Upskill first in almost every agency under about 20 people. Training costs a fraction of a salary, and the people who know the clients do the AI work. Hire a specialist only when you have AI build work (automations, integrations, AI services you sell) to fill at least three days a week, or clients already paying for it.
The mistake is treating it as either-or. Most agencies that get this right do both in sequence: upskill everyone to use AI well in their own jobs, make one existing person the internal AI lead, and bring in specialist help by the project until the specialist work is steady enough to justify a role. A specialist hired into a team that hasn't learnt the basics spends their first six months answering "how do I" questions.
How the two options compare
| Criterion | Upskill the team | Hire an AI specialist |
|---|---|---|
| Cash cost | Low: many good courses are free; business AI seats cost about $20 to $25 each a month | A full salary plus employment on-costs, recruitment and equipment |
| Time cost | High and spread out: 15 to 25 hours per person over three months | Recruitment takes months; the specialist then needs to learn your clients |
| Client knowledge | Stays with the people doing the work | Has to be learnt; the specialist depends on account teams |
| Everyday AI use (drafts, reports, research) | Strong: the team applies AI inside its own work | Weak on its own: one person can't do everyone's drafting |
| Building automations and integrations | Limited unless someone has an aptitude for it | Strong, and the main reason to hire |
| Selling AI services to clients | Hard to deliver reliably | Makes it possible |
| Risk if it goes wrong | Uneven adoption; some people never change habits | An expensive hire with too little to do, or who leaves with the know-how |
Read down the table and a pattern shows. Upskilling wins on everyday AI use, which is most of the value in most agencies. Hiring wins on building things. So the question becomes: how much building work do you actually have?
What an agency AI specialist would spend the week on
Before deciding, write the job down as if you were advertising it. If you can't fill the list with real, current work, you don't have a role yet. A filled-in example for a fifteen-person agency, illustrative:
ROLE: AI and automation lead
Hours of real work available now (estimated per week):
Maintain and extend client reporting automations ...... 6 h
Build lead-routing and CRM automations for 3 clients .. 5 h
AI chatbot builds sold to clients (2 live, 1 pipeline) 6 h
Brand-voice Projects and prompt library upkeep ........ 2 h
Internal support and training for the team ........... 3 h
Evaluating new tools and features .................... 1 h
TOTAL ................................................ 23 h
Would this be billed to clients? About 11 h of the 23
Could an existing person do it with training? Prompt library and
support, yes. Automations and chatbot builds, not without months
of learning.Twenty-three hours of real work, half of it billable, is close to the three-days-a-week line. Six hours, mostly "keeping an eye on new tools", is not a job; it is a few hours of an existing person's week. Which agency tasks AI handles well can help you list the realistic work.
Putting a number on each option
Salaries vary too much by place and seniority to quote one here, so use your own figures in the same structure. With illustrative placeholders:
| Upskilling a ten-person team, first year | Amount |
|---|---|
| Course fees (OpenAI Academy and Claude Academy are free) | $0 |
| Staff learning time: 10 people x 20 hours x $40 internal cost | $8,000 of time |
| Internal AI lead: 4 hours a week for 13 weeks, then 2 hours a week | About 130 hours, $5,200 of time |
| Business AI seats: 10 x $20 a month on annual billing | $2,400 |
| Occasional outside help for one automation build | Use your quote |
| Total | About $15,600, most of it time rather than cash |
| Hiring a specialist, first year | Amount |
|---|---|
| Salary (placeholder; use your local rate) | $65,000 |
| Employment on-costs (ask your accountant; 20% used here) | $13,000 |
| Recruitment fee, if using an agency (usually a percentage of salary; ask) | Use your quote |
| Equipment and extra software | $3,000 |
| Team still needs basic training | Most of the $8,000 above |
| Total | About $85,000 or more, mostly cash |
The hire only pays if the specialist's billable work and the time they save the team cover that cost. At the fifteen-person agency above, 11 billable hours a week at a $90 charge-out rate is about $47,000 a year (at 47 working weeks), leaving roughly $38,000 to be covered by time saved and new work won. Possible, but not guaranteed. At an agency with six hours of real work a week, it can't add up. An AI consultant versus an in-house hire runs the same sums against outside help.
