Chapters · 9
What is an AI agent, and how is it different from a workflow or a chat assistant?
An AI agent decides its own next step within the tools and limits you give it; a workflow follows steps you wrote in advance and starts on a trigger; a chat assistant waits for a person to type.
| Chat assistant | Workflow with AI steps | AI agent | |
|---|---|---|---|
| Who decides the steps | The person, turn by turn | You, in advance | The model, inside your limits |
| What it can touch | What you paste in | Only what each step names | Any tool on its list, as often as it chooses |
| Best agency jobs | One-off drafts and questions | Repeat jobs: reports, updates, reminders | Diagnosis, research, prep packs |
| Main risk | One client's data pasted into another's chat | A wrong step repeats on every client | Loops and wrong tool calls |
| Cost shape | A seat price | A few AI calls a run | Many calls, each re-reading the context |
Here "agentic" means agents doing the agency's own work, not buying media. Our rule: if you can write the steps down, build a workflow, and use an agent only where the next step depends on what it finds. Eleven of the 15 cards below are workflows; four hand one part to an agent.
AI agents for marketing agencies: 15 workflows at a glance
The set starts from the most common agent tasks: in Basis's 2026 survey of agencies, brands and publishers, reporting and analytics came first, then campaign briefs, then content (Basis, 2026, via PPC Land).
| # | Workflow | Built as | Human gate | Cost per run |
|---|---|---|---|---|
| 1 | Monthly report draft | Workflow | Account manager approves | Medium |
| 2 | Weekly client update | Workflow | Every send | Small |
| 3 | Performance watch and alert | Code, then agent | Buyer makes any change | Large, per alert |
| 4 | Budget pacing check | Workflow | Buyer changes budgets | Small, per flag |
| 5 | Ad copy variants | Workflow | Lead picks, client approves | Medium |
| 6 | Creative brief | Workflow | Lead signs off | Medium |
| 7 | UGC script | Workflow | Lead and client approve claims | Medium |
| 8 | SEO content brief | Agent | SEO lead approves | Medium to Large |
| 9 | Ranking and technical watch | Workflow | SEO lead triages | Small |
| 10 | Competitor watch | Agent | Strategist reviews | Large |
| 11 | Onboarding intake to client record | Workflow | Lead confirms each fact | Medium to Large, once |
| 12 | Meeting notes to tasks and promises | Workflow | Call owner approves | Large |
| 13 | Out-of-scope request check | Workflow | Lead decides | Small |
| 14 | Renewal prep pack | Agent | Director and owner | Large |
| 15 | Overdue invoice reminder draft | Workflow | Finance sends | Small |
Cost per run uses three run sizes we estimated, priced at Claude Sonnet 5.5 list prices and worked out below: Small $0.015 to $0.07, Medium $0.08 to $0.30, Large $0.35 to $1.15.
Download all 15 spec cards (Markdown, printable)
AI agent permissions and client data: the guardrails every agency needs
Every agent needs six things before it touches a client: one client per run, the least access that does the job, a human in the loop on every send and spend, hard limits, a log, and a safe way to fail.
Gates exist because agents make mistakes: in Salesforce researchers' May 2025 benchmark, the best agents completed about 35% of multi-turn CRM tasks (Salesforce AI Research, 2025). Newer models may do better, and the tasks were simulated CRM work, not agency work.
- One client per run. An agent working for client A reads only client A's records, documents and accounts, and never falls back to another client's files. The market research bodies MRS and ESOMAR, writing for research agencies, start from the default that client data belongs to the client and ask them to write down which data may go into which AI tools.
- Least privilege. Give each agent its own connection at the lowest level that works, usually reporting access, never an admin's login. The client onboarding checklist lists the lowest role per platform.
- A person approves anything that reaches a client or spends money (defaults in the table below).
- Spend caps and step limits. Cap each run in dollars, tool calls and minutes, and each AI key per month. A large agent run on Opus 5.5 can reach $2.30 at list price, so a $2 cap stops it first.
- An audit log of every run. Trigger, data read, tool calls, model, tokens, cost, output and approver. Blank raw tool data after a set period.
