Food & Beverage AI agent · Drive-thru voice
Drive-Thru Voice-Ordering AI Agent
Take the order at the lane as speech and not as a voiceprint, hand any allergen question to the shift team, and report the human-assist rate as a measured number.
- Typical delivery
- 4–6 weeks
- Deployment
- Your stack
- Speech, not ID
- Allergen held
- After launch
- Agent Care
What this agent does
Takes the order, never the voiceprint
In
- 01
When a car reaches the board, the disclosure runs before anything is captured, on signage and aloud.
- 02
Where the lane is captured, it is captured as speech to text, with no speaker template built or stored.
Reason
- 03
When the order is spoken, items are matched to the menu configured for that restaurant and daypart.
- 04
Where an upsell is offered, nothing joins the order without an explicit yes from the customer.
- 05
When the board carries calorie declarations, the agent stays inside the combinations the board declares.
Decide
- 06
Where a question turns on allergens or cross-contact, the lane stops and a person takes it over.
- 07
When confidence drops or the lane stalls, the order goes to the shift team and the handover is logged.
Out
- 08
Where audio is retained, it is kept to the configured schedule and never trains a shared model.
- 09
When write actions run, they run inside the approval boundaries agreed during implementation.
Product statement
The agent takes the order; the crew handles the allergen question and every lane it hands over, and the operator answers for what the customer was served.
Example workflow
One car, greeting to window
- 1Car at the boardLane signage and the audible disclosure, before the first utterance is captured
- 2Order capturedItems, modifiers and quantities, each matched to the menu on the board
- 3Order assembledItems, modifiers, flags and confidence
- 4Controls appliedMenu and price checks, upsell-consent checks, allergen held-act checks, disclosure rules by market and confidence threshold
No human action required
Stages 1 to 4 run unaided, and no allergen question is answered at any of them — the agent is taking an order, and the crew's lane opens at the confidence gate.
Goes to the crew to confirm.
Adds a shift-manager read first.
Crew confirmation
The order is held on screen with its modifiers, its flags and the confidence.
Confirm · Correct · Hand to the manager
Confirmed — sent to the kitchen
- 6Point of sale updatedOnly where write access and approval policy allow it
- 7Outcome evaluatedCrew corrections, handover reasons, remakes at the window and complaints raised after the car left
Corrections
Every crew correction is counted in the evaluation.
What should not run autonomously
Human approval stays in control
Outside the boundary — human approval required
8 items
- Telling a customer an item is free of an allergen.
- Judging cross-contact in a fryer or on a surface.
- Identifying a returning customer by their voice.
- Adding an item the customer did not agree to.
Automation boundary
Agent acts unaided
- ✓Take the spoken order and match it to the board's menu.
- ✓Play the disclosure before the first utterance is captured for the named owner.
- ✓Hold the lane for a person when the question turns medical.
- ✓Keep audio to the configured schedule, out of shared training.
Any write happens inside the boundaries agreed at implementation, never ahead of the confirmation.
- Publishing an autonomy rate no measurement supports.
- Declaring calories for a combination off the board.
- Deciding a complaint at the window is settled.
- Changes to retention, notice or handover rules.
Example output
One order, annotated
Everything the agent captured is attached to the notice that ran before it.
Lane output · single order
Illustrative example- Order
- Peak-hour order
- What the agent captured
- Two combinations and a modifier, read back before the window
- Items captured
- 4 items
- Notice given
- Board sign and audible pre-roll
- Confidence
- 92%
- What happens next
- Crew confirmed at the window
As received
Taken from the lane audio as transcribed — no speaker template is built, matched or stored.
Source records used
- Board menu and prices
- Lane disclosure log
- Transcript of the order
Why it stops here
The lane cannot see the fryer or the prep surface and a person still decides.
What the score decides
Below the configured threshold the order picks up a shift-manager read first.
Action
ConfirmCorrectHand to the managerValue
Where AI adds value
The same four claims, placed at the point in the workflow where each one applies.
Where the value lands
Value 01 – 04
Every car
From the board menu
Take from the lane
Draw on the menu, the prices, the daypart and the notice rule for the state the restaurant sits in.
Take the order, not the voice
Routine orders arrive captured, priced and matched to the board.
