AI agents in retail are moving beyond chatbots. They search assortments, negotiate constraints, recommend products, monitor shelves, adjust promotions, resolve service issues and coordinate operational work. This gallery maps 60 agents across the retail value chain and shows the control, evidence and escalation each one needs before it can be trusted with real decisions.
Not every AI feature is an agent. An agent takes action toward a goal with some degree of independence — it decides, not just suggests. A recommendation engine ranks products; a shopping agent decides what to put in a basket and why. The difference matters because the moment software acts on its own, the design problem shifts from accuracy to trust.
Retail already has the agents. Trust decides which ones you let act.
Twenty retail playbooks. Sixty agents. The build list is the easy part — every category, store-ops, and digital team can already name the agents they want.
The hard part, the part that decides whether any of them ever ships, is the one question a gallery never answers:
which of these can you let act without a human in the loop, and which would you never?
This is that gallery — the full sixty — but scored on the axis retail actually cares about: consequence. Every agent below carries a heat reading and the control posture it earns. Because in retail, "the agent did it autonomously" is a feature when it drafts a report and a liability when it rewrites a live price.
A list of sixty retail agents is a wish, not a roadmap. Anyone can write it. What separates the teams that ship from the teams that pilot forever is not the model, the data, or the vendor — it's whether they've decided, action by action, who pulls the trigger.
Retail makes that question unusually sharp. The actions hiding inside these agents are some of the hottest in any industry. They move price. They move money. They touch live inventory. And they reach customers and suppliers by name. A "SKU diagnostics" agent that drafts a consolidation report and a "competitor price match" agent that rewrites your shelf-edge prices show up in the same paragraph of every vendor deck — and they could not be more different in what they cost you when they're wrong. One earns a shrug. The other earns a margin investigation and an angry call from a regional manager.
So auxfirst reads this gallery the way we read every agent system: not by what it can do, but by what it's allowed to do unsupervised. That reading is the product.
We score every agent action with auxfirst field model № 03, the Action Heat Ladder. Most teams argue about agent autonomy with vibes — "that one feels risky." The ladder replaces the vibe with five questions you can answer about any action, in any function.
An action is as hot as its hottest dimension. No averaging. Four cool dials never buy back one hot one — a perfectly reversible, narrow, internal action that signs a contract is still a contract.
Each band maps to exactly one control posture:
| LOW | auto-run | Let it run; audit on incident. |
| LOW-MED | sampled review | Runs autonomously; a human audits a slice on cadence. |
| MEDIUM | propose + approve | The agent proposes; a named human approves before anything happens. |
| HIGH | named approver | A specific, accountable person signs off every time. |
| CRITICAL | human executes | The agent prepares; a human does the deed. |
"Push a 40% markdown to every store, live, tonight."
Reversibility 3 (you can re-price up, but the units sold at 40% are gone) · Blast radius 4 ("every store") · Exposure 3 (public shelf price, customers, a headline if it's an error) · Commitment 3 (margin given away) · Authority 4 (writes live pricing systems)
CRITICAL A human executes. Even if four dials were cool, "every store, live" is enough on its own.
MEDIUM Now the same verb, cooled by design: "draft a phased markdown schedule for the buyer to approve, capped at 40%, one category, staged." Reversibility, blast, and authority all drop — propose + approve. Same agent. Different design. The difference is the whole game.
Forbidding hot actions is easy. Cooling them so you keep the leverage is the work. Five design moves do most of it in retail, and you'll see them named throughout the gallery:
The heat band tells you how much oversight an action needs. But a safe agent isn't automatically a trusted one. Below the posture, every agent in this gallery — cool or critical — needs the same five things before a merchant, a manager, or a customer will rely on it:
Map every agent up the Trust Architecture as it earns it — Functional → Contextual → Judgment → Advocacy. Promote the ones that earn trust; demote the ones that burn it.
The twenty playbooks cover the full sweep of modern retail. They group into five operating pillars — and each pillar has its own trust signature, a characteristic shape of risk that tells you how aggressively you can automate it.
1 · Merchandising & Brand Strategy. Mostly analytical and proposal-shaped — diagnose, brief, simulate, recommend. The heat is low, which hides the real trap: not runaway action but quiet authority. A consolidation list nobody questions becomes policy. Keep these honest with Confidence Cues and an Escape Hatch on every recommendation. The one hot spot is co-op billing — the moment an invoice leaves the building, money and an outsider are in play.