- Safe failure. A step that already sent something never re-runs by itself, and a run refuses to draft from data older than a day, naming the stale source.
| Action | Examples | Default gate |
|---|---|---|
| Read | Pull numbers, read a transcript | None; logged |
| Draft | Report text, brief, reply | None to write; a person decides what to use |
| Internal change | Create a task, update a record | Allowed; logged, reversible, spot-checked weekly |
| Client-facing send | Email, Slack message, report link, post | A person approves each send, on the exact version |
| Spend or account change | Budget, bid, campaign status, new ad | A person makes it; the agent's login is read-only |
| Delete | Records, files, campaigns | Off for agents |
Download the agent permission policy template (Markdown)
AI agent use cases for agencies: 15 build specs
The cards run from reporting to invoices; monthly figures assume an example agency with 18 retainers.
1. Monthly report draft
| Trigger | The 1st of the month per client, after the overnight data pull |
|---|---|
| Data it reads | Two months of stored figures per channel, goals, last approved report, call notes |
| Agent or step | Workflow. Code makes every figure; one AI step writes the words; a check drops any line with a number code did not make |
| Human gate | Account manager approves before the client sees it |
| Output | Draft summary: what moved, why, what we do next (the full reporting method) |
| Cost per run | Medium; 18 clients $1.44 to $5.40 a month |
| ROI metric | Minutes from data ready to approved; share approved with light edits |
| Failure modes | A misread figure; a source that stopped overnight; a cause nobody confirmed (write "cause not known") |
| Variants by agency type | Performance: CPA, ROAS. SEO: clicks, pages that moved. Social: reach, top posts. Creative: work shipped |
2. Weekly client update
| Trigger | Thursday afternoon in the client's time zone |
|---|---|
| Data it reads | Tasks done and due, items waiting on the client, key numbers, call notes |
| Agent or step | Workflow. Code gathers the lists; one AI step writes done, next, and what we need from you |
| Human gate | Every send |
| Output | Draft email or Slack message |
| Cost per run | Small; 78 runs $1.17 to $5.46 a month |
| ROI metric | Updates out by Friday noon; "where are we on X?" messages from clients |
| Failure modes | A moved task reported as done; a date nobody agreed; another client's item |
| Variants by agency type | Performance: spend against pace. SEO: pages published. Social: posts awaiting approval. Creative: rounds with the client |
3. Performance watch and anomaly alert
| Trigger | Daily; AI runs only when a code rule trips, such as CPA over target three days running |
|---|---|
| Data it reads | Daily figures by campaign, targets, a 28-day baseline, the team's change log |
| Agent or step | Code detects; a read-only agent checks recent changes, splits the data and tests tracking |
| Human gate | Media buyer makes any change in the platform |
| Output | Alert: what moved, likely cause with evidence, proposed fix |
| Cost per run | Large; 10 alerts $3.50 to $11.50 a month (540 Small daily checks instead: $8.10 to $37.80) |
| ROI metric | Hours from anomaly to action; false alerts a week |
| Failure modes | Alert fatigue; broken tracking read as a performance drop; a confident cause with no evidence |
| Variants by agency type | Performance: CPA, ROAS, spend. SEO: clicks or indexed pages falling. Social: engagement drops |
4. Budget pacing check
| Trigger | Daily, per account with a monthly budget |
|---|---|
| Data it reads | Month-to-date spend, budget, days gone, planned heavy weeks, agreed changes |
| Agent or step | Workflow. Code works out pace; a small AI step notes flagged accounts. Example: $6,200 of $10,000 by day 15 of 30 is 124% of pace; landing on budget needs $253.33 a day, not $413.33 |
| Human gate | Buyer changes budgets; going over the agreed budget needs the client's written yes |
| Output | Off-pace accounts with the daily spend that lands on budget |
| Cost per run | Code: none. Note: Small per flagged account |
| ROI metric | Accounts ending within your threshold of budget |
| Failure modes | An old budget figure; heavy weeks ignored; platform reporting delay; mixed currencies |
| Variants by agency type | Performance and paid social: core. SEO: content hours against the retainer. Creative: production spend against the quote |
5. Ad copy variants from the brief
| Trigger | By hand, when a brief or test plan is approved |
|---|---|
| Data it reads | Brief, brand voice, banned words, past winners, character limits, claims the client may not make |