Send the crew the ones that stall
Allergen questions and stalled lanes are handed over, so crew time goes where the harm and the queue sit.
Build an evidence trail
The order, the disclosure given before it and the person who stepped in stay on the transaction.
Integrations
Typical integrations
Five system groups connect to the same agent. Which of them are in scope is decided in discovery.
Point of sale
- Oracle Simphony
- NCR Aloha
- Toast
- PAR Brink
Lane hardware
- Headsets
- outdoor speakers
- Order confirmation boards
Menu and pricing
- Menu management
- dayparts
- Calorie declarations on the board
Agent
Lane order taking
- Gives the notice
- Takes the order
- Hands to the crew
Operations and reporting
- Lane audio
- order display
- Point-of-sale
- kitchen display
Observability & evaluation
- OpenTelemetry
- 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 window
The controls are nested. What survives all of them appears in the map underneath.
L6 · Outermost — last line of defence
Inward → L1 · closest to the model
- L1 – L2
- Test whether an order may stand
- L3
- Puts the lane in the crew's hands
- L4 – L5
- Show the order and the disclosure behind it
- L6
- Holds the lane for a person when signals degrade
How Nestack evaluates it
Evaluate the ordering workflow — not only the order that reached the window.
Coverage runs the whole depth of the workflow, and every layer is cut by slice.
Surface — the order the customer receives
- E1Final-output evaluationDid the captured order match what the customer actually asked for?
- E2Step-level evaluationDid the agent use the right menu, price and daypart?
- E3Tool evaluationDid it write the correct items to the correct check?
- E4Confidence calibrationDo low-confidence orders actually attract more crew corrections?
- E5Slice evaluationHow does performance change across specific lane conditions?
- E6Business outcomeHow many orders needed a correction, or a remake after the window?
Floor — the order the operator answers for
Failure modes
Where each failure originates in the agent
Seven ways a lane goes wrong, placed by stage.
Agent lifecycle
Direction of processing →
01 · Retrieval
1 mode
- BM-03
Stale menu read
An item or price is read from a menu the board has moved past.
Stage gathers
Lane audio, board menu, prices and daypart rules
02 · Reasoning
2 modes
- BM-04
Silent yes assumed
An upsell is added without an explicit yes from the customer.
- BM-06
Allergen answered
The lane assesses a question that belongs with a person.
Stage proposes
Items, modifiers, flagged questions and confidence
03 · Tool / write
2 modes
- BM-02
Order sent early
A flagged order reaches the kitchen before the crew sees it.
- BM-05
Item duplicated
The same item is added twice when a customer repeats it.
Stage writes
Only where write access and approval policy allow it
04 · Output
1 mode
- BM-01
Notice not played
Capture begins before the notice the lane is configured to give.
Stage returns
The order the crew confirms and the kitchen makes
05 · Change / Version
1 mode
- BM-07
Silent capability drift
A model or vendor change adds speaker features nobody asked for.
Stage tracks
Model, prompt, menu and notice version
- Sev-1 · a speaker feature ships
- Sev-2 · a wrong order reaches the window
- Sev-3 · source degrades, lane routes to the crew
Affected slices
The lane average can hide one daypart
Read these against the base rate. Most orders are short, quiet and unmodified, and that population is what sets the headline; the peak-hour queue is a small share of the orders and a large share of the corrections. Nestack reports the slice.
Slice performance — reported separately, not only in aggregate
Illustrative example| Slice | Failure rate | Lift | Lift vs. threshold | Status |
|---|---|---|---|---|
| Peak-hour queues | 4.5% | 3.9× | Review | |
| Heavy background noise | 3.0% | 2.6× | Review | |
| Long customised orders | 2.3% | 2.0× | Watch | |
| Short unmodified orders | 0.8% | 0.7× | Normal |
Bar: crew-correction-rate lift vs. short-order 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 standing test
Explaining a wrong order does not close the cycle. A standing regression case does, and that suite is what the next order taken is measured against.
Improvement cycle · five stages
Switchback — the path turns at Improve and returns at Learn
- 01
Detect
Crew-correction rate rises in a lane slice.
- 02
Diagnose
If the pattern holds across a daypart, the transcripts and the board behind them are read together until the cause narrows to one.
- 03
Improve
The fix is versioned against the orders that produced it.