2 · Pricing, Margin & Supplier Relationships. The hottest pillar in retail. Almost every action moves price or money, and most reach a customer or a supplier by name. This is where "the agent did it autonomously" becomes a chargeback dispute or a price-error headline. Default posture is high: propose-and-approve at minimum, a named approver for anything that writes a live price or issues a claim. Margin-floor caps, draft-only invoices, and sampled audits are non-negotiable here.
3 · Physical Stores & Workforce Operations. Heat splits cleanly. Anything pointed at SOPs, audits, training, and internal alerts is cool and ships fast. Anything pointed at people — building schedules, flagging staff for loss prevention — is hot for a reason no dashboard shows you: a false positive has a human cost. Never let an agent conclude guilt or finalize a roster alone. Loop In Experts; Clarify Before Commit.
4 · Omnichannel & Digital Experience. The deceptive pillar. Most actions are reversible — you can roll back a ranking, a banner, a synonym — which tempts teams to over-automate. But reversible isn't the same as low-stakes when the audience is every customer and the lever quietly steers margin. The move here is to embrace autonomy and engineer the cooling: allowlisted components, instant rollback, discount caps, sampled review. Reversibility is the thing you exploit, not the excuse you use.
5 · Governance & Performance Management. Almost entirely cool. These agents read, diagnose, draft, and chase; they never act on the outside world. Auto-run the lot. The only discipline that matters is Transparency, Tapered — show the reasoning behind every "why the KPI moved" before anyone trusts it to brief the executive room.
Jump straight to the category you own — each links into the gallery below.
| Category | What these agents do | Playbooks | Agents |
|---|---|---|---|
| Merchandising & brand strategy | Assortment and SKU rationalization, category insight, line review, content enrichment, own-brand development | 4 | 12 |
| Pricing, margin & supplier relationships | Markdown and promotion planning, competitor response, margin guardrails, supplier negotiation and terms | 3 | 9 |
| Physical stores & workforce operations | Task allocation, staffing and scheduling, shelf and planogram compliance, incident handling, store standards | 5 | 15 |
| Omnichannel & digital experience | Shopping and discovery, customer service and returns, loyalty and next-best action, fulfilment and delivery exceptions | 7 | 21 |
| Governance & performance management | Performance reporting, policy enforcement, audit trails and the oversight layer above every other agent | 1 | 3 |
↔ scroll table on mobile · 20 playbooks · 60 agents in total
Sixty agents, regrouped under the five pillars. Each carries its heat reading and the posture it earns. Where an action is born hot, we name the de-escalator that cools it enough to ship.
Trust signature: low heat, quiet authority. The risk isn't action — it's a recommendation no one challenges.
PlaybookAssortment Optimization & SKU Rationalization
Goal: maximize category productivity and cut the low-performing tail — without losing the customer who came for the item you dropped.
Scores every SKU on margin, units, and turns, then drafts consolidation proposals for the merchant to weigh. Turns the quarterly tail-cull from a spreadsheet marathon into a standing recommendation.
Sampled review. It only ever drafts; the merchant approves any delete. Draft-only output (D1) keeps a high-blast action — removing a product — cool.
Reads basket and loyalty data to answer the question that makes or breaks a delist: will those sales transfer to what's left, or walk out the door? Simulates the revenue impact before anyone touches the range.
Auto-run. Read-only simulation; it models, it doesn't act.
Runs pilot stores against controls during a range change, tracking sales transfer, margin, and out-of-stocks — and flags a pilot that's drifting before it costs a season.
Auto-run. Internal alerts and a scorecard; nothing leaves the building.
PlaybookMerchandising Calendar & Line Review
Goal: hold the cross-functional calendar together and walk into every line review already briefed.
Watches design, sourcing, QA, and vendor-pitch milestones and raises the alarm when a task slips toward the launch date — auto-rescheduling the dependent, non-critical work.
Sampled review. It writes to your PM tool, but only reshuffles non-critical items (D6 allowlist) and every move is reversible (D3).
Pulls multi-year sales, margin, returns, and supplier performance into a briefing dossier and lands it in the merchant's inbox two weeks before the review — so the meeting starts at insight, not data-gathering.
Auto-run. Read-only data, internal audience.
Extracts cost, lead time, MOQ, and certifications from vendor bid PDFs and builds a like-for-like scorecard against your category margin target.
Auto-run. It ranks and recommends; the award stays human.
PlaybookPrivate Label Strategy & Launch
Goal: grow the own-brand portfolio to its margin target without breaking the co-packer relationship.
