←
Perspectives · Marketing Technology

Who's watching your AI agents?: AI agents in marketing, and the access you give them

TCS, Cognizant, HCLTech and Tech Mahindra have spent October telling investors that their next line of business is keeping AI agents in check. Their clients are large enterprises, but the question applies to any business that has let AI agents in marketing touch its ad accounts, CRM or inbox. This piece covers what those agents can change, who is checking, and five controls a smaller business can set this week.

Published
Reading time
6 min
Topic
Marketing Technology
By
Kaushik Bhagwat
01The news

Why India's IT firms are selling agent oversight

Companies that rushed AI agents into their operations now want to know what those agents are doing, and the large IT services firms have products ready.

On its earnings call on 8 October, TCS described an "agentic control plane" as a key layer for its clients: one place to govern agents, track their cost and performance, and manage them from launch to retirement. Tech Mahindra used its 15 October call to talk about traceability and red teaming. HCLTech announced on 28 September that it plans to buy Robotiq.ai, a Croatian firm that coordinates agents across business software.

Cognizant went furthest. Its AI Nexus, unveiled on 23 October, gives each agent its own identity, permissions and memory, logs every task, and asks a person to sign off on important decisions. Elets CIO has the full round-up, drawing on a report in Mint.

Phil Fersht of HFS Research named the problem these products address. Agents tend to inherit the access rights of the person or app that set them up, and most of the controls on them are still manual.

02Your stack

Where AI agents in marketing already have access

A small business does not need an enterprise rollout to have agents acting on its accounts.

51.1%
Marketing organisations using or building AI agents, August 2026 (20% a year earlier)
6.5%
Marketers who measure AI's return against set benchmarks

The figures come from a survey of 184 marketing and advertising professionals by the ad software firm Basis, reported by PPC Land. The same survey found that 65% of respondents do not measure AI's return, plan to start, or do not know.

Ad accounts are the clearest case. Since April, Meta has offered AI connectors, in open beta, that let Claude and ChatGPT read an ad account and act on it: create campaigns, ad sets and ads, change budgets and targeting, and pause or switch on ads. Jon Loomer's review of the connector notes that it takes on all the access of the Facebook user who connects it, which can include accounts that clients or partners have shared with that user.

Then there are the agents most owners never set up. CRM tools draft and send follow-up emails, website chat tools answer buyers in the brand's name, and the ad platforms themselves now make changes by default.

03Defaults

The changes ad platforms make without asking

Three changes this month show the platforms acting on advertisers' accounts unless the advertiser opts out.

From 12 October, Google Ads switched on automated promotions for some advertisers' Search and Performance Max campaigns. Google reads coupons and deals from the advertiser's own website and shows them with the ads. Each promotion runs for under 24 hours and renews while the offer is still on the site, so a Diwali banner left up after the sale keeps feeding the ads. The opt-out sits in account-level automated assets.

On 28 October, Meta turned on product browsing in existing Facebook and Instagram ads. A person who taps an ad may see product details and suggestions inside Meta's apps before reaching the advertiser's website. Advertisers who missed the opt-out can still switch it off for each ad in the Advantage+ Creative and Destination settings.

On 30 October, Meta's revised Commercial Terms take effect. A new section on AI features makes the advertiser "solely responsible" for changes Meta's AI makes to its ads, including features Meta switches on itself. Nobody signs the new terms: running ads after the notice counts as acceptance. PPC Land reports that agencies have found Meta's AI adding sitelinks, translating copy and generating images in ads where nobody had enabled those features.

Each of these is a change inside your account, made by the platform, and each needs someone on your side to check the setting.

04The risk

What an AI agent can change in your accounts

An agent can do anything the login it uses can do, and it does it faster than a person.

Take a founder who connects an AI assistant to Meta with their own admin login and asks it to "scale what's working." The assistant can raise budgets on three campaigns, duplicate the best ad into new audiences and pause the rest. Each step may be sensible. Together they can double the week's spend before anyone opens Ads Manager.

The same applies to a CRM agent that sends a discount to the wrong segment, or a website chat tool that quotes last year's price. In each case, the mistake is visible to customers before it is visible to the business.

05Controls

Five controls for AI agents in your ad accounts

The enterprise products sell identity, permissions, logs and sign-off. A small business can set up each of these with tools it already pays for.

  1. Start with read access.Let an agent report on performance for a few weeks before you let it change anything. Most of the value in the early weeks comes from analysis.
  2. Give each agent its own login.Add the agent through a separate user or a business partner role with the lowest permission that does the job. Then you can remove it without touching your own access, and its changes are easy to tell apart.
  3. Make it draft, and you launch.Meta's connector creates campaigns paused, so a person has to switch them on in Ads Manager. Keep it that way, and ask CRM and email tools to queue messages for approval in the same way.
  4. Cap the spend.Set an account spending limit in Meta and a daily budget ceiling in Google Ads that an agent cannot exceed. A cap limits how much a bad instruction can cost.
  5. Read the log every week.Google Ads keeps a change history that shows who changed what and when, and Meta keeps a similar record. A weekly read also shows what the platform defaults above have changed.
06India

What India's data law means when an agent touches customer data

Under the Digital Personal Data Protection Act, the business that decides why customer data is used stays responsible for it, whoever or whatever does the processing.

The Act calls that business the data fiduciary, and it remains responsible for processing done on its behalf by a data processor. An AI tool that reads your CRM or sends emails to your customer list is processing personal data for you. EY's summary of the Act and Rules covers the duties, including contracts with processors and access logs.

The Rules were notified on 13 November 2025, and most duties apply from 13 May 2027. They require reasonable security safeguards, and the penalty for failing to take them goes up to ₹250 crore. They also ask businesses to keep processing logs for at least a year. An agent with its own login and a change history to read is easier to account for than one running on the founder's password.

07Where to start

Where a small business should start

Start with a list of every tool that can act on your accounts.

Write down each AI tool, plug-in and automation that can read or change your ad accounts, CRM, email platform and website. Next to each, note whose login it uses, what it can change, and who checks its work. Most businesses find at least one tool that nobody remembers connecting.

Then switch off what you are not using and apply the five controls to the rest. Setting up and checking the tools connected to a website, CRM and ad accounts is part of what we do. For the other side of AI in marketing, how assistants decide which brands to name, read our guide to generative engine optimisation.

08Questions

Questions people ask

What is an AI agent in marketing?

Software that takes actions in your marketing tools, as well as answering questions. An AI agent might create and pause ads, change budgets, update CRM records or send emails, using the access granted when it was connected. A chatbot that only answers questions is not an agent in this sense.

Is it safe to connect ChatGPT or Claude to my Meta ad account?

It can be, with limits. Meta's connector creates new campaigns paused, so a person has to launch them. Connect it with the lowest permission that does the job, set an account spending limit, and check the change history each week.

How do I turn off Google's automated promotions?

In Google Ads, go to Assets, open account-level automated assets, then Advanced settings, and set automated promotions to off. To see what Google has added, filter assets by Promotion and look for Google AI in the Added by column.

Does the DPDP Act apply to AI tools that use my customer data?

Yes. Under the Act, your business stays responsible for personal data processed on its behalf, including by software and vendors. Most duties under the DPDP Rules apply from 13 May 2027, with penalties of up to ₹250 crore for failing to protect personal data.

Have a brief?

If AI tools are connected to your website, CRM or ad accounts, we can check what they can touch and set the limits.

Start a project →