Most useful for DTC ecommerce teams wanting practical AI agent workflows for marketing analytics, attribution, and saving money.
Best for
A DTC ecommerce brand around $5M to $800M GMV considering a talk or advisory session.
Not for
A consumer brand seeking a paid creator placement for product demos, reach, or engagement.
Key takeaways
One post lists 15 AI agent use cases already running in production DTC brands.
An agent found $16K per day of Amazon DSP spend was not incremental.
An apparent retention dip traced to a 3.6x spike in failed payments.
Check first: Confirm whether they accept sponsorships or paid placements, plus rates, deliverables, and audience terms.
Main ideas
The limit on AI agents is imagination, not capability · Sort agent work by whether it unlocks money or removes work · Attribution should fit how customers buy, not a template or a black box
David Dokes writes about applying AI to the work of DTC ecommerce teams, especially marketing analytics, attribution, and operational workflows.
They move beyond general claims about AI to show tasks agents can perform: finding ad spend that may not be incremental, drafting campaign materials, and flagging payment problems.
They also write about building attribution models around customer journeys and giving AI shared definitions for company metrics.
Follow their work for practical examples of how ecommerce teams can use AI—and for the measurement and data-governance questions those uses raise.
David Dokes connects specific business problems to concrete AI workflows.
The posts distinguish finding money a team may be wasting from reducing manual work, and include figures such as $16,000 a day in Amazon DSP spend identified as not incremental and a 3.6-fold spike in failed payments behind an apparent retention dip.
They also address a less visible prerequisite: AI can calculate correctly but use the wrong definition of a metric, so teams need shared context.
Together, these posts show a practical range—from identifying where to act, to shaping attribution, to setting the metric definitions AI should follow.
David Dokes writes about putting AI agents to work inside direct-to-consumer ecommerce businesses, mainly in posts introducing features of the Polar measurement product.
Read more
Two arguments run through the material: the obstacle to using agents is usually imagination rather than capability, and an agent earns its place by either finding money a team could not see or removing work a person used to do.
A parallel thread covers measurement: attribution models should be built around how customers actually buy, and definitions of metrics such as conversion rate should be shared company context rather than one person’s private setting.
The examples come from brands in the $5M to $800M GMV range and include the specific numbers and tools behind each one.
The limit on AI agents is imagination, not capability
Dokes argues that teams are not short of AI tools: nearly every ecommerce team they meet is already building dashboards on Claude or ChatGPT.
Read more
But very few can name a task an agent could take on Monday morning.
The framing is blunt: companies do not have an AI problem, they have an imagination problem.
The useful work is cataloguing concrete tasks and the data each one needs, then handing them over, rather than buying more capacity.
That is why they publish a series of real agent workflows instead of general advice.
The argument draws on a first pass through roughly 9,500 conversations across omnichannel brands from $5M to $800M GMV in the US, UK, and Australia, in the weeks after launching an agent that sits in Slack and connects to a company’s numbers and tools.
The opening anecdote comes from a customer whose yearly number keeps growing while the team does not.
Dokes sees an opportunity to do more without growing headcount at the same rate.
Sort agent work by whether it unlocks money or removes work
The examples in Dokes’s field-guide post fall into two buckets.
Read more
“Unlock” cases surface money the team could not see — non-incremental ad spend, wasted search terms, a retention problem that is really a payments problem.
“Reduce” cases take over work a person used to do, such as drafting campaigns or checking supplier documents.
Each use case documents what the person asked, what the agent did, the output, the data and tools it needed, and where a human stayed in the loop.
That structure turns a vague pitch into a task a brand can budget for.
Under unlock: proving that $16K per day of Amazon DSP spend was not incremental, then cutting that spend by 98% with sales down only 2.5%; a Google Ads audit where the VP of Growth approved four of five recommendations, including $6.7K per month of junk search terms and a campaign running without an effective ROAS floor; and a “retention dip” that turned out to be a 3.6x spike in failed payments. Under reduce: building campaign drafts and flows directly in Klaviyo; reviewing factory inspection reports against the tech pack and drafting the supplier email, 25 in three days; and sending leadership a daily actuals-versus-forecast flash.