The freelance bill is the other number to watch, because it tells you when project help has quietly become a job. With illustrative figures: automation freelancers at $75 an hour, bought for 15 hours a week, cost about $53,000 a year over 47 weeks. That is most of the way to the $78,000 of salary and on-costs above, without the recruitment fee but also without the documentation, availability and client knowledge an employee builds up. When the freelance spend passes about two-thirds of a salary and client work is waiting on the freelancer's diary, start writing the job description.
A 90-day upskilling plan that actually changes habits
Training fails when it is a one-off afternoon. This plan spreads the hours and ties every session to live client work. Filled in for a ten-person agency, illustrative:
| Weeks | What happens | Hours per person |
|---|---|---|
| 1 to 2 | Everyone completes one free foundations course (OpenAI Academy or Claude Academy, matching your chosen tool); AI rules read and discussed | 4 |
| 3 to 6 | Weekly one-hour session on real work: each role practises its own task (reports, drafts, ad variants, pitch outlines) | 6 |
| 5 | Checking session: the team hunts for errors in five AI drafts | 1 |
| 7 to 10 | Each person picks one recurring task, redoes it with AI for four weeks, logs time before and after | 6 |
| 11 to 12 | Show and tell; working prompts go into the shared library | 2 |
| 13 | Review: who is ahead, who is stuck, what the agency builds next | 1 |
The weeks 7 to 10 log is where individual progress becomes visible. One account manager's entries for monthly client reports, illustrative:
| Week | Reports done | Minutes each | Needed a second pass? | Note |
|---|---|---|---|---|
| Before (from time tracker) | 4 | 170 | No | Commentary written from scratch |
| 7 | 4 | 150 | 2 of 4 | Draft invented a cause for a traffic dip |
| 8 | 4 | 115 | 1 of 4 | Added "say if the cause is unclear" to the prompt |
| 9 | 3 | 95 | 0 of 3 | Prompt shared with the other account managers |
| 10 | 4 | 90 | 0 of 4 | Stable |
Week 7 is slower than people expect and faster than before, and the drop after the prompt fix is typical. Someone whose log shows no change by week 9 needs a one-to-one session, not a second course.
Twenty hours per person over a quarter is enough for most people to change how they work. The week 13 review is where the hire question gets answered with evidence: by then you know which tasks the team handles comfortably and which keep needing someone with build skills. An illustrative review note from a ten-person agency:
WEEK 13 REVIEW
Comfortable now: report commentary (all 3 account managers),
first drafts (3 writers), ad variants (paid-media lead)
Still stuck: 2 people rarely use it; 1:1 sessions booked
AI work we couldn't do this quarter:
- client asked for a website chatbot: declined
- lead-routing automation for 1 client: freelancer, 9-day wait
- reporting automation broke twice: freelancer fixed, 3 h each
Estimate of build work per week if we said yes to all: 8-10 h
Decision: no hire yet. Keep freelancer on a monthly retainer;
revisit at week 26 or if build work passes 20 h a week.Eight to ten hours of build work is a retainer, not a role. The same note at twenty-plus hours, with clients waiting, would point the other way. Training staff to use AI in a small business has more on running the sessions.
Picking the internal AI lead
The internal lead is the piece of the hybrid most agencies choose badly. They pick the most enthusiastic person, who is often the one most willing to trust the output. A better test is a short practical exercise, the same one you'd give a specialist candidate. Hand two or three volunteers this task:
Here is a real (anonymised) client brief for a 600-word website post
and last month's performance figures. In 45 minutes:
1. Write the prompt you'd give the team for the first draft.
2. Run it and show us the draft.
3. List everything in the draft you would check or change before a
client saw it, and why.A weak answer, illustratively, produces a polished draft and a short list: "tighten the intro, add a call to action". A strong answer produces a plainer draft and a longer list: "the second paragraph claims a 40% industry growth figure with no source, so remove or source it; the tone is closer to our other fintech client than this one; the brief said no mention of pricing and the draft mentions 'affordable plans'; the figures section rounds conversions up from 33 to 'nearly 40'."
The strong candidate is the lead. The skill that matters most in the role is noticing what AI gets wrong, because that is what they will be teaching everyone else. Give the chosen person protected time (three to four hours a week during the first quarter) and make it part of their review, or the role quietly disappears under client work.
Two agencies, two decisions
A seven-person content and social agency. Its AI work is almost entirely drafting, editing, research and reporting commentary. Nothing it sells needs integrations. The owner had interviewed two "AI specialists" before running the numbers. The decision was to upskill: the 90-day plan, a senior writer as AI lead with three hours a week protected, and a freelancer for one reporting automation. First-year cash cost was about $1,700 in seats plus the freelancer's fixed fee, against a hire that would have cost many times that in cash alone.