| Agent or step | Workflow. One AI step writes variants per angle; a check tests limits, banned words and claims |
| Human gate | Copywriter picks; client approves before anything goes live |
| Output | Table: headline, body, angle, hypothesis, character counts |
| Cost per run | Medium per batch |
| ROI metric | Minutes per approved variant; tests that find a winner |
| Failure modes | Claims the client cannot make; one idea in five wordings; wrong market |
| Variants by agency type | Performance: search and social ads. SEO: titles and meta descriptions. Social: captions. Creative: taglines |
6. Creative brief from client context
| Trigger | By hand, when a request or campaign is approved |
|---|---|
| Data it reads | Brand, audience, offer, past results and lessons, the request, specs, dates, budget |
| Agent or step | Workflow. One AI step fills a fixed template and lists gaps as questions instead of guessing |
| Human gate | Account lead signs off before the creative team starts |
| Output | One-page brief: objective, audience, message, mandatories, deliverables, dates, questions |
| Cost per run | Medium |
| ROI metric | Revision rounds per deliverable; briefs sent back for missing facts |
| Failure modes | Guesses filling gaps; lessons from another client; old brand facts (show each fact's source) |
| Variants by agency type | Performance: ad sets and sizes. Social: monthly themes. Creative: concept brief |
7. UGC script (creator brief)
| Trigger | By hand, when a creator is booked |
|---|---|
| Data it reads | Product facts, allowed claims, customers' words from reviews, format, the creator's style, usage rights |
| Agent or step | Workflow. One AI step writes three hooks and one or two scripts with shot notes; a check tests claims and the ad disclosure |
| Human gate | Lead approves; client approves claims |
| Output | Hook, body, call to action, shot list, disclosure, do-not-say list |
| Cost per run | Medium |
| ROI metric | Scripts filmed without a rewrite; days from booking to script |
| Failure modes | Sounds like an ad, not a person; unsupported claims; no disclosure |
| Variants by agency type | Performance: hooks from the test plan. Social: organic creator posts. Creative: hero scripts |
8. SEO content brief
| Trigger | By hand, when a topic is approved for the month |
|---|---|
| Data it reads | Query and intent, top results' headings from your SEO tool, the client's pages on the topic, Search Console queries |
| Agent or step | Agent with read-only research tools, or a workflow if the data is pasted in |
| Human gate | SEO lead approves before a writer starts |
| Output | Keywords, intent, outline, questions, facts needing sources, what top results miss |
| Cost per run | Medium with pasted data; Large when the agent researches |
| ROI metric | Minutes per brief; writer questions per brief |
| Failure modes | Invented search volumes (no number without an export); a keyword another page owns; a copied outline |
| Variants by agency type | SEO: core. Performance: landing-page briefs. Creative: long-form campaign content |
9. Ranking and technical watch
| Trigger | Weekly, after the Search Console pull and your rank tracker's and crawler's exports |
|---|---|
| Data it reads | Clicks and position by page; tracked positions; crawl errors, noindex and canonical changes |
| Agent or step | Workflow. Code flags drops against a 28-day baseline; a small AI step groups them into tickets |
| Human gate | SEO lead triages; the account lead sends developer requests |
| Output | Tickets: what dropped, which pages, likely cause, evidence, fix |
| Cost per run | Small per digest |
| ROI metric | Days from drop to ticket; days from ticket to fix |
| Failure modes | Normal swings flagged; a Google update blamed on the site; a crawl that never ran |
| Variants by agency type | SEO: core. Performance: broken landing pages in live ads. Creative: launch checks |
10. Competitor watch
| Trigger | Weekly or monthly, for clients who pay for it |
|---|---|
| Data it reads | A fixed list of competitors and sources: ad libraries, pricing pages, posts, shared rankings |
| Agent or step | Read-only agent held to the source list and an output template |
| Human gate | Strategist reviews before anything goes in a client deck |
| Output | At most five items: what changed, link, why it matters, suggested response |
| Cost per run | Large; 18 clients monthly $6.30 to $20.70, weekly $27.30 to $89.70 |
| ROI metric | Findings used in client plans or meetings a month |
| Failure modes | Made-up activity (every claim needs a link); noise; a page that plants instructions (prompt injection); scraping against a site's terms |
