 – W3

Evaluate W4 – W5

Pilot & Launch W5 – W6

Week focus W1Counter workflow discovery, term mapping and the automation boundary W2Source integration and the prompting baseline W3Prompting workflow, confidence logic and authorisation controls W4Evaluation suite, enjoined-claim checks and failure-mode testing W5Point-of-sale integration, a pilot store and targeted corrections W6One trading period run under the store team, then handover

Reading the bandEach bar covers only the weeks its work is named in. The fifth week doubles because the work genuinely overlaps.

At the end of W6The period closes validation and Agent Care picks up the monitoring.

DurationSix-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.

Next step · Telecom AI agent

Build a retail copilot around the terms you already publish.

Start at the counter, with the sale as it actually runs. You bring the price cards, the promotion terms and whoever signs off what an advisor may say; we set the automation boundary.

Discuss this agent →

Nestack Agents · Retail frontline copilotAGT-TL-06 · Agent Care available after launch