- 04
Verify
The release stops until every affected case is passing.
- 05
Learn
The case stays in the suite, and the lane playbook is amended.
Learn → Detect
The return edge. The next order meets a suite one case longer.
Typical build scope
Twelve workstreams across six weeks
The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, sources, ordering workflow, evaluation, integration, then production validation and handover.
Workstream
Week 1Week 2Week 3Week 4Week 5Week 6
- 01Lane workflow discovery and boundary definition with operations.Week 1
- 02Point-of-sale and menu source assessment.Week 1, Week 2
- 03Notice, retention and handover mapping for each market.Week 1, Week 2
- 04Lane capture and transcription setup.Week 2
- 05Order assembly and menu binding.Week 2, Week 3
- 06Confidence scoring and handover routing.Week 3
- 07Crew confirmation workflow.Week 3, Week 4
- 08Point-of-sale and lane integration.Week 5
- 09Allergen and disclosure cases.Week 4, Week 5
- 10Guardrails and handover controls.Week 4
- 11Order-trail instrumentation.Week 5, Week 6
- 12Deployment, documentation and Agent Care handover.Week 6
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 restaurant, one lane | ProductionProduction point of sale | AdvancedMultiple markets / brands |
|---|---|---|---|
| Introduced at Pilot | |||
| Order taking on your board menu | ✓ | ✓ | ✓ |
| Crew confirmation | ✓ | ✓ | ✓ |
| Order-accuracy baseline | ✓ | ✓ | ✓ |
| Introduced at Production | |||
| Reporting by daypart | — | ✓ | ✓ |
| Handover workflow in your systems | — | ✓ | ✓ |
| Approved write-back | — | ✓ | ✓ |
| Point-of-sale integration | — | ✓ | ✓ |
| Introduced at Advanced | |||
| Multi-market notice rules | — | — | ✓ |
| Multi-stage operations approvals | — | — | ✓ |
| High lane volume | — | — | ✓ |
| Multi-market lane controls | — | — | ✓ |
| Build price | From $5,000 | From $8,000 | Custom quote |
- Final build price
- Confirmed after discovery based on integrations, workflow complexity, lane volume, approval controls and deployment requirements.
- Separate from build
- Build 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 bring
We build with it
01Your board menu and the dayparts it runs on
→Menu ingestion and item mapping
Week 102Lane recordings you already keep
→Order-taking baseline, transcription setup and item binding
Week 203Your notice, retention and consent posture
→Notice, retention and handover mapping by market
Week 104Access to relevant APIs, feeds or exports
→Point-of-sale, menu and lane assessment, then integration setup
Week 205Orders you would not want served
→Allergen cases and the evaluation suite
Week 406What the lane may never answer on its own
→Confidence scoring, handover routing, guardrails and lane controls
Week 307Shift managers to take the handovers
→Crew confirmation workflow, then pilot and production validation
Weeks 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
W1
W2
W3
W4
W5
W6
- DiscoveryW1
- BuildW2 – W3
- EvaluateW4 – W5
- Pilot & LaunchW5 – W6
Week focus
W1Lane workflow discovery, notice mapping and the automation boundary
W2Menu integration and the order-taking baseline
W3Ordering workflow, confidence logic and handover controls
W4Evaluation suite, allergen guardrails and failure-mode testing
W5Point-of-sale integration, pilot lanes and targeted corrections
W6One trading period run under the shift manager, then handover
- Reading the band
- Each band sits on the weeks it truly needs, so the fifth holds evaluation and launch together.
- At the end of W6
- The period closes validation and Agent Care picks up the monitoring.
- Duration
- Six-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.
Next step · Food & Beverage AI agent
Build a lane agent that takes the order and leaves the voice alone.
The aural notice rule that binds a lane is Maine's, not California's bot law, which reaches online accounts only — so the notice is configured by state rather than assumed. If you can give us one lane, a board menu and a shift team to hand over to, we will measure the human-assist rate and publish what it is instead of an autonomy claim.
- Typical delivery
- 4–6 weeks
- Deployment
- Your lanes
- Allergen held
- Speech, not ID
- After launch
- Agent Care available
Nestack Agents · Lane ordering
AGT-FB-06 · Agent Care available after launch