Attribution should fit how customers buy, not a template or a black box
Dokes frames the choice brands face on attribution as two bad options: rigid templates such as first-click, last-click, linear, or U-shaped, which are understandable but fixed, or black-box models that claim to adapt but offer little visibility or control.
Read more
Either way, budget decisions rest on a model that does not fit the business or that nobody can explain. Their alternative is to build the model around the actual buying journey and describe the credit rules in plain language rather than configure a preset, so the result can be shown to the people who approve spend.
The three steps shown are: connect Claude to Polar through an MCP connection; visualize the buying journeys captured by the Polar Pixel, including channels, touches, and time to purchase; then describe how credit should be assigned and create the model by prompting.
David Lawson, named as Head of Ecommerce at the brand VALABASAS, is quoted on building custom models with the team: “I've gotten insights that every other model and view and platform missed.”
Metric definitions are company context, not a personal file
An AI can get the calculation right and still use the wrong metric.
Read more
Ask about conversion rate and you may mean Shopify’s figure while the assistant returns the one from Google Analytics.
Dokes argues that a preference should not live in someone’s private assistant project or a file on their laptop; it needs to be shared across the company, with permissions controlling who can create and edit it.
This makes agent setup a governance decision about a single source of truth, often the reason AI answers disagree with the numbers leadership trusts.
Labels are created inside Polar or directly from Claude, for instance by asking Polar to “Create a Source of Truth label.”
The team assigns the metrics it wants AI to use, such as conversion rate or sales.
Those metrics then feed dashboards or an agent set up to monitor a chosen rate in Slack.
Prompt it, and let the output land where the team already works
Across the three posts, the same delivery pattern repeats: prompt an assistant to build a custom attribution model or ask Polar to create a label, connect it to an ecommerce data platform, and send the result to a dashboard, marketing tool, or Slack message.
Read more
Dokes argues that the hard part of adoption is connecting existing systems and writing clear instructions, not learning a new interface.
The Slack agent connects to a company’s numbers and tools; the attribution model is created by prompting Claude through an MCP connection; the Labels post ends with preferred metrics available as shared context for dashboards, questions, and agents.
The three posts describe features introduced over an unspecified period, so this reads as a consistent method rather than a shift. Publication dates were not included in the material reviewed, so no change in emphasis over time can be established.
Start here if you want to know what an AI agent does for an ecommerce team on a Monday morning.
Read more
This post is the clearest statement of David Dokes's working thesis and the most concrete set of examples they have published, so it answers both "what is this person about" and "what would an agent actually do for us."
David Dokes describes reading every channel during the first weeks of Polar Operator, an AI agent that sits in Slack and connects to a company's numbers and tools.
The work spanned omnichannel brands from $5M to $800M GMV in the US, UK, and Australia, covering roughly 9,500 conversations.
The opening anecdote is a customer saying the yearly number keeps growing while the team does not.
They draw the opportunity from that: do more without growing headcount at the same rate.
The conclusion is blunt: companies do not have an AI problem; they have an imagination problem.
Nearly every ecommerce team already builds dashboards on Claude or ChatGPT, but few can name a task an agent could take over on Monday morning.
The post sorts the work into two buckets and gives named numbers for each.
Under UNLOCK, examples include proving that $16K per day of Amazon DSP spend was not incremental, then cutting that spend by 98% with sales down only 2.5%; auditing a Google Ads account where the VP of Growth approved four of five recommendations, including cutting $6.7K per month of junk search terms and addressing a campaign running without an effective ROAS floor; and discovering that a “retention dip” was actually a 3.6x spike in failed payments.