| Variants by agency type | Performance: ad creative and offers. SEO: rankings and new content. Social: formats. Creative: launches |
11. Onboarding intake to client record
| Trigger | Deal won, or the intake form comes back |
|---|---|
| Data it reads | Intake answers, sales notes, proposal and scope, website, brand files |
| Agent or step | Workflow. One AI step proposes facts (offer, audience, competitors, brand rules, goals, contacts), each with its source |
| Human gate | Account lead confirms facts row by row before any agent reads them |
| Output | Confirmed facts, plus missing items and access still owed |
| Cost per run | Medium to Large, once per client |
| ROI metric | Days from signed to kickoff-ready; questions asked twice in the first 30 days |
| Failure modes | Old website facts; sales promises filed as scope; a guess recorded as fact |
| Variants by agency type | Performance: pixels, conversion definitions. SEO: Search Console, CMS access. Social: handles, approvers. Creative: brand kit |
12. Meeting notes to tasks and promises
| Trigger | When the notetaker's transcript is ready |
|---|---|
| Data it reads | Transcript, attendees, the client record, open tasks and promises |
| Agent or step | Workflow. One AI read pulls decisions, tasks with owners, promises with dates and scope asks; a check matches each to a transcript line |
| Human gate | Call owner approves before anything is filed; the recap email is a separate send |
| Output | Tasks, promises (ours and the client's), a flagged scope ask, a draft recap |
| Cost per run | Large, since transcripts are long; 40 calls $14 to $46 a month |
| ROI metric | Promises with an owner and date within 24 hours; promises broken a month |
| Failure modes | Wrong speaker labels; "we'll look into it" turned into a deadline; a two-client call filed to one |
| Variants by agency type | The same for all; creative teams also count feedback as revision rounds |
13. Out-of-scope request check
| Trigger | A client request is logged from email, Slack, a form or a call |
|---|---|
| Data it reads | The request; the signed scope (included, allowances, exclusions, revision limits); this month's usage; prices |
| Agent or step | Workflow. One AI step quotes the matching clause and suggests included, uses allowance, extra or not offered |
| Human gate | Account lead decides; a price goes to the client only after approval |
| Output | Suggested answer with the clause; for an extra, a draft change order (scope-creep scripts and form) |
| Cost per run | Small |
| ROI metric | Extras priced or declined against extras done free; change-order value a month |
| Failure modes | Vague scope read too kindly or too harshly; an allowance already used up |
| Variants by agency type | Performance: new channels. SEO: extra pages. Social: extra posts. Creative: extra rounds and sizes |
14. Renewal prep pack
| Trigger | A set number of days before the notice date, such as 60 |
|---|---|
| Data it reads | A year of results against goals, delivered against sold, promises kept and broken, what calls said, margin (finance only) |
| Agent or step | Read-only agent drafts results, wins, misses with reasons, and scope and price options |
| Human gate | Account director and owner choose before anything reaches the client |
| Output | Internal brief and a draft client review outline |
| Cost per run | Large, a few times a year per client |
| ROI metric | Renewal decisions made before the notice date; hours to build the pack |
| Failure modes | Cherry-picked wins; margin in the client version; a broken promise the pack forgets |
| Variants by agency type | Performance: CPA and ROAS trend. SEO: a year of traffic. Social: growth. Creative: projects and rounds |
15. Overdue invoice reminder draft
| Trigger | An invoice passes its due date by a set number of days |
|---|---|
| Data it reads | Invoice, payments recorded, billing contact, past reminders, any open dispute |
| Agent or step | Workflow. Code checks dates and that no payment is recorded; a small AI step drafts a reminder at the right firmness |
| Human gate | Finance or the owner sends, every time |
| Output | Draft email with the invoice details and your accounting tool's payment link |
| Cost per run | Small |
| ROI metric | Days sales outstanding; share paid within seven days of the first reminder |
| Failure modes | A payment that arrived but was not recorded; wrong currency; a reminder during a dispute |
| Variants by agency type | Retainers chase monthly; creative and project work chase milestones |
How much does an AI agent cost to run?
Most agency runs cost cents at list prices; review time costs far more than the model.