The freelancer's brief did the job a permanent hire's documentation habits would otherwise do. Besides the working automation, it asked for:
- everything built in accounts the agency owns, with the freelancer added as a user, never the other way round;
- a one-page runbook: what each step does, what triggers it, and the three most likely ways it breaks;
- a recorded 30-minute handover call with the AI lead, who then made one small change alone to prove they could;
- a fixed price for the build and an hourly rate for fixes, so a broken report in month eleven had a known cost.
When the reporting automation did break, after a platform changed an export column name, the AI lead fixed it from the runbook in twenty minutes without calling anyone.
An eighteen-person performance and CRM agency. Clients were already asking for AI chatbots on their websites and automated lead routing into their CRMs, and the agency had turned down two such projects in a quarter. It still upskilled the whole team first, but it also hired. It wrote the job down, found 26 hours a week of real work with 14 billable, and recruited someone with automation experience rather than a general "AI" title. The specialist's first month was spent on the two turned-down projects, not on training colleagues, because the team had already had its training.
If you do hire, what to test for
- A build they can show and explain. Ask candidates to walk through one automation or AI integration they built: the tools, what broke, and how they found out. Vague answers about "AI strategy" are a warning.
- Documentation habits. Ask to see how they documented that build. If the answer is "it's all in my head", expect trouble when they leave.
- Cost awareness. A good specialist knows that automation platforms bill per task or credit, and that AI calls add up. Ask how they'd keep a client's automation bill predictable. A useful test question: "A Zapier automation has four action steps and runs on 300 new leads a month. Which plan does it need?" A strong candidate works out that each successful action step is a task, so that is 1,200 tasks a month, past the 750 on Zapier's Professional entry tier, and suggests moving a check into a filter step, which uses no tasks, before upgrading.
- Client-facing judgement. They will explain limits to clients. Give them a scenario where a client wants a chatbot to quote prices, and listen for whether they push back. A weak answer: "Sure, we'd load the price list into it." A strong one asks where prices come from, how often they change, and what happens when the bot quotes an out-of-date figure, then suggests the bot gives price ranges from an approved page and hands firm quotes to a person.
Where agencies go wrong either way
- Hiring a title. "Prompt engineer" and "AI strategist" roles often turn out to be one person writing prompts for others. Whether a small business should hire a prompt engineer explains why that rarely lasts. Hire for the build skills you need, such as automation platforms and CRM integrations.
- Hiring before the work exists. If the job list above comes out at under two days a week, you are hiring on hope.
- Letting the know-how leave with one person. A realistic failure: a specialist builds a reporting automation spanning three tools, documents none of it, and leaves after ten months. A client report breaks in month eleven and nobody can fix it. Whoever builds, the agency should own written documentation of every automation. Hiring your first AI automation specialist covers the handover terms to agree up front.
- Upskilling without protected time. Twenty hours of learning won't happen in the gaps between client deadlines. It shows up quickly: an agency that left training to "quiet weeks" found at week 6 that three of ten people had finished the foundations course, and the shared prompt library held two prompts, both from the same person. Put the sessions in diaries and treat them like client meetings.
- Expecting a hire to change culture. One specialist can't make nine people change habits. That comes from the directors using AI visibly themselves and asking about it in reviews.
A quick way to decide for your own agency: if your week 13 review shows that the unmet AI work is mostly "we'd like to use it more", keep upskilling. If it is "clients want things built and we keep saying no", that is when the specialist role writes itself.
Further reads
- AI Rollout Plan for a Ten-Person Marketing Agency — A ten-person agency's first quarter with an internal AI lead.
- How Much Time and Money Does AI Staff Training Take? — More detail on the hours behind the upskilling plan.
- AI or a New Hire? How to Decide Before You Recruit — The broader version of this decision for any small business.
- How to Choose an AI Training Provider for Your Team — If you want outside training rather than free courses.
- AI Content Workflow for Agencies: From Brief to Approved Draft — The workflow an upskilled content team should be running.
- How to Hire an AI Automation Freelancer on Upwork or Fiverr — The project-by-project alternative to a permanent hire.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI Academy help article; Claude Academy course pages; OpenAI ChatGPT Business and Anthropic Claude Team pricing. Checked September 2026. Salary, on-cost and hour figures in the examples are illustrative placeholders, not market rates.