Under REDUCE, examples include building campaign drafts and flows directly in Klaviyo; reviewing factory inspection reports against the tech pack and drafting the supplier email—25 in three days; and sending leadership a daily actuals-versus-forecast flash.
The promised field guide lists, for each use case, what the person asked, what the agent did, the output, the data and tools it needed, and where a human stayed in the loop.
For brands planning a campaign
This is a natural spine for a talk or advisory session.
The unlock-versus-reduce split gives a brand a way to sort its backlog before making tooling decisions.
The figures ($16K per day of non-incremental DSP, $6.7K per month of junk search terms, and a 3.6x failed-payment spike behind an apparent retention problem) are specific enough to open a discovery conversation with a growth or finance lead.
For a brand in the $5M to $800M GMV range, the post frames the real constraint as imagination rather than capability—a useful starting point for a workshop that ends with a written list of Monday-morning tasks.
Read this for the clearest example of how Dokes argues about measurement and for a named customer reaction to the result.
Read more
It shows the difference between templated attribution and a model a brand can explain.
The post argues that attribution tools usually give a brand two bad choices: rigid templates such as first-click, last-click, linear, or U-shaped, which are understandable but fixed; or black-box models that claim to adapt but offer little visibility or control.
Either way, budget decisions rest on a model that does not fit the business or cannot be explained.
Their alternative is to build the model around how customers actually buy.
The process has three steps: connect Claude to Polar through an MCP connection; visualize the buying journeys captured by the Polar Pixel, including channels, touches, and time to purchase; then describe how credit should be assigned and create the model by prompting.
The post says the model was built by prompting Claude and names VALABASAS as a brand that built custom models with the team.
It quotes VALABASAS's Head of Ecommerce, David Lawson: “I've gotten insights that every other model and view and platform missed.”
A video link offers to build a custom attribution model in under five minutes.
For brands planning a campaign
Attribution arguments decide budgets, so this is the post to hand a brand stuck between a template nobody trusts and a vendor model nobody can read.
The three-step process—connect, visualize journeys, describe the credit rules—provides a usable agenda for a working session with a growth team.
The VALABASAS quote gives a buyer a peer reaction rather than a vendor claim.
Ask for the five-minute walkthrough and supporting measurement behind the quote.
A short, precise example of a problem many teams notice only when an AI answer is confidently wrong.
Read more
It is a strong starting point for discussing how a team agrees on what its numbers mean.
Labels are shared context that tells AI which metrics to use and when.
The problem is plain: an AI can get the calculation right but still pick the wrong metric.
Ask about conversion rate and it may return Google Analytics' figure when you mean Shopify's.
They argue that preferences should not live in someone's private Claude project or a file on their laptop; they need to be shared across the company, with permissions controlling who can create and edit them.
The process has three steps: create labels inside Polar or from Claude, for example by asking Polar to “Create a Source of Truth label”; assign the metrics the team wants AI to use, such as conversion rate or sales, to that label; then use the label in a dashboard or set up an agent to monitor the chosen conversion rate in Slack. The team's preferred metrics become shared context for dashboards, questions, and agents.
For brands planning a campaign
This is the governance piece, and it is easy to underrate.
Any brand that wants agents reading its numbers has to settle definitions first, or agents may confidently report the wrong version of a familiar metric.
The permissions detail makes this a starting point for a conversation with analytics or data leadership, not just marketing.
It also works as a short prerequisite module before the money-finding examples: agree on the metrics, then let the agent look for the gap.
Prompt-first interfaces, Slack as the working surface3
AI agents in production DTC brands: what to point them at first1
Building custom attribution models by prompting1
Shared metric definitions before agents get access1
AI agents in production DTC brands: what to point them at first
David Dokes writes about AI agents already running inside direct-to-consumer ecommerce businesses.
Read more
In one post, the examples are split into two groups.