Scans categories for the price-and-margin gap between national brands and the entry tier — the signature of a private-label opening — and builds the development backlog.
Auto-run. Internal opportunity list.
Parses co-packer bids into fully-loaded cost — materials, production, freight — against your target, and surfaces the most cost-effective, contract-compliant partner.
Auto-run. Recommendation only.
Tracks velocity, repeat rate, and review scores on a new own-brand line against its business case, and calls for a promotional boost when launch milestones slip.
Auto-run. Internal tracking and recommendation.
PlaybookPromotion Strategy & Promo Planning
Goal: turn promotions into investments you can measure, not discounts you hope work.
Computes true promotional ROI — net of cannibalization, the margin handed to full-price buyers, and pull-forward — and grades every campaign repeat, modify, or kill.
Auto-run. Internal analysis and a grade.
Checks that vendor-funded promotions actually ran as contracted, then prepares the co-op claim for finance.
Named approver. The moment a claim leaves the building it's money and an outside party (escalators E1, E2). Draft-only output (D1) cools it: the agent prepares, finance files.
Reads the promo calendar for clashing offers, shipping bottlenecks, and margin-draining overlaps, and recommends the fix before customers feel the fatigue.
Auto-run. Internal recommendation.
Trust signature: the hottest pillar in retail. Price, money, and a named counterparty are in almost every action. Default to high.
PlaybookMarkdown Optimization
Goal: get the timing and depth right so you recover value and clear stock without renting the sale.
Tracks weekly sell-through against plan and flags the slow-movers early — with a recommended first markdown date, while there's still season left to act.
Auto-run. Internal alert and recommendation.
Estimates price elasticity to recommend the depth that actually maximizes gross margin — 20 vs 40 — and proposes a phased schedule for sign-off.
Propose + approve. It proposes a price action; a buyer approves before anything moves.
Checks — via shelf photos or POS price scans — that marked-down stock is actually priced down and moved to clearance, scores each store, and nudges managers on missing tags.
Sampled review. It nudges staff; no pricing or customer action is taken autonomously.
PlaybookRetail Pricing Architecture
Goal: govern price tiers, zones, and rules to protect both margin and price image.
Watches competitor prices and recommends a match or a deliberate differential against your margin floor — adjusting live e-commerce prices, or flagging the change for a store manager.
Named approver. Writing a live, public price is production, exposure, and commitment at once. Cool it with margin-floor hard caps (D4) and sampled audits (D5) — and notice the design fork: "flag for the manager" is a cool action; "auto-adjust the site" is a hot one. Choose deliberately.
Groups stores into price zones by local income, competitor density, and elasticity, and proposes zone updates that capture margin you're leaving on the table.
Propose + approve. An annual, reviewable proposal — not a live write.
Guards the price perception of your high-visibility KVIs while recovering back-margin on everything else, proposing changes that keep KVIs inside your competitive index.
Propose + approve. It proposes price moves; pricing signs them off.
PlaybookVendor Terms & Trade Spend Optimization
Goal: stop value leaking out of supplier agreements — audit compliance, capture rebates, enforce SLAs.
Tracks year-to-date volume against tier thresholds and tells the buyer exactly when a rebate is in reach — "500 more cases of SKU X by Friday unlocks 3%."
Auto-run. Internal alert; the order decision stays human.
Reconciles vendor invoices, logistics fees, and receiving against contract terms to catch overcharges and unpaid allowances — and drafts the dispute letter to vendor accounting.
Named approver. A dispute letter is an external, money-bearing claim (E1, E2). Draft-only (D1) keeps it cool: the agent writes, a buyer sends.
Tracks on-time-in-full against your SLA policy, calculates the penalties, surfaces chargebacks, and arms buyers ahead of contract renewals.
Named approver. Chargebacks bind money and the supplier relationship and are hard to walk back. Draft-and-review (D1, D5) before any penalty is issued; never auto-fire a chargeback.
Trust signature: heat splits. Process and knowledge agents are cool and ship fast. Anything pointed at people is hot — a false positive has a human cost.
PlaybookNew Store Opening (NSO)
Goal: land the launch on time across thousands of dependent tasks.
Tracks the critical path — permits, construction, electrical — and drafts the emergency status update to regional ops the moment a high-priority task slips.
Auto-run. Internal drafts and tracking.
Coordinates shelving, registers, signage, and opening inventory against construction status — flagging shipping delays and chasing carriers for fresh ETAs.
Propose + approve. Internal flags are cool; reaching out to a freight carrier touches an outside party, so route external contact through a human.
