Direct Answer
The useful lesson is not that AI has replaced six marketers. It is that six recurring workflows can now be split into a human judgment layer and an automated production layer. In Sam Oh's walkthrough, AI handles keyword expansion, SERP filtering, draft structure, competitor collection, brief assembly, prospect research, email drafting, and thumbnail exploration. The human still chooses the goal, supplies the expertise, defines quality, reviews evidence, and approves anything public.
That distinction is the difference between a dependable marketing system and a fast way to scale mistakes. Automate the repeated labor; keep accountability attached to a person.
Watch: Six Marketing Jobs Delegated to AI
Credit: This guide is based on Sam Oh's Ahrefs video, the supplied transcript, Letaido's public product information, and Ahrefs' published workflow research. The control model, implementation sequence, risk levels, and scorecard are editorial additions.
The Operating Model: Judgment Upstream, Labor Downstream
| Workflow | Let AI handle | Keep human | Risk |
|---|---|---|---|
| Keyword research | Expansion, filtering, clustering, SERP collection, report assembly | Business relevance, ranking realism, priorities, final roadmap | Medium |
| First drafts | Transcription, structure, transitions, initial prose | Angle, experience, factual review, voice, final edit | Medium |
| Competitor monitoring | Collection, deduplication, enrichment, matching to existing pages | What matters, whether to respond, strategic interpretation | Low |
| Content briefs | SERP retrieval, page extraction, topic and entity mapping, first brief | Search intent, originality, evidence plan, writer direction | Medium |
| Outreach preparation | Prospecting, rule-based vetting, contact lookup, draft creation | Audience, offer, claims, compliance, approval, relationship | High |
| Thumbnail concepts | Variations, composition studies, rough visual execution | Promise, taste, brand fit, accuracy, final design | Low |
The risk is not determined by how impressive the output looks. It is determined by the consequence of being wrong. A weak thumbnail concept is easy to discard. An irrelevant email sent to hundreds of people can damage sender reputation and trust.
1. Delegate Keyword Research, Not the Roadmap
Sam's Letaido workflow starts with a niche, expands seed terms into a much larger candidate set, filters irrelevant meanings, evaluates intent, reviews competing pages, groups topics, and produces an actionable report. In the golf example, the system removes unrelated meanings of driver instead of making a marketer clean the list manually.
This is a strong automation target because the data collection is repetitive and the output is inspectable. But a keyword report is not a strategy. The final selection still needs product relevance, realistic authority, click potential, commercial value, editorial capacity, and a reason your page would deserve to exist.
A safer acceptance test
- Trace every recommended topic to current search data and a visible SERP.
- Explain why the query belongs to the business and what action it should support.
- Reject branded noise, homonyms, publisher-only results, and formats the team cannot execute well.
- Sample at least 20 accepted and 20 rejected terms before trusting the filter at scale.
2. Replace the Blank Page, Not the Point of View
For scripts and articles, Sam starts with a voice conversation: the topic, argument, examples, and stories come from him. Claude organizes that material into an outline and helps assemble sections. A project containing older scripts, a style guide, and writing principles gives the model useful constraints. Sam then edits heavily and feeds the differences back into the system.
The repeatable pattern is source, structure, edit, learn. The author supplies original thinking and approved source material. AI proposes structure and prose. A human verifies every claim and rewrites until the work is accurate and recognizably theirs. The edits become training examples for the next run.
| Stage | Required input | Approval question |
|---|---|---|
| Source | Voice notes, interviews, research, examples, firsthand experience | Is the core idea actually ours? |
| Structure | Audience, desired action, format, successful references | Does the sequence make the argument easier to understand? |
| Draft | Style constraints, banned phrases, evidence boundaries | Is every specific claim supported? |
| Edit | Named owner and review checklist | Would that person stand behind every sentence? |
| Learn | Generated draft plus approved revision | Can the correction become a reusable rule or example? |
3. Turn Competitor Monitoring Into Triage
The competitor app described in the video gathers newly published posts into one inbox, records attributes such as author and publication date, enriches saved items with Ahrefs metrics, and matches each item to the closest pages on the user's own site. The marketer reviews a small decision queue instead of repeatedly checking dozens of tabs.
The right output is not a stream of competitor URLs. It is one of four dispositions: dismiss, watch, refresh, or build. Require the system to explain what changed, whether the site already covers the topic, and why action is warranted. This prevents competitor monitoring from becoming automated imitation.
4. Generate Evidence Plans, Not Generic Briefs
The demonstrated brief workflow reads current ranking pages and assembles keyword metrics, SERP composition, possible headlines, a slug, subtopics, named entities, frequently asked questions, and internal-link suggestions. That is useful groundwork, but a list of common headings is not yet a defensible page.
Add four fields before assigning the brief:
- Reader decision: what should someone understand or do after the page?