"Unlock" cases find money a team could not see: proving $16K/day of Amazon DSP spend was not incremental (spend cut 98%, sales down 2.5%), auditing a Google Ads account to remove $6.7K/month of junk search terms and a campaign running without an effective ROAS floor, and tracing a "retention dip" to a 3.6x spike in failed payments.
"Reduce" cases remove work a person used to do: drafting Klaviyo campaigns and flows, checking factory inspection reports against the tech pack and drafting the supplier email (25 in three days), and sending leadership a daily actuals-vs-forecast flash.
Each case is documented with what the person asked, what the agent did, the output, the data and tools it needed, and where a human stayed in the loop.
Dokes describes reading about 9,500 conversations across omnichannel brands from $5M to $800M GMV in the US, UK, and Australia in the first weeks after launching an AI agent that sits in Slack and connects to a team's numbers and tools.
Each example follows a repeatable structure: the request, the agent's action, the output, the tools and data required, and the human checkpoint.
For brands planning a campaign
For a brand weighing AI agents, this is a set of concrete tasks with numbers attached, not a vendor pitch. A talk or advisory session can start with which tasks pay back first, what data each needs, and where a person should stay in the loop — useful for a team that has tried chat tools but does not know what to hand an agent on Monday morning.
A post introducing custom attribution models in Polar describes them as built by prompting Claude.
Read more
Dokes frames the choice teams face as rigid templates (first-click, last-click, linear, U-shaped) or black-box models that adapt without offering visibility or control.
The steps shown are: connect Claude to Polar through an MCP, visualize the buying journeys captured by the Polar Pixel — channels, touches, and time to purchase — then describe in words how credit should be assigned.
The post cites a customer example: VALABASAS built custom models with the team, and its head of ecommerce said the insights covered things other models and platforms missed.
Prompt-driven model specification: connect the assistant to the measurement platform, inspect journey data, then describe the credit rules in natural language instead of configuring a fixed template.
For brands planning a campaign
Suited to brands that distrust template attribution but cannot explain a vendor's black box to finance or leadership. A session can show how to specify a model in plain language, what journey data it depends on, and how to sanity-check the result before shifting budget.
Shared metric definitions before agents get access
A post on Labels describes a context problem: an AI assistant can get the calculation right and still pick the wrong metric — for example, using the conversion rate from Google Analytics when the team means Shopify's.
Read more
Labels record which metrics a team wants used and when, are created inside Polar or directly from Claude, and carry permissions controlling who can create and edit them.
Once defined, they feed dashboards built on those metrics or agents monitoring a chosen rate in Slack.
The post's point is that this preference should be shared company context, not buried in one person's private assistant project or a laptop file.
The post covers metric definitions, not broader data quality or warehouse work; ask whether those are in scope.
For brands planning a campaign
Relevant to any brand giving agents or dashboards access to marketing data.
It frames setup as a governance decision: agree on the source of truth for each metric, make it shared, and control who can change it.
That can unblock teams when AI answers disagree with the numbers leadership trusts.
Prompt-first interfaces, Slack as the working surface
Across the three posts, the same delivery pattern recurs: work is configured by prompting an assistant (build a custom attribution model, ask Polar to create a label) connected to an ecommerce data platform through an MCP.
Read more
The output lands where the team already works — a dashboard, a Klaviyo draft, or a Slack message. The agent introduced in the field-guide post lives in Slack and connects to a company's numbers and tools.
Prompt the assistant to configure or trigger a process; the data platform provides the numbers, and Slack or the existing marketing tool delivers the output.
For brands planning a campaign
For a brand, the adoption cost sits less in new tooling than in connecting existing systems and writing clear instructions. An advisory session can map which of the brand's tools need connecting before any of the showcased use cases can run.
Across the three reviewed LinkedIn posts, David Dokes's brand connections centre on Polar, the ecommerce data and AI-agent product discussed in all three.