Reads local demographics and competitor footprints to stand up geo-targeted social and search campaigns timed to the opening — allocating budget to localized ad groups.
Named approver. It spends money in public (E1, E2). Hard budget caps (D4) and sampled review (D5) let it run the mechanics while a human owns the spend.
PlaybookPlanogram & Space Optimization
Goal: allocate shelf space to productivity, not habit — and prove the shelf matches the plan.
Compares linear-foot allocation to sales and margin contribution and flags the categories carrying space they haven't earned — or starving for facings.
Auto-run. Internal diagnostic.
Tailors the corporate planogram to each store's real fixtures and local demand, adjusting facings to fit without a designer in the loop.
Propose + approve. It writes a production artifact across many stores (blast radius), so approve the policy and sample the output (D5, D6) before it sets execution everywhere.
Matches shelf photos — from staff or robots — against the master planogram to score execution and tasks associates with the fixes.
Sampled review. It assigns corrective tasks internally; audit a slice to keep the vision model honest.
PlaybookStore Labor Scheduling & Productivity
Goal: match labor to demand — and to non-selling work — without breaking budget or the law.
Forecasts hourly traffic and sales — weather included — and outputs the staffing the floor actually needs, straight into the scheduling tool.
Sampled review. It writes a recommendation, not a roster.
Builds shift schedules that respect the labor budget, local rest rules, and stated preferences, and manages shift-swap approvals.
Propose + approve. Schedules bind people, cost, and legal compliance; the manager signs. Encode the budget and the law as hard caps (D4) the agent cannot exceed.
Pushes the right task at the right moment — a BOPIS queue, a stockout to replenish, a price tag to change — to associate devices during non-peak hours.
Sampled review. Internal task dispatch within guardrails.
PlaybookStore Operations Standardization
Goal: make the operating system the same in every store, every shift.
Lets associates ask, in plain language, how to do the thing — "how do I process a tax-free return?" — and answers from the live ops manual, cutting onboarding time.
Auto-run. Read-only Q&A.
Reviews opening and closing checks, endcap photos, and safety logs to score compliance, and reminds store leaders about the checklist they missed.
Auto-run. Internal reminders.
Assigns bite-sized training off the back of each associate's task performance or a role change, and tracks completion.
Auto-run. Internal scheduling.
PlaybookShrink & Loss Prevention
Goal: cut total retail loss — audit the anomalies, enforce the controls — without treating staff as suspects.
Audits register logs for the patterns that signal loss — void clusters, serial refunds, under-scanning — and compiles a case for loss prevention to review.
Propose + approve. It surfaces a case; a human investigates. An agent must never conclude guilt about an employee — false positives carry a human cost. Loop In Experts; keep a person between the flag and any consequence.
Reconciles warehouse manifests, supplier invoices, and receiving reports to catch the admin errors that quietly become write-offs.
Auto-run. Internal discrepancy reports and notifications.
Maps shrink rates onto the planogram to flag high-risk locations — cosmetics by the exit — and suggests relocation or security pegs.
Auto-run. Internal suggestion to ops.
Trust signature: reversible, which is the trap. Roll-back-able actions still reach every customer and quietly steer margin. Automate boldly, then engineer the cooling.
PlaybookE-commerce Conversion Rate Optimization (CRO)
Goal: find and remove the friction killing digital conversion.
Watches for real-time drops — a form, a page, a payment gateway quietly failing — and drafts the Jira ticket, with repro steps, for engineering.
Auto-run. Internal alert and a drafted ticket.
Tailors the homepage, category pages, and banners to referral source, browsing history, and purchase propensity — serving components straight to the live layout.
Sampled review (born HIGH). Live, customer-facing changes are hot — but reversibility plus an allowlist of approved components (D6) and instant rollback (D3) cool them to a sampled-review action. This is reversibility used as leverage, not licence.
Sizes the sample, tracks significance to keep false positives out, declares winners, and kills losing variants to protect revenue.
Sampled review. Auto-killing a losing variant is safe inside statistical guardrails; sample the calls to confirm the agent isn't ending tests early.
PlaybookOmnichannel Fulfillment
Goal: route every order to balance shipping cost, store capacity, stock, and the promise you made the customer.
Picks the most profitable node — DC, store A, store B — for each online order in real time, costing shipping and split-shipment risk, and writes the decision into the OMS.
Approve the policy, automate the instances. Writing production at scale is hot — so a named owner approves the routing policy and its caps once, and the individual per-order decisions then auto-run safely inside it.