The method: for each AI call, multiply input tokens by the input price and output tokens by the output price, divide by 1,000,000, and add up the calls. On Claude Sonnet 5.5 ($2 input, $10 output per million tokens), a run that reads 40,000 tokens and writes 3,000 costs $0.08 + $0.03 = $0.11. If 30,000 of those input tokens are a cached prefix read at $0.10 per million, it costs $0.003 + $0.02 + $0.03 = $0.053, after a cache write of $0.075 that lasts five minutes (Anthropic pricing, checked 4 November 2026).
Agents cost more because every tool call is a new AI call that re-reads everything so far. Example: 15 calls, each sending a 10,000-token prefix plus 10,000 tokens of work so far, writing 8,000 tokens in all, is 300,000 input tokens: $0.60 + $0.08 = $0.68 on Sonnet 5.5. Cache the prefix (one write at $0.025, 14 reads for $0.014, while the other 150,000 tokens still cost $0.30) and it falls to $0.42.
| Run size (tokens in / out) | Haiku 5.5 | Sonnet 5.5 | Opus 5.5 | GPT-5.6 Luna | GPT-5.6 Terra |
|---|---|---|---|---|---|
| Small: 5k to 25k / 0.5k to 2k | $0.0008 to $0.0035 | $0.015 to $0.07 | $0.03 to $0.14 | $0.0016 to $0.0074 | $0.016 to $0.074 |
| Medium: 30k to 120k / 2k to 6k | $0.004 to $0.015 | $0.08 to $0.30 | $0.16 to $0.60 | $0.0084 to $0.031 | $0.084 to $0.31 |
| Large: 150k to 500k / 5k to 15k | $0.018 to $0.058 | $0.35 to $1.15 | $0.70 to $2.30 | $0.036 to $0.12 | $0.36 to $1.18 |
List prices checked 4 November 2026 (Anthropic and OpenAI); Haiku 5.5's applies to prompts up to 100,000 tokens. GPT-5.6 Sol is left out because its price that day was a temporary promotion. Models count the same text as different numbers of tokens, so the rows compare prices, not which model is cheaper. Re-check prices when you budget.
To choose the tools, see AI tools for agencies, compared by job.
How do you measure the ROI of an AI agent?
Measure minutes saved per run against review time, upkeep and model cost, from a baseline you record before the agent goes live.
In Basis's 2026 survey of 184 marketing and advertising professionals at agencies, brands and publishers, "just 6.5% of marketing organizations measure the ROI of their AI investments against defined benchmarks, and 65.1% either do not measure it at all or are unsure whether anyone at their organization is doing so" (Basis, October 2026, via MarTech Cube). That measures practice, not how well agents work.
Each card names its ROI metric. Beside it, record baseline minutes per task, review minutes per run, runs a month, model cost and upkeep hours. Then:
Net value a month = runs × (baseline minutes − review minutes) ÷ 60 × loaded hourly cost − model cost − upkeep hours × loaded hourly cost.
Track quality too: drafts approved with light edits, errors caught at the gate, and errors that reached a client (target zero).
The model bill is under 1% of the value; review time decides the result. So measure your own baseline: in AgencyAnalytics' 2026 survey, 46% of the 473 agency professionals who answered build a client report in under 30 minutes (AgencyAnalytics, 2026; a reporting-software vendor's audience, many already automating), so some agencies have less to save than they assume.
Which AI agent should an agency build first?
Start with a high-hour, low-risk card where a mistake stays inside the agency: the report draft, meeting notes or the scope check. Leave anything near ad spend for later.
- Record the baseline for two cycles before switching anything on.
- Run it on one client in a test mode that holds every send, and read every output.
- Add clients only when errors at the gate fall and none reach a client.
If your numbers sit in a dozen tools, map them first with the agency tech stack guide.
Where RevSmith fits (and where it does not)
RevSmith makes this guide and appears in it. We list it where it fits and say where it does not.
If every card you want needs the same client record, permissions and run log, you need one shared layer, not 15 tools: see what an AI operating system for agencies is, and when it is overkill.
In RevSmith an agency builds its own agents and workflows. An agent has a purpose, instructions, tools, skills and a client scope, and by default stops at 60 tool calls, 20 minutes or $2 a run. It runs with the rights of the member who owns it and reads only its own client's Client Brain (its confirmed facts) unless it is an all-clients agent. "Allow sending" is off by default for every step and agent; in a chat, any change an agent wants to make in an app waits as a button you press.