Read more
They write as part of the group behind Polar and describe its features, one named customer, and an integration with Claude. Ask Dokes to clarify their role at Polar.
Polar
Disclosed partnership
Dokes posts as one of the people behind Polar, an ecommerce analytics and AI-agent product: “We recently launched Polar Operator, an AI agent in your Slack, connected to your numbers and your tools,” and “We're bringing marketing measurement and AI together.”
Read more
They also describe Polar's custom attribution models and its Labels feature.
Two posts describe Polar features that run on or connect to Claude: custom attribution models “Built by prompting Claude” through Polar's MCP, and Labels that can be created “directly from Claude.”
David Dokes writes about where AI agents pay off inside ecommerce teams and the measurement plumbing that has to be right first.
Read more
Start with the post cataloguing 15 agent use cases already running at DTC brands, split into work that finds money the team could not see and work that takes over a task a person used to do.
Each case covers the request, the agent's action, the output, the data and tools it needed, and where a human stayed in the loop.
Then read the two measurement posts: one on building a custom attribution model by describing a buying journey in plain language, and one on Labels, which keep a team's agreed metric definitions — such as which conversion rate counts — in shared company context instead of one person's assistant project.
The posts introduce features of Polar, a measurement product, and the customer results are the vendor's account; ask for independent validation before relying on them.
These are short LinkedIn posts rather than a course.
Start with the field-guide post; the two measurement posts work in either order.
Writes for the same reader — ecommerce and paid-media teams — about how AI tools change day-to-day media buying.
Read more
One post reviews GoMarble AI, a chat tool that connects Meta, Google, and Shopify accounts to run AI analysis and reports, and judges the output rather than trusting it, including where it makes shaky assumptions such as applying ROAS to lead generation.
The other translates Meta's four new algorithm systems into what marketers should stop and start doing.
Like Dokes, the emphasis is on concrete practice: prompt quality, what data a tool needs, and which recommendations are safe to act on.
Noah King mostly evaluates tools built by other companies and explains platform changes as they land, including a review disclosed as paid.
David Dokes writes about workflows on the measurement product their own posts introduce, drawing examples from customer conversations rather than hands-on product tests.
Go to Noah King for an outside read on whether a given AI marketing tool is worth paying for; go to Dokes for how an internal agent workflow is designed and where a person stays in the loop.
Covers the same measurement problem in the same market: numbers on the dashboard that do not mean what the team thinks they mean.
Read more
One post argues AI answers and AI-driven discovery are invisible to click-based attribution and land as "direct traffic"; the other walks through an affiliate channel showing 12x blended ROAS that failed an incrementality test, and explains why the over-attribution concentrates on customers who would have bought anyway. Both are close-read arguments with a mechanism and a required action, the register Dokes uses too.
Will O'Brien diagnoses broken measurement across existing channels and pushes for incrementality testing, largely as an independent commentator on industry shifts, without tying the argument to a product being sold.
Dokes starts from the opposite end, showing how a model or an agent is configured and what data it needs.
For a team deciding whether its reported numbers can be trusted, Will O'Brien is the more direct read; for building the replacement workflow, Dokes is.
Writes about agentic commerce reaching ecommerce checkout and the same measurement blind spot Dokes addresses.
Read more
Posts cover Google's Universal Commerce Protocol going live, a four-point diagnostic for making a Shopify store readable and buyable by AI agents, and the warning that when an agent skips the landing page and the pixel, a paid-media sale is recorded as direct traffic. The reader is the same Shopify and DTC operator, and the guidance comes as checklists to act on.
Martin K. concentrates on store-side readiness — product taxonomy, checkout handshakes, whether agents can find and buy an item at all — with short, urgent posts and little case data.
Dokes stays inside the brand's own analytics stack, covering the agent tasks and metric definitions a team controls after the purchase is possible.
If the open question is whether agents can transact with your store, start with Martin K.; if it is what to point an agent at once they can, start with Dokes.