Tracks time-to-pick and time-to-handover, catches the bottleneck or the understaffed shift, and alerts the manager when pick times blow past SLA.
Auto-run. Internal alert prompting human reallocation.
Calculates the fully-loaded cost to fulfill every transaction — pick labor, packaging, shipping — and exposes the deliveries and zip codes that lose money, proposing fee or listing fixes.
Propose + approve. It proposes commercial changes — restrict a listing, adjust a fee — and a human decides.
PlaybookOmnichannel Merchandising
Goal: make price, promotion, and availability consistent across the shelf and the screen.
Scans web and app prices against shelf tags and POS to catch discrepancies — and can trigger the correction to digital or the registers.
Named approver. Auto-correcting prices across channels is a live, public price write. Run it in flag-only mode (D1) or inside tight tolerance caps (D4) with sampled audits; "detect" is cool, "correct" is hot.
Predicts web vs store velocity and rebalances safety stock and shared pools so neither channel starves.
Sampled review. Reversible internal writes; sample them and cap the swing.
Spots the bulky, low-margin, low-turn items that belong online-only as drop-ship, and builds the quarterly "shift to digital" list.
Auto-run. Internal recommendation.
PlaybookOn-Site Search & Navigation Optimization
Goal: help customers find it — better taxonomy, fewer dead ends, more search-to-cart.
Audits the searches that return nothing and links the synonyms that fix them — "sweatpants" to "joggers" — updating the live dictionary for high-confidence matches.
Sampled review (born MEDIUM). It writes live search config, but a confidence threshold plus reversibility keeps it cool; sample the additions.
Ranks results by blending relevance, margin, availability, and purchase likelihood — generating the order customers see in real time.
Sampled review. Reversible and bounded, but it quietly steers both discovery and margin — so audit a slice and watch for margin creeping ahead of relevance.
Learns how customers actually filter each category, promotes the filters they use, and hides the dead ones.
Sampled review. Low-stakes, reversible UI tuning.
PlaybookRetail CRM & Lifecycle Marketing
Goal: move from batch-and-blast to lifecycle journeys that grow retention and CLV.
Spots the customers slipping away — longer gaps, smaller baskets — and triggers a tailored win-back with a dynamically sized offer.
Sampled review (cooled from HIGH). External customer comms plus a discount are the two classic escalators (E1, E2). Cap the discount depth (D4) and allowlist the templates (D6), and the sends drop to sampled review.
Decides the right move per segment — cross-sell, review request, loyalty signup — and feeds it into the CRM's outbound engine.
Sampled review. A recommendation layer driving customer comms; sample the actions it injects.
Coordinates onboarding, birthday, abandonment, and loyalty triggers across email, SMS, and push, dialing frequency to avoid fatigue.
Sampled review. External sends, but inside approved content and frequency caps (D4, D6).
PlaybookRetail Loyalty Program Design & Optimization
Goal: design tiers and rewards that lift incremental margin from members — without runaway liability.
Simulates the financial liability of outstanding points and the margin hit of different reward thresholds, and raises the alarm if liability crosses your line.
Auto-run. Internal simulation and alert.
Catches the members one visit from leveling up — or slipping down — and sends the incentive: "spend $15 more to unlock Gold."
Sampled review. Customer comms carrying an incentive; cap the incentive and approve the templates, then sample.
Runs vendor-sponsored loyalty promos, keeps co-op funding matched to points awarded and redeemed, and drafts the reimbursement invoice.
Named approver. Vendor invoices bind money to an outside party. Draft-only (D1); finance sends.
PlaybookReturns Reduction & Reverse Logistics
Goal: find what drives returns, stop the abuse, and route returns down the cheapest recovery path.
Reads return reasons, review comments, and sizing complaints to find the defective or mislabeled products, and alerts QA and the buyer to high-return items.
Auto-run. Internal alert.
Profiles return behavior to flag serial returners and likely wardrobing, and can apply stricter return rules — a fee, no open-box.
Named approver. Penalizing a customer is external, reputationally sensitive, and hard to undo if you're wrong. Flag for review (D1), never auto-penalize; a human applies any rule. Clarify Before Commit.
Routes each return to its most profitable end — restock, outlet, liquidate, recycle, return-to-vendor — and generates the disposition instruction at the receiving desk.
Sampled review. Liquidation is irreversible, so auto-run the low-value dispositions on an allowlist (D6) and route high-value items to review.
Trust signature: cool throughout. These agents read, diagnose, draft, and chase — they never touch the outside world. Auto-run, and demand the working.
