Read one before you buy.

M1. Know what every channel actually returned yesterday.

The job. One truthful answer, every morning, to the only marketing question that matters: what did we make for what we spent, per channel, instead of four ad dashboards that each claim the same sale.

The tool. Buy, then wire: a third-party attribution platform (the WorkMagic and Triple Whale class of tool) connected to Shopify, Amazon, Meta, and TikTok, validated against your store of record so you trust it, plus direct spend pulls from each ad platform's reporting API so the spend number never waits on anyone.

Roughly. The platform is a subscription you buy from its vendor. The wiring, weekly validation, and reporting integration is $2,000 to $4,000 to build.

Hand your developer this. Connect the attribution platform to Shopify, Amazon, Meta, and TikTok Shop and reconcile its order totals against the store of record weekly. Pull spend per platform directly from each reporting API, normalized to date, platform, and spend. Merge sources so the freshest one wins by date, and footnote any platform that has not synced rather than showing a silent zero.

M2. Capture an email from every conversation that does not end in a sale.

The job. A visitor who talked to the agent for five minutes and left is a lead, not a loss. Capture the email, with consent, and land it in the welcome flow.

The tool. A capture tool inside the chat agent wired to the email platform's API (ours is Klaviyo), subscribing to both a dedicated chat-leads list and the main welcome-flow list.

Roughly. $1,000 to $2,000.

Hand your developer this. Build a capture tool the agent can call with email, optional phone, and consent state. Subscribe through the ESP API to both lists, queue failed captures for retry instead of dropping them, and never enroll a phone number in SMS without explicit opt-in, because a captured phone number is a property, not consent.

M3. Write down what the brand is allowed to say, once, in machine-readable form.

LIVEIt is the corpus the storefront agent answers from today.

The job. Every AI system in the company must sell from the same approved facts: product ingredients transcribed from the actual labels, dosing rules, the store's guarantee and return policy, brand voice, and the compliance lines that must never be crossed. For this brand: products were developed for Cesar Millan and are never described as formulated by him, and no answer ever tells a customer to stop veterinarian-prescribed medication.

The tool. A versioned content corpus with a repeatable chunk builder, feeding retrieval for every agent. Ours holds the full catalog, the official usage guide, and every product label.

Roughly. $2,000 to $4,000.

Hand your developer this. Keep every approved fact in versioned plain-text files, compile them into retrieval chunks with one script, and make every AI system answer from the corpus rather than from its own memory. Build this before the agent; the agent is only as safe as the corpus it answers from.

S1. A salesperson on the storefront who never sleeps and cannot make things up.

LIVEOn betterdogsupplements.com today. Go ask it something.

The job. Answer product questions instantly at any hour, recommend by symptom and dog weight, build the cart with live variant IDs, track orders, and hand off to a human the moment it should. In supplements, a wrong answer is worse than no answer, so wrong answers have to be blocked by design, not discouraged by prompt.

The tool. An LLM agent (we use Claude) with retrieval over the approved corpus (M3) and a tool loop: search corpus, product lookup, add to cart, track order, escalate. A second, cheaper model runs as a deterministic verifier that blocks any reply containing a factual claim not grounded in the retrieved corpus. When the verifier rejects an answer, the customer gets a safe fallback and a human, never a guess. The system fails closed.

Roughly. $10,000 to $15,000 for an own-developer build. We also sell a productized version, quoted in writing; the quote appendix covers that.

Hand your developer this. A two-model loop where the frontline model proposes and the verifier disposes, failing closed to human handoff on any mismatch. Cart writes go through the storefront's own cart API, and every escalation opens a ticket in the helpdesk with the full transcript attached.

S2. Sell the bundles customers actually buy, first.

The job. The agent should lead with the bundles real customers choose, ranked by real sales, and never push twice after a customer declines.

The tool. A nightly script that queries 90 days of the store's own order data and writes a ranked recommendation file the agent reads at runtime.

Roughly. $1,000 to $2,000.

Hand your developer this. Query 90 days of line items, dedupe by product handle, rank bundles and singles separately, and write a small JSON file the agent loads. Encode the selling rule in the prompt: bundle first, single on decline, never repeat the pitch.

S3. Make cancellation a conversation.

The job. A subscription cancellation should trigger a question, not a shrug. Learn the reason, answer it with the right save offer, and log every reason so the product team sees the pattern.

The tool. A cancellation survey wired into the subscription platform, reason-specific save offers (pause, swap, delay), and a monthly reasons report.

Roughly. $1,500 to $3,000.

