Why this matters
Small teams rarely decide to adopt AI; individuals open a tool, it works, and soon much of the week runs through prompts nobody has written down. The AI in marketing guide explains why that matters: AI moves the bottleneck from producing a first version to checking it. The AI applications section gives the method applied here: judge each task by its value, the cost of reviewing the output and the risk if a wrong output slips through.
The agency and its workload
Hirewell Lane Recruitment is an invented agency placing office, finance and warehouse staff locally. Its marketing team is a manager and an assistant. It uses a recruitment CRM holding every candidate and client record, an email platform, a social scheduler and a website content management system. It has no customer data platform (CDPUnified first-party customer profiles.); the CRM is the record of people. Each tool belongs to one of the marketing technology categories, and AI features can arrive inside any of them as well as in standalone assistants.
Hirewell Lane: the marketing team's recurring weekly tasks
| Task | Hours a week | Notes |
|---|---|---|
| Job adverts written from client role briefs | 6 | Salary, location and requirements come from the client |
| LinkedIn headlines and opening lines for job posts | 2 | Posted daily; the manager wants more versions to compare |
| Weekly candidate email of new vacancies | 2 | Built from the live vacancy list |
| Resizing the jobs-fair banner and weekly graphics for each channel | 2 | Same artwork, several sizes |
| Sorting website enquiries into candidate, client and supplier | 3 | Free-text form; some contain CVs |
| Summary of the weekly web and ads report for the directors | 1 | Directors read it on Monday |
| "Candidate of the week" spotlight posts | 2 | Assistant pastes the CV into a free chat tool for a draft |
| Placement case studies from client interviews | 3 | Each includes quotes from the hiring manager |
| Replies to public reviews of the agency | 1 | Some reviews are angry |
| Choosing which roles get paid promotion | 1 | Budget split agreed with the directors |
Automate, assist or keep human
- Automate. The tool produces the output, used with light or sampled checks. Only for internal or easily verified output where a mistake costs minutes.
- Assist. The tool drafts and a person checks every output against a source before use.
- Keep human. A person does the work; a tool may tidy prose but never supplies facts or judgement.
Personal data is a separate test across all three: a task can be low risk in its output and still be wrong to give a tool because of what goes into the prompt.
Task
- Make a table with columns for task, label, why, likely failure and review step.
- Judge each task on value, review cost and risk, and let that decide the label.
- For every automate or assist task, name the likely failure, the check that catches it and who runs it.
- Mark every task where personal data would enter a prompt and say what must change first.
- Choose the measure that will judge the pilot after a month.
Worked answer
| Task | Label | Why | Likely failure | Review step |
|---|---|---|---|---|
| Job adverts | Assist | High value; the client brief is a checkable source | Invented benefits; salary or requirements drifting from the brief | Recruiter who owns the role checks every fact against the brief |
| LinkedIn headlines | Assist | Cheap to judge, low risk | Over-promising hooks; the offer changed too | Manager checks each against the advert; vary one element at a time |
| Candidate email | Assist | Grounded in the vacancy list, which is easy to compare | A role that has closed, or one added from nowhere | Assistant ticks every role against the live list before sending |
| Resizing graphics | Automate | Output compared with the original at a glance | Logo or date cropped out of a size | Contact sheet of every size checked before scheduling |
| Sorting enquiries | Assist, after a data fix | Saves time, but the text holds CVs and contact details | Personal data in an unapproved tool; a client misfiled | Only in a tool approved for personal data; a person reviews uncertain calls |
| Report summary | Automate, trial only | Internal and reversible | A figure that does not match the report | Manager spot-checks figures; drop it if checking takes as long as reading |
| Candidate spotlight | Keep human, stop now | A CV pasted into a free chat tool leaves the agency's control | The vendor may keep the data | Stop; write spotlights from what the candidate agreed to share |
| Case studies | Assist | High value, but quotes must be real | Invented or misattributed quotes | Quotes checked against interview notes; client approves |
| Review replies | Keep human | Tone and accountability matter most; a wrong word is public | Defensive or generic replies | Manager writes and signs every reply |
| Paid promotion choices | Keep human | A spending decision tied to client commitments | Not applicable | Directors and manager decide as now |
The automate column is short on purpose: only resizing and an internal summary are checkable at a glance. Most of the week lands in assist, because it reaches people under the agency's name. The spotlight posts are the urgent finding: the problem is the input, not the output. Judge headline versions on applications started, not click-through rate (CTRShare of impressions that result in a click.) alone, since clicks from unsuitable candidates cost recruiters time.
The pilot KPIA metric tied directly to an objective. is time per acceptable output, review included, with errors caught at review.
Reading the pilot after four weeks
Job adverts before: 6 hours for 12 adverts, so 30 minutes each. With AI: 10 minutes to draft plus 12 to check, so 22 minutes each; 5 errors caught across 48 adverts. Report summary: 15 minutes to check against 20 to read the report, so the team drops it. Graphics: resizing for 4 channels now takes 25 minutes a week including the contact sheet, against 2 hours before.
Common mistakes
- Labelling by tedium. Enquiry sorting is dull and full of personal data.
- Automating published work. Anything candidates or clients see needs a person checking it against a source.
- Forgetting the input. A harmless output can rest on a prompt that should never exist.
- Review steps with no owner. Name who checks, against what.
- Counting drafting time only. A saving that ignores review time is not a saving. The beginner level overview shows how the next module turns assist tasks into briefs and reviews.
