AI agents can now browse, pull data, write, run checks and hand you a finished draft without step-by-step prompting. Whether they can do a job is rarely in doubt any more. Whether they should, and who checks the output, is. Hand agents the work where mistakes are cheap and easy to spot. Keep the decisions where a mistake damages the brand, breaks a law or sends the business in the wrong direction.
The disclosure rules near the end apply at the time of writing, October 2026.
Assistant, automation or agent?
People use the three words loosely. They describe different amounts of freedom:
- An assistant answers when asked. You prompt, it drafts, you decide.
- An automation follows fixed rules you wrote. If X happens, do Y.
- An agent is given a goal and works out the steps, often using tools such as a browser, a spreadsheet, an analytics account or an ad platform, then returns a result or takes an action.
That freedom is why agents are useful and why they are risky. An assistant that writes a bad headline wastes a minute. An agent with write access to your ad account that “optimises” budgets on a broken conversion event can waste a month of spend before anyone notices.
For each task, then, ask two things before you ask whether AI is any good at it: what can the agent touch, and what happens if it gets it wrong?
Cost of error × ease of checking
Most delegation choices come down to two questions.
- If the output is wrong, how much damage does it do? A wrong number in an internal draft is cheap. A false claim in a live ad is not.
- How quickly can a person check it? Reviewing ten ad variants takes five minutes. Checking whether a market-entry recommendation is right takes weeks.
Where errors are cheap and checking is fast, delegate freely. Where errors are costly but checking is fast, delegate with review. When checking is slow and errors are expensive, keep the work in human hands, because nobody will catch the mistake in time.
Jobs worth handing to agents
Research and desk work
Competitor pricing pages, feature comparisons, review mining, collecting public data for a report, summarising customer interviews you already transcribed. Agents gather and organise quickly. They also invent or recycle outdated facts, so ask for a source link next to every claim and spot-check a sample before anything goes outside the company.
Reporting
An agent can pull weekly numbers from analytics, ad platforms and the CRM into one sheet, write the first-draft commentary and flag movements outside normal ranges. For small teams this is often where most hours come back. Keep agents read-only on the source systems. A person writes the “so what” for leadership.
QA and monitoring
Is every live link carrying the right UTM parameters? Do conversion tags fire on key pages? Does each landing page load and match its ad, and is any campaign spending against a broken goal? These checks are tedious and rule-based, and a person can verify the output in minutes, which is exactly the profile you want. If you are scaling spend, run a tracking audit before scaling and have an agent repeat the checks weekly. A browser extension such as our Pixel & Tag Checker covers the manual side.
Creative variants
Headline, body and call-to-action variants, resized assets, alternative hooks for short video. Generation costs almost nothing now, so review and testing become the bottleneck. Give the agent your brand rules and past winners, ask for variants that differ in one idea at a time, and feed them into a structured creative testing process instead of launching everything at once.
Lifecycle copy drafts
Agents write decent first drafts of onboarding emails, win-back messages and in-app prompts to a brief you supply, and they can turn out a version for every segment. Before anything is scheduled, a person checks tone, claims, offers and the personalisation logic. The welcome email sequence tactic is a good place to start.
Jobs to keep in human hands
Strategy and prioritisation
Which market to enter, which channel to bet on, what to stop doing. An agent can prepare the inputs: market sizing, competitor summaries, channel benchmarks. It does not know your cash position, your team’s capacity or your appetite for risk, and its answers drift towards the average of what is already written online. Strategy is the one place you need to be different from the average. For ranking a roadmap, the ICE and RICE lesson will serve you better than a model’s list.
Positioning and brand voice
Who you are for, what you stand against, how you sound. Drafts can show you options. The final positioning should come from customer research and from someone who will defend it for years.
Final claims
Performance figures, savings, health or financial benefits, “best” or “number one”, testimonials, reviews: anything customers will rely on. Generated text can invent a statistic or overstate a benefit and sound completely sure of itself. In the US, the FTC’s rule on fake reviews, announced in August 2024, explicitly covers reviews that misrepresent who wrote them, “such as AI-generated fake reviews”. Never let an agent create reviews or testimonials.
Spending and irreversible actions
Raising budgets, changing bid strategies, sending to your whole list, deleting audiences or data. Agents recommend. People approve.
Task risk table
Treat the levels as a starting point. In regulated sectors (health, finance, children’s products), move most rows up one level.
| Task | What the agent does | Risk level | Human role |
|---|---|---|---|
| Competitor and market research | Gathers and summarises public information with links | Low | Spot-check sources before external use |
| Weekly performance report | Pulls data, drafts commentary, flags anomalies | Low | Read-only access; human writes conclusions |
| UTM and link QA | Checks every live link against naming rules | Low | Review exceptions |
| Tag and conversion QA | Tests that events fire on key pages | Low | Fix issues it finds |
| Data clean-up (CRM, lists) | De-duplicates, standardises fields | Medium | Approve before merging or deleting |
| Ad creative variants | Drafts copy and image variants to a brief | Medium | Review every variant for claims and brand fit |
| Lifecycle email and push drafts | Writes segment versions to a brief | Medium | Check claims, offers and personalisation logic |
| SEO content drafts | Drafts articles from an outline and sources | Medium to high | Edit heavily; add original insight; fact-check |
| Budget and bid changes | Recommends or applies changes | High | Recommend only; human approves |
| Sending to customers | Schedules or sends messages | High | Human approves every send |
| Public claims, reviews, testimonials | Writes claims customers rely on | High | Human writes or verifies; never generate reviews |
| Strategy, positioning, pricing | Proposes direction | High | Human decides; agent prepares inputs only |
The SEO row is rated higher for a reason. Google’s spam policies give “using generative AI tools or other similar tools to generate many pages without adding value for users” as an example of scaled content abuse. Drafting with AI is fine; Google objects to volume published without value.