Hand your developer this. Intercept subscription cancellation with a reason survey, present reason-specific save offers within the subscription platform's rules, and log every reason and outcome to a monthly report.

CS1. Same brain, every channel: the helpdesk on AUTO.

LIVEEmail, social DMs and comments, and SMS are answered by AI in full AUTO mode through our helpdesk today.

The job. Every conversation, on every channel, answered fast and accurately in one brand voice by the same grounded agent that runs the storefront, with humans working exactly one "needs human" view instead of a raw inbox.

The tool. Custom helpdesk automation built on the helpdesk's REST API (ours is Gorgias), sweeping on a short cadence (ours runs every 2 minutes), with a sentinel job that catches stuck replies, sender filters so platform notifications and vendor pitches never get a public reply, and a handoff packet on every escalation: the data the agent already verified, plus the draft it discarded and why, so the human never starts from zero.

Roughly. $6,000 to $10,000 on top of the agent build, because the brain is shared with S1.

Hand your developer this. Poll the helpdesk API on a cadence, classify each ticket by lane and sender, filter robot senders by address pattern, draft or auto-send per lane policy, tag everything the AI touched, dedupe escalations per ticket per day, and route all uncertainty into one shared human view.

CS2. Let customers help themselves with orders and subscriptions.

The job. "Where is my order" and "cancel my subscription" resolved inside the conversation, safely, without a human in the loop.

The tool. Order tracking with email verification against the store API, and subscription changes through the subscription platform's external API, with a fail-safe path that registers the request for a human whenever a write fails.

Roughly. $2,000 to $4,000.

Hand your developer this. Verify identity by matching the order email before revealing anything. Mine the order number from the email subject line when the customer forgot to include it, and fall back to an email-only lookup of the latest order. Treat every failed subscription write as a logged human task, never a silent drop.

CS3. Give the AI its own QA department.

The job. Nobody on your team will read every AI conversation, so a system has to.

The tool. An audit judge (a second model) that replays recent conversations on a schedule (ours ran every 2 hours) against a written rulebook and logs verdicts as flags on an admin dashboard, plus a taught-rules mechanism that turns any human correction into a permanent standing instruction, plus a daily brief listing what was handled, handed off, and flagged.

Roughly. $2,000 to $4,000.

Hand your developer this. A scheduled job that scores conversations against explicit written rules, stores structured verdicts, and gives staff a one-click way to teach a permanent rule. Ours caught our own launch-night mistakes in its first run. That is exactly what it is for.

O1. Inventory forecasting. This is something we do not sell.

The job. Know your stockout dates, reorder dates, and replenishment quantities per SKU, with supplier lead times baked in.

The kind of vendor that does it. A dedicated inventory-planning platform built for e-commerce. We pay for one ourselves (Flieber) and pipe its forecast into our dashboard. Buy this from a company whose whole business is inventory math; do not have an agency custom-build it, including us.

Roughly. The platform's pricing is the vendor's to quote. The integration on our side, feeding its forecast into the dashboard, is under $1,000 of scripting.

Hand your developer this. Consume the planning platform's exported report on a schedule, map its per-SKU status and reorder dates into the dashboard's inventory section, and flag any product whose optimal order date has passed.

O2. Clean product data, enforced by a robot.

The job. Every agent, feed, and dashboard downstream is only as good as the catalog. A missing ingredient list or metafield does not stay invisible; it surfaces as a wrong answer to a customer.

The tool. A scheduled audit script over every product and variant, with an emailed exception list and a publishing block on missing critical fields.

Roughly. $1,000 to $2,000.

Hand your developer this. Schedule an audit that walks every product and variant, validates required fields and metafields, emails an exception list, and blocks publishing when critical fields are missing.

O3. Turn team feedback into permanent agent rules.

The job. When a moderator catches the agent doing something wrong, the fix gets taught once and kept forever, instead of living in someone's memory.

The tool. A feedback channel in the team's project tool (ours is ClickUp), read on a schedule by an automation that applies each correction as a taught rule through the agent's admin API, posts confirmation back to the channel, and opens an engineering ticket when the fix needs actual code.

Roughly. $1,000 to $3,000.

Hand your developer this. Read the channel with a cursor so nothing is processed twice, apply corrections through an authenticated admin endpoint, reply in the channel with what was done, and separate "teach a rule" from "fix the code" explicitly.

O4. Give every automation a dead-man switch.

The job. An automation that silently stops is worse than no automation, because everyone believes it is still running.

The tool. OS-level scheduled jobs instead of app-dependent timers wherever possible, a freshness watchdog that texts a named human when any data feed goes stale past its budget, and once-per-day send flags so a retry can never double-send a report. Ours guards the morning report in this plan today.