- Backed today: card 1 (code makes every figure, AI writes the words, a person approves; RevSmith drafts it on the 1st and never sends it by itself) and card 11 (AI suggests facts from the sources you add; a person confirms row by row; a won deal does not start it). Live reads from ad platforms are not yet proven on a customer account.
- Pieces only: cards 2 to 8, 12, 13 and 14. Monitors compare measures with targets but have not run on a live ad account; a person records the scope answer. Card 12 is the Meeting follow-through pack, where nothing is filed until a person approves; its notetaker read is not yet proven on a customer account, and its call read is not yet proven on a real AI model.
- Not built: rank tracking and crawling (card 9 gets only Search Console's daily read), competitor watching (card 10) and reminders (card 15: invoices you record read overdue; nothing is sent).
- Edge to know: ad changes, such as a budget or a campaign switched on, are not treated as sending, so "Allow sending" does not hold them, and a send in an app outside the social and email-marketing group can slip past if its tool name has no send word. Give agents reporting-level ad access if spend must stay with a person.
RevSmith is not a PSA, a time tracker or an ad platform. We set terms with each agency during onboarding; book a demo to ask.
Sources
- Forrester with the 4A's, "Forrester: Nine In 10 US Marketing Agencies Use AI To Cut Costs At The Expense Of Creativity", press release for The State Of AI Inside US Marketing Agencies, 2026, 24 June 2026. Sample: nearly 200 US agency decision-makers at VP level or above, polled April 2026, as reported by MediaPost (7 July 2026). Population: US marketing agencies. https://www.forrester.com/press-newsroom/forrester-nine-in-10-us-marketing-agencies-use-ai-to-cut-costs-at-the-expense-of-creativity/ and https://www.mediapost.com/publications/article/416324/report-agencies-using-ai-to-boost-margins-are-sac.html. S53 (Verified-primary for "nine in 10"), S54 (Verified-primary), S57 (Verified-secondary).
- Digital Applied (sells AI services), "Agentic AI adoption survey 2026: 250 agencies", 26 April 2026. Sample: 250 marketing and development agencies, $1M to $50M revenue, 8 to 180 staff, US, EU and APAC, January to March 2026. https://www.digitalapplied.com/blog/agentic-ai-adoption-survey-2026-250-agencies. S106 (Verified-primary).
- Basis, AI and the Future of Marketing, fourth edition, October 2026, via MarTech Cube (21 October 2026) and PPC Land. Sample: 184 marketing and advertising professionals at agencies, brands and publishers, August 2026. https://www.martechcube.com/marketers-double-daily-ai-usage-but-performance-measurement-falls-behind/ and https://ppc.land/chatgpt-drops-to-74-5-of-marketers-as-claude-climbs-to-72-3-basis-finds/. S60 (Verified-secondary), S63 (Verified-secondary; task order only), S64 (Changed: population and date).
- Salesforce AI Research, "CRMArena-Pro: Holistic Assessment of LLM Agents Across Diverse Business Scenarios and Interactions", arXiv 2505.18878, May 2025. Sample: a benchmark of 19 expert-validated tasks (4,280 instances) in sales, service and configure-price-quote, on models available in May 2025; AI models, not agencies. https://arxiv.org/pdf/2505.18878. S141 (Verified-primary).
- Gartner, forecast that over 40% of agentic AI projects will be canceled by the end of 2027, 25 June 2025, via The Star (Reuters). An analyst forecast, not a measurement. https://www.thestar.com.my/tech/tech-news/2025/06/25/over-40-of-agentic-ai-projects-will-be-scrapped-by-2027-gartner-says. S138 (Verified-secondary).
- AgencyAnalytics (sells reporting software), 2026 Marketing Agency Benchmarks Report, 30 June 2026. Sample: 494 marketing agency professionals (45% owners or leaders), mostly US, Australia, UK and Canada, February to April 2026; 473 answered the report-time question. https://agencyanalytics.com/agency-benchmarks-2026. S74 (Verified-secondary), S69 (Changed: unit).