David Dokes has a LinkedIn following of 23,871 as of 8 October 2026, relevant to a B2B ecommerce audience.
Read more
The three posts target people who run ecommerce growth, analytics, and retention: examples include a customer conversation with a brand team, recommendations approved by a VP of Growth, and a quote from a head of ecommerce at a named DTC brand.
The content is aimed at practitioners and budget owners, not general consumers, but the available information does not show who reads the page or their age, gender, geography, job seniority, or employer.
If a campaign depends on reaching a specific market or role, ask them for platform demographic data or audience survey results, plus recent results for posts covering comparable topics and depth, to assess typical reach rather than rely on follower count.
A DTC ecommerce brand at roughly $5M–$800M GMV wants to use AI agents for marketing analytics and attribution and is considering David Dokes for a talk or advisory session on real agent workflows that find and save money.
Read more
This is a strong fit, for specific reasons.
The published work addresses what the brand wants.
One post walks through 15 agent use cases running in production DTC brands, split into two useful buckets.
Under "unlock" — money a team could not see — Dokes reports proving $16K per day of Amazon DSP spend was not incremental, cutting that spend 98% with sales down only 2.5%; auditing a Google Ads account where a VP of Growth approved four of five recommendations, including $6.7K per month of junk search terms and a campaign running without an effective ROAS floor; and tracing an apparent "retention dip" to a 3.6x spike in failed payments.
Under "reduce" — work a person used to do — the examples include building Klaviyo campaign drafts and flows, reviewing factory inspection reports against the tech pack and drafting the supplier email 25 times in three days, and sending leadership a daily actuals-versus-forecast flash.
Each case follows a repeatable structure: what the person asked, what the agent did, the output, the data and tools it needed, and where a human stayed in the loop.
That is exactly what a brand needs before handing an agent live tasks.
Attribution gets direct treatment.
A second post argues that teams are stuck between rigid templates (first-click, last-click, linear, U-shaped) and black boxes that adapt without giving visibility or control.
It describes building a model by prompting: connect the assistant to the measurement platform, inspect the buying journeys captured by the pixel — channels, touches, time to purchase — then describe in words how credit should be assigned.
David Lawson, head of ecommerce at VALABASAS, says the insights covered things other models and platforms missed.
That is one customer's stated experience, not independent measurement, but it is a peer reaction rather than a vendor claim.
The governance gap that often blocks analytics use cases is also addressed.
A third post on Labels explains that an AI can get the calculation right and still pick the wrong metric — returning Google Analytics' conversion rate when the team means Shopify's — and describes making the team's preferred metrics shared company context, with permissions on who can create and edit them.
For a brand trying to make agent answers match the numbers leadership trusts, that is the prerequisite conversation, and they have already framed it.
The level and audience line up too.
The writing addresses practitioners and budget owners: customer conversations, recommendations approved by a growth lead, and a head of ecommerce quote.
The opening claim — that companies do not have an AI problem so much as an imagination problem, and that most teams can build dashboards but cannot name a task for an agent on Monday morning — gives a workshop a frame that can end with an action list.
Before booking, agree on rates, deliverables, session format, and how to inquire; the posts do not specify them.
Follower count is 23,871 on LinkedIn, but the posts give no demographic or per-post engagement data.
If the brand wants the session promoted to their following, confirm audience composition.
The use-case post promises a series; ask for later instalments, which are not included here.
If the brand's stack differs from the tooling shown in these posts, confirm the advice transfers before committing.
When to choose a specialist
A different hypothetical campaign: a consumer brand — say a home-goods or supplement label — wants to buy a paid creator placement, sponsoring a short series of posts with product demonstrations aimed at a broad shopping audience and judged on reach, engagement, and direct conversions.
Read more
The posts suggest this is a poor fit.
The posts contain no sponsorship offer.
The three posts are product and workflow write-ups from inside the team that builds Polar, describing an agent, custom attribution models, and shared metric labels.