Roughly. $1,000 to $2,000.

Hand your developer this. Schedule at the OS level, write a per-day sent flag before any send, and run one independent watchdog process whose only job is to alert a phone number when a data file's age exceeds its freshness budget.

F1. Per-product profitability, per channel, with real costs.

LIVEIt runs on our own cost sheet today.

The job. The true contribution of every SKU on every platform after landed cost, platform fees, shipping, and pick-pack, so pricing, ad budgets, and reorders stop being decided on gross revenue.

The tool. Cost-sheet ingestion keyed by SKU, bundle decomposition into component units, and per-channel fee and shipping models, working with the sample-order exclusion in F3. The cost sheet is the prerequisite; the engine is the easy part.

Roughly. $2,000 to $4,000.

Hand your developer this. Ingest a per-SKU landed-cost sheet, decompose bundles into component jars before costing, model each channel's fees separately, publish profit per product per platform to the dashboard, and reconcile totals against accounting monthly.

F2. The 7am morning report: yesterday, explained, before coffee.

LIVEIt arrives by email and SMS every morning, with the analysis already written.

The job. Yesterday's business in the CEO's inbox before the first meeting: revenue split by channel, spend per platform, subscription movement, and a short AI-written analysis of what changed and why, readable in ninety seconds. A text-message version goes to the phone.

The tool. Deterministic send scripts that read a pre-built data file. The slow part, the AI analysis, is generated ahead of time by the hourly refresh, so the 7am send takes seconds and never waits on a model. Delivery is confirmed, not assumed.

Roughly. $2,000 to $4,000.

Hand your developer this. Separate compute from delivery. Pick the report day by presence of settled sales, not by calendar. Hold the send behind QA gates that catch broken values, and when a platform has not synced its spend yet, omit the row rather than print a guess.

F3. Stop letting free samples poison the P&L.

LIVEBuilt as part of the profitability work above.

The job. Influencer sample orders ship at zero dollars and silently distort average order value, unit economics, and channel P&L. Exclude them from revenue math and count them where they belong: as a visible marketing cost.

The tool. An order-tagging rule at import plus a filter applied in every report, with samples broken out on their own line so the spend stays visible instead of disappearing.

Roughly. $1,000 to $2,000.

Hand your developer this. Tag zero-dollar sample orders at import, exclude them from revenue, AOV, and unit economics in every report, and surface them on a separate samples line so the cost stays visible.

F4. Subscription truth.

The job. Active subscribers, cancellations, and monthly recurring revenue from the systems of record, not estimates.

The tool. The subscription platform's external API for the storefront program, and the Amazon SP-API Replenishment reports for Subscribe and Save, each metric labeled with its source.

Roughly. $1,000 to $3,000.

Hand your developer this. Pull status counts and contract line items from the subscription API, pull Subscribe and Save seller metrics from the Replenishment report, and never present the two programs as one number without saying so.

C1. One screen the leadership team actually trusts.

LIVEIt is the first thing opened here every morning.

The job. One password-protected, phone-friendly page that answers "how did we do" without anyone pulling a spreadsheet: revenue by channel, advertising by platform, subscriptions, inventory risk, and a weekly scorecard with traffic-light comparisons. It is the surface every other item in this plan reports into.

The tool. A warehouse-connected web dashboard. Ours reads a commerce data warehouse, Shopify, the Amazon SP-API Sales and Traffic reports, Meta's API, and Klaviyo, refreshes hourly, sits gated on its own subdomain, and puts a source badge on every number.

Roughly. $5,000 to $10,000.

Hand your developer this. A single-page app reading one data-file contract, rebuilt hourly by one pull script per source, deployed behind a password, mobile-friendly, with a printed source badge on every card so nobody has to ask where a number came from.

C2. One written set of rules for what the numbers mean.

The job. A dashboard is only trusted if the definitions are written down and enforced in code, and an alert system that cries wolf gets ignored until the real fire burns.

The tool. A data contract living in the repository next to the code: platform-reported revenue and modeled attribution shown side by side and never blended; channels reconciled so nothing double counts (our TikTok Shop orders check out through Shopify, so adding both would count them twice); and a per-channel rulebook so known sync lags are footnoted on the page as lags, while only genuine anomalies escalate to a phone.

Roughly. $2,000 to $4,000, mostly discipline spent during the dashboard build.

Hand your developer this. Write the definition of every metric into the repo beside the code that computes it, label modeled numbers as modeled, encode each channel's known reporting quirks as annotation rules rather than alarms, and when two sources disagree, show both with their names on them.