Guardrails to set before you scale
Review at the points where it counts
Decide in advance which outputs a person must approve, and make approval a step someone actually takes. A workable rule: anything external, anything that spends money and anything that touches customer data gets a named reviewer.
The brand brief goes in first
Every agent gets the same written brief: voice, banned words, claims you can and cannot make, legal lines, product facts with sources. Most off-brand output comes from agents nobody gave the rules to.
Least-privilege access
Read-only by default. Give agents their own accounts or API keys so their actions are logged and can be revoked. Keep them away from payment details and full customer lists unless the task truly needs them.
Data privacy
Find out what your AI provider does with the data you send (is it retained, is it used for training?) and pick business settings that fit your obligations. Send aggregated or pseudonymised data where you can, and check that the use fits the consent you collected. In India, the Digital Personal Data Protection Rules were notified on 14 November 2025 with an eighteen-month phased compliance period, so review your consent notices and stated purposes now.
A log you can audit
Record what each agent did, with which inputs, and who approved the output. Something will go wrong eventually, and when it does you want the cause in minutes, not days.
Platform and legal rules on AI disclosure
These rules change often. What follows is accurate as of October 2026; check the source before each launch.
Meta requires advertisers to self-disclose when they use AI to create or edit certain social issue, electoral or political ads, and adds information to the ad and to its Ad Library. In a June 2026 update, Meta also said it is adding more information to ads it detects were made with third-party AI tools, shown in an “About this ad” tool, and that it labels organic content it detects as AI-generated with an “AI info” label.
On Google Ads, election ads that use synthetic or digitally altered content that “inauthentically depicts real or realistic-looking people or events” must be disclosed by ticking the “Altered or synthetic content” box in campaign settings. Google’s policy generates the in-ad disclosure for some formats; for others the advertiser must add a clear disclosure as well.
TikTok’s Community Guidelines require creators to label AI-generated or significantly edited content that shows realistic-looking scenes or people. Its ad policies are separate, so check those for paid content.
In the EU, the AI Act’s Article 50 transparency obligations have applied since 2 August 2026. According to ICTRecht’s summary, deployers must disclose deepfakes (realistic AI-generated or manipulated image, audio or video), and advertising can be in scope. AI-generated text published to inform the public on questions of public interest must also be disclosed, unless it has had human editorial review and someone holds editorial responsibility.
India amended its IT Rules on 10 February 2026, in force from 20 February: synthetically generated content must be clearly labelled and carry provenance metadata. According to Hogan Lovells, large social platforms must ask users to declare whether content is synthetic and label it. Post AI-made realistic content in India and you should expect to be asked.
One rule covers all five. If AI made a realistic person, voice, scene or event that did not happen, disclose it. If it only helped you write or edit, check the platform’s policy, though routine edits usually need no disclosure.
Before you give an agent write access
- You have listed the team’s recurring tasks, scored each on cost of error and ease of checking, and started with the two low-risk tasks that eat the most hours.
- A one-page agent brief exists: brand voice, banned words, approved claims, legal lines, links to product facts.
- A read-only reporting or QA agent has run for four weeks and you have measured the hours it saved.
- Every external output and every action that spends money has a named reviewer.
- Your creative sign-off checklist asks “AI-made realistic content?” and you know the disclosure rule on each platform you advertise on.
Questions people ask
What marketing tasks can AI agents do well?
Tasks with clear inputs, checkable outputs and low cost of a mistake: desk research, weekly reporting, tracking and link QA, drafting ad and email variants, and tidying data. They save the most time where a human can review the output quickly.
What should marketers not hand to AI agents?
Strategy, positioning, pricing decisions and anything that makes a final claim to customers, such as health, financial or performance claims. Agents can help prepare inputs for these decisions, but a person who understands the business and carries responsibility should make them.
Do you have to disclose AI-generated ads?
It depends on the platform, the ad type and the country. Google and Meta require disclosure for realistic synthetic content in election and political ads, TikTok requires labels on realistic AI-generated content, and the EU AI Act requires deepfakes to be disclosed from 2 August 2026. Check each platform's current policy before you launch.
Is it safe to give an AI agent access to customer data?
Only with care. Give agents the minimum data they need, prefer aggregated or pseudonymised data, check how your AI provider handles data you send, and make sure your use fits the consent you collected under laws such as GDPR or India's DPDP Act.
Sources
- Meta Newsroom: How Meta is preparing for the 2026 US midterms
- Google Ads Policy: Disclosure requirements for synthetic content (July 2024)
- TikTok Community Guidelines: Integrity and Authenticity
- ICTRecht: AI Act transparency obligations now in effect
- Hogan Lovells: India introduces mandatory labelling for AI
- FTC: Final rule banning fake reviews and testimonials
- Google Search Central: Spam policies
- PIB: DPDP Rules, 2025 notified