- Original contribution: what data, experience, test, example, or viewpoint will the page add?
- Evidence map: which claims require first-party, official, expert, or current sources?
- Stop condition: what would make the team decline to create the page?
A brief should help a skilled writer make better choices. It should not force every article into the average structure of the existing SERP.
5. Automate Outreach Preparation Carefully
Sam describes an agent that begins with an interview in Slack, applies his domain-quality criteria, finds contacts, and creates personalized drafts in Gmail. He retains a final send step as a guardrail. The video says this workflow achieved better response rates than a previously hired agency, but it does not publish the campaign definition, sample size, dates, or raw results. Treat that as a creator-reported comparison, not a general benchmark.
The phrase "worth millions" is also a statement about the potential scope of outreach across links, sales, and partnerships. It is not an audited revenue result from this workflow.
Use a staged permission ladder
- Research only: produce candidate organizations and explain the fit.
- Draft only: prepare messages without opening email access.
- Create drafts: write into a dedicated mailbox, but never send.
- Approved batches: a human reviews every recipient and message before sending.
- Limited automation: only after measured pilots, suppression rules, monitoring, and a fast kill switch.
Keep opt-outs, bounced addresses, do-not-contact records, sensitive categories, and existing relationships outside autonomous decisions. Follow the relevant law and email-provider rules for each audience and location.
6. Use Image AI as a Translation Layer
The final workflow does not primarily save time. Sam uses a thumbnail generator to turn a title, short description, and photo into three early concepts, then iterates on the strongest one in Gemini before handing a legible direction to a designer. AI translates an idea he could visualize but could not sketch clearly.
This is an excellent use of generative media: expand the option space and improve communication before expensive finishing work begins. Keep the final promise honest, check likeness and brand rights, and let the designer solve hierarchy, legibility, composition, and production quality rather than treating the generated image as automatically finished.
A Four-Week Implementation Plan
| Week | Work | Exit criterion |
|---|---|---|
| 1. Baseline | Choose one workflow, collect 10 normal examples, record time, cost, errors, and acceptance criteria | The team can describe a good output without mentioning the tool |
| 2. Assisted pilot | Run the agent manually with read-only or draft-only access | At least 80% of outputs are usable after bounded review, with no critical errors |
| 3. Feedback loop | Label failures, turn repeated corrections into rules, test difficult edge cases | Quality improves on a held-out set, not only the examples used to tune it |
| 4. Controlled operation | Schedule the workflow, assign an owner, log actions, cap spend, and define rollback | The process saves net time after review and remains easy to stop |
Start with competitor triage or brief assembly. Add account connections only when the pilot proves they are necessary. Letaido's public site says workspaces begin with no permissions and expose granular permissions, activity logs, and communication controls; use those controls deliberately rather than connecting the entire stack on day one.
Measure the System, Not the Demo
| Measure | Question | Failure signal |
|---|---|---|
| Net time | Does automation save time after review, correction, and maintenance? | The team spends longer fixing than doing |
| Acceptance rate | How much output survives human review? | High volume with low usability |
| Evidence quality | Can every material claim be traced to an appropriate source? | Invented facts, stale pages, or unsupported certainty |
| Business result | Did the workflow improve qualified traffic, responses, conversions, or production capacity? | Activity rises while outcomes stay flat |
| Control | Are permissions minimal, actions logged, and external communication reviewable? | Broad access, silent sending, or no rollback |
| Learning | Do corrections improve future runs? | The same failures repeat every week |
Watch the Related Videos
Complete AI SEO Course for Beginners
Build an AI Keyword Research Tool
The That Fit Friend SEO Recovery
Video Chapters
| Time | Section | Time | Section |
|---|---|---|---|
| 00:00 | Six jobs Sam no longer does manually | 04:28 | Competitor monitoring becomes triage |
| 00:24 | Automated keyword research | 06:25 | Generate a structured content brief |
| 02:05 | Voice-led first drafts | 08:20 | Outreach research and drafts |
| 04:03 | Feed edits back into the system | 10:10 | Thumbnail concept generation |
Sources and Further Reading
- Ahrefs: I Let AI Do These 6 Marketing Jobs (primary video and supplied transcript)
- Letaido: Official Product, Workflow, Security, and Pricing Information
- Ahrefs: Automated SEO, Agent Workflows, and Human Approval
- Ahrefs: AI Keyword Research Workflows and Prompts
- Ahrefs: The State of AI in Content Marketing
YouTube lists the primary video's publication date as 26 August 2026. This article was reviewed on 27 September 2026. Product features, model availability, integrations, permissions, prices, and app listings can change. Letaido's official site currently lists a $99 monthly plan and a possible free month for eligible Ahrefs customers; verify the current offer before purchasing. Workflow results and comparisons attributed to the video remain creator-reported unless a public underlying dataset is linked.