They do not state rates, deliverable types, disclosure practice, audience terms, or how a brand would request a paid collaboration.
Resolve those details directly; the posts do not establish whether Dokes accepts sponsorships.
Do not base a placement on an assumed rate card.
The content points to a different reader.
The examples address ecommerce growth, analytics, and retention practitioners: a VP of Growth approving audit recommendations, a head of ecommerce commenting on attribution models, and a leadership team receiving a daily forecast flash.
A consumer product demonstration targets shoppers, and the posts do not show that audience.
Reach is not documented either.
The available follower figure is 23,871 on LinkedIn, a professional network, with no per-post engagement or demographic breakdown.
That is modest and insufficient to assess a reach-and-conversion buy.
The subject matter compounds the mismatch: prompting a custom attribution model and agreeing on a source-of-truth metric are not visual demonstrations, so a sponsored consumer post would have to be invented from scratch rather than built from what Dokes publishes.
If the consumer brand still wants to approach, ask directly for sponsorship terms, past placement examples, disclosure practice, and audience demographics.
Until those are confirmed in writing, this campaign should go to creators whose documented offer matches it.
A worked campaign example
Worked angle for the analytics and attribution need: a 60-minute session with a growth and analytics team, followed within a week by a one-page working document the brand owns.
Open with the sorting exercise, taken from the production use-case post.
Give the team ten minutes to write their own backlog and split it into "unlock" — finding money they cannot currently see — and "reduce" — work a person does today.
The published examples give them calibration: non-incremental ad spend, junk search terms, a failed-payment spike hiding behind a retention metric on the unlock side; Klaviyo drafts, supplier emails, and a daily forecast flash on the reduce side.
The point of the exercise is to leave with three candidate tasks rather than a general enthusiasm for agents.
Second block: pick the first task and specify it in the structure Dokes uses publicly — what the person asks, what the agent does, the output, the data and tools required, and where a human stays in the loop. Ask the brand to name the tool connections each candidate task would need, because in the published material the work runs by prompting an assistant connected to the company's numbers and tools, with output landing in Slack, a dashboard, or an existing marketing tool.
Third block: agree what the agent is allowed to answer.
This is the Labels problem — an AI can get the calculation right and pick the wrong metric, returning Google Analytics' conversion rate when the team means Shopify's.
Have the team name the source of truth for the three metrics their chosen task depends on, decide who can edit those definitions, and make them shared rather than sitting in one analyst's private project.
Putting governance before the money-finding examples means the first agent output is trusted rather than argued with.
Fourth block: attribution decision.
Use the framework from the custom-model post — rigid template versus a model built by prompting the platform to reflect how customers actually buy, after inspecting channels, touches, and time to purchase.
The brand should leave with a decision on which route to test and a sanity check they will run before shifting budget.
Note that the five-minute walkthrough referenced in that post is not included in the posts here, so ask Dokes to demonstrate the build live rather than promising the recording.
Close by writing the deliverable in the room: three tasks, each with an owner, the data and tool connections it needs, the human checkpoint, and a first-review date.
Add the metric definitions agreed in block three and the attribution route chosen in block four.
That document is the handoff.
Why this angle holds together: it follows the order of the published work — pick the task, connect the data, agree the definitions, then decide how the result will be judged — and every block ends in an artifact the brand keeps.
It also respects the tradeoff.
Dokes writes from inside a product team, so the session will be shaped by that tooling; a brand on a different stack should confirm before booking that the patterns transfer, and should treat any product-specific step as a demonstration rather than a neutral comparison.
If the brand wants hands-on implementation rather than a session, that is a different engagement and should be scoped and priced separately.
Choose Noah King when the brand’s first question is which AI analytics or reporting tool is worth paying for, rather than how to run agents inside its own team.