- Anthropic, Claude API pricing, checked 4 November 2026: Sonnet 5.5 $2 input, $10 output, $0.10 cache reads per million tokens (live page opened 5 November 2026); Opus 5.5 $4 and $20; Haiku 5.5 $0.10 and $0.50 up to 100,000-token prompts; cache writes 1.25x input (five minutes). https://platform.claude.com/docs/en/about-claude/pricing. S90 (Changed: cache reads $0.10, the live page; input and output Verified-primary), S91, S92, S95 (Verified-primary).
- OpenAI, model pages, checked 4 November 2026: GPT-5.6 Terra $2 input and $12 output per million tokens, Luna $0.20 and $1.20 (both since the cuts of 30 July 2026). GPT-5.6 Sol is not priced on this page because its price that day was a promotion running until at least 21 November 2026. https://developers.openai.com/api/docs/models/gpt-5.6-terra, https://developers.openai.com/api/docs/models/gpt-5.6-luna, https://developers.openai.com/api/docs/models/gpt-5.6-sol. S113, S114 (Verified-primary); S112 (Verified-primary, for the promotion only).
- MRS and ESOMAR, "Client Data, AI, and the Boundaries of Consent: A Practical Note for Industry Leaders", 2026 (no date on its first page; the PDF file is dated 2 June 2026). Guidance for market research agencies; no figure used. https://www.mrs.org.uk/pdf/MRS%20and%20Esomar%20Client%20Data%20AI%20Statement%20for%20ISSUE.pdf. No S-ID (guidance only).
- Nine in 10 US marketing agencies use generative AI, and half use agentic AI for marketing execution (Forrester and the 4A's, June 2026).
- 41% of the 250 marketing and development agencies Digital Applied surveyed in early 2026 had shipped at least one AI agent (Digital Applied, April 2026; the publisher sells AI services).
- Just 6.5% of marketing organizations measure the ROI of their AI investments against defined benchmarks (Basis, October 2026; 184 respondents at agencies, brands and publishers).
- In Salesforce researchers' May 2025 benchmark of CRM tasks, the best AI agents completed about 58% of single-turn tasks and about 35% of multi-turn ones; newer models may do better (Salesforce AI Research, 2025).
- At list prices checked 4 November 2026, a run that reads 40,000 tokens and writes 3,000 costs $0.11 on Claude Sonnet 5.5 (Anthropic pricing).
Frequently asked questions
What does "agentic AI" mean for a marketing agency?
Agentic AI is software that decides its own next step toward a goal and uses tools to take it. In advertising news it often means AI that buys media. Inside an agency it means agents doing delivery work: diagnosing a CPA spike, drafting a renewal pack, turning a call into tasks. Most of it runs best as a workflow with one AI step.
How do I keep one client's data away from another client's agent?
Run each agent for one client at a time, with only that client's accounts, documents and records. Never let it fall back to another client's files, and file lessons under the client they came from. Write the rule into a permission policy, test it by asking the agent about another client, and check the log for the accounts each call touched.
Should an AI agent use my admin login?
No. An admin login lets an agent reach every client and change anything, so one bad tool call does the widest damage. Give each agent its own connection at the lowest level that works, usually reporting access, and keep spend changes with a person. If your tool runs agents as a team member, keep that member's access narrow, because the agent inherits it.
How much does an AI agent cost to run per month?
At list prices checked 4 November 2026, this page's estimated run sizes cost from under two cents to just over a dollar each on Claude Sonnet 5.5. An example agency with 18 retainers would spend $1.44 to $5.40 a month on monthly report drafts. Costs grow with runs and with agent loops, which re-read their context on every call, so cap each run and each AI key.
How should agencies explain AI data and security to clients?
Tell clients four things in writing: which tasks AI drafts; that a named person approves anything sent to them or spending their money; that their data is used only for their account, never to help another client; and how long run logs are kept. Put it in the contract, and be ready to show their account's log when they ask who accessed what.
What goes wrong when agencies deploy AI agents?
The usual failures are the wrong client's context, stale data, invented numbers (hallucinations), double sends, runaway loops and a login that quietly lost access. The checks: one client per run, a freshness rule, code-made figures, never re-running a step that sent, limits, and an alert when a connection fails. Gartner predicts over 40% of agentic AI projects will be canceled by the end of 2027 (Gartner, June 2025).