PlaybookRetail KPI Dashboard & Weekly Business Review (WBR)
Goal: move from reporting what happened to diagnosing why — and tracking what gets done about it.
Catches the weekly miss — category X down 5% — and queries across traffic, margin, inventory, and promo to find the actual driver, drafting the diagnostic that explains why.
Auto-run. Internal analysis. Hold it to Transparency, Tapered: show the working before anyone trusts the "why."
Pulls marketing, store ops, e-commerce, and merchandising into the weekly briefing deck and summary — in the inbox 24 hours before the meeting.
Auto-run. Internal document to leadership.
Tracks the actions assigned in the WBR, checks progress, chases the owners, and escalates the stalled ones to directors.
Auto-run. Internal follow-up — turning the review from a status meeting into accountability.
The heat reading tells you how hot a single action is. Zoom out and the 60 agents fall into four control classes — and the class determines the design requirements, because the useful question is not what can this agent do but what happens when it is wrong.
| Control class | Typical agents | What a mistake costs | Required controls |
|---|---|---|---|
| Customer-visible, immediate | Service, returns, shopping and discovery | Trust and reputation, instantly and sometimes publicly | Confidence cues, value thresholds, fast human escalation, easy customer redress |
| Commercial, compounding | Pricing, promotion, markdown, supplier terms | Margin, quietly, until someone runs the numbers | Hard guardrails, variance alerting, evidence trail, revert path |
| Operational, recoverable | Store tasks, shelf compliance, scheduling, maintenance | Wasted effort and store-team trust in the system | Confidence display, easy dismissal with feedback, false-positive monitoring |
| Planning, delayed | Forecasting, allocation, replenishment | Stock position weeks later, hard to attribute | Uncertainty display, scenario visibility, planner override, post-hoc audit |
↔ scroll table on mobile
The pattern worth noticing: the agents that feel safest to automate — planning and operations, with no customer in sight — are often the ones where mistakes are hardest to detect and slowest to surface. Customer-facing errors announce themselves. A forecasting agent drifting quietly for six weeks does not.
Every agent in this gallery needs an answer to five questions before it goes live: what may it do alone, how does it show uncertainty, what does it remember, how does a person take over, and how would you reconstruct what it did? Those are AUX questions. → What is Agentic User Experience?
Three factors, in this order.
How much work does this actually absorb, at what volume? Be honest about frequency — a high-value agent that runs eleven times a month is a demo, not an operational asset.
Is the process documented, are the exceptions known, is the data reachable, and is there someone who will own the agent's behavior? An undocumented process does not become clearer when you automate it — it becomes faster and harder to audit.
What does a wrong action cost, and how quickly would you find out? This determines the control model, not whether to proceed.
How to read the combination: start where value is high, readiness is real, and consequence is recoverable. Save high-consequence domains for after your organization has learned to supervise an agent — because the supervision capability, not the model, is what you are actually building. → The five-layer agentic organization playbook
A gallery tells you what's possible. It doesn't tell you what to build first, what your data can actually support today, or who in your organization is accountable when an agent acts. That last part is where most retail AI programs stall — not on the model, on the operating posture. That's the work auxfirst does.
Map a real retail workflow as it runs today, then design where and how agents act within it: intent, autonomy levels, human handoffs, the interaction patterns the agents will follow. You leave with an agent interaction blueprint and an intent-and-autonomy map of what the agent owns versus what stays human.
Bring your real action inventory. We run every verb through the five dimensions, write the band and the posture next to each, and design the de-escalators that cool the hot ones — and you walk out with the rubric your team operates from.
Already shipping a retail copilot or agent? We diagnose it against the AUX framework — intent, memory, transparency, control — and hand back a Trust Scorecard, the failure modes, and a prioritized fix list.
Stress-test a built agent against real cases, edge cases, and failure modes before it touches a live price or a live customer. Go / no-go, with the fixes ranked.
Keep evolving the system as it scales and breaks: promote the agents that earn trust up the ladder, demote the ones that burn it, and re-score on every incident.
The discipline underneath all of it is one rule from the ladder: a heat model is only worth something when it changes what an agent is allowed to do unsupervised. List the verbs, score the five dials, read the band, design the cooling, review on incidents. The function changes; the shape of the ladder does not.
You don't need sixty agents. You need the handful that move your numbers — designed so you can trust them to act, and cooled so the hot ones can still run. We'll map your real action inventory onto the ladder in a single working session and hand you the posture for every one.