Read more
The reviewed work is a hands-on test of a commercial AI reporting tool: connecting Meta, Google, and Shopify accounts in a few clicks, asking which campaigns to scale or cut, then probing creative analysis and next-generation ad concepts.
The write-up sorts the audience explicitly — teams wanting ad accounts, ecommerce data, and AI in one place, versus agencies and media buyers needing cohort-level nuance — and names the monthly price band.
The same creator translates platform shifts, with a long plain-English walkthrough of Meta’s 2025 ranking, matching, and sequence systems and do-and-don’t lists aimed separately at enterprise businesses, large consumer brands, and small Shopify stores.
The tradeoff is tool evaluation versus operating depth.
Both posts assess someone else’s product or platform; neither shows an agent workflow built around a specific brand’s numbers and tools.
The task structure a session would need — what the person asked, what the agent did, where the output landed, and which human checkpoint applied — is not demonstrated.
One post was a compensated partnership that disclosed the arrangement and included the reviewer’s own criticisms, so read its judgements as stated opinion.
Ask for a sample session, rates, advisory availability, reach, and audience composition before booking.
Choose Will O’Brien when the brand’s core problem is that it does not trust the numbers it is deciding from.
Read more
The reviewed work is measurement criticism with the mechanics attached: a performance lead at a large apparel brand calling affiliate ROAS "fake," the last-click and self-reported attribution chain behind a 12x blended figure, the structural split between the team running affiliate and the team running measurement, and the reported result that 20%–50% of affiliate spend was not incremental once tested. A second post applies the same reasoning to AI-assisted discovery — why recommendation-driven visits arrive as direct traffic, and why branded search becomes a downstream signal of AI visibility rather than an upstream one.
The tradeoff is audit versus build.
O’Brien argues what the measurement stack gets wrong and prescribes incrementality testing and quarterly visibility audits; the posts do not lay out agent tasks a team could run on Monday morning or how to govern shared metric definitions across functions.
A brand that already accepts its attribution is inflated and wants help designing and running agent workflows should ask for that operating detail before engaging.
Choose Martin K. when the urgent question is whether AI agents can find and buy the brand’s products at all.
Read more
The posts cover the Universal Commerce Protocol and agentic commerce as a readiness checklist: activating the business-agent setting so inventory is treated as writable rather than read-only, enabling direct checkout so an agent can inject orders into the store backend, rewriting product taxonomy with conversational attributes so agents can match products to queries, and piloting direct offers. A companion post states the measurement consequence directly — an agent that skips the landing page and the pixel makes paid-driven sales look like direct traffic — and points to intent-level measurement and incrementality partners as the response.
The tradeoff is channel readiness versus internal analytics.
This work addresses the storefront, product-data, and purchase-path side of agentic commerce, plus the attribution blind spot that creates, rather than the analytics and attribution tasks a brand’s own team would hand to an agent.
The format is a diagnostic checklist with strong framing, so it suits a technical readiness sprint but offers less evidence of repeatable workflows and measured client outcomes.
Ask for examples of those workflows and outcomes before engaging.
David Dokes's three LinkedIn posts here cover a field guide to 15 AI-agent use cases already running at DTC brands, an introduction to custom attribution models in Polar, and a post on Labels for shared metric definitions.
All three introduce features of Polar, the measurement product Dokes writes about, so the customer results quoted in them are that team's account rather than independent evaluation.
The individual posts do not display publication dates, so shifts in emphasis over time cannot be assessed; a LinkedIn profile snapshot dated 8 October 2026 records 23,871 followers, and the posts here show no per-post engagement or audience demographics.
Similar-creator comparisons include two LinkedIn posts each from Noah King, Will O'Brien, and Martin K.
No books, shows, talks, transcripts, longer-form writing, or third-party profiles surfaced for Dokes.
The posts here do not state sponsorship terms, deliverable formats, or an inquiry route, and do not include later instalments of the promised use-case series or the five-minute attribution walkthrough referenced in one post.
