30 Oct 2026 · 9 min read
The People-Heavy Parts of Programmatic: Where Manual Work Still Sits and What Agents Take On
AI agents can take on most of the repetitive coordination work in media buying, from RFPs and insertion orders to deal ID setup, pacing checks, reconciliation, make-good requests and reporting. People keep the judgement: the objective, the approval of commitments that matter, the disputes and the client relationship.
TL;DR. Programmatic automated the auction, not the work around it. This workflow map lists nine people-heavy tasks in a typical programmatic or direct deal, with how each works now, what an agent can take on and what stays with a person. The pattern is consistent: agents take the chasing, copying and checking, while people approve the deals that matter.
Why is so much of programmatic still manual?
In May 2026, AdExchanger reported that the day-to-day work of running campaigns "is still largely manual, from copying data between systems and reconciling reports to running QA checks". Tanja Mimica, COO and co-founder of Kovva, told the publication that "inside of the platforms, it's AI everywhere", yet "buyers are still in spreadsheets for most of the day".
The gap is not new. In 2020, Digiday quoted Alexis Faulkner, who then headed Mindshare's performance division Fast in the UK: "Automation hasn't made the operational process of running campaigns any easier."
New agentic standards address this in-between work. IAB Tech Lab's AAMP (Agentic Advertising Management Protocols) and AgenticAdvertising.org's AdCP (Ad Context Protocol) both define structured steps for proposals and buying; which protocol does what covers how they relate.
How should the work be split between agents and people?
A workable test is whether a task can be checked against a written record of what was agreed. RFP chasing, IO drafting, deal setup and reconciliation pass that test; choosing an objective, accepting a make-good instead of a credit or protecting a seller relationship do not.
Earned autonomy describes how the split moves: an agent starts at Observe, then Recommend, Draft, Human-approved action and finally Bounded automatic action, widening only as its record supports it. As a worked example, Alkimi, an enterprise platform for agentic advertising, offers a Gym where buyers and sellers build their own agents, set the rules they work to and spar against counterparty agents before any real money moves. A team could start an agent on RFP chasing at the Observe stage, before it sends anything.
Which parts of media buying can AI agents automate?
The nine tasks below follow a typical direct or programmatic guaranteed deal from brief to final report.
1. Sending and chasing RFPs
How it works now. The request for proposals is still largely a human conversation. An AdExchanger Sell Sider column noted in 2018 that "buyers and sellers still meet and discuss all of the requirements of an RFP".
What an agent can take on. Turning a brief into a structured request, sending it to shortlisted sellers and chasing late replies. IAB Tech Lab's draft OpenProposal specification, part of AAMP 3.0, defines a custom proposal "built in response to a buyer's agent brief", and AdCP's task reference includes request_proposals to "request seller-authored plans".
What stays with a person. The brief itself: the objective, the audience, the budget and which sellers are worth asking.
2. Comparing proposals
How it works now. Sellers describe their products and packages in their own ways, so someone has to line responses up before comparing them. IAB Tech Lab's AAMP hub describes OpenProposal as standardising how publishers "define advertising products and packages and to represent them in RFP responses".
What an agent can take on. Reading every response against the brief and the mandate, and flagging gaps. The OpenProposal draft's summary view is "what a buying agent reads when screening many sellers' shelves", and its README states that of 71 fields, only 3 require a round trip with the counterparty.
What stays with a person. Choosing between close offers and deciding which exceptions are worth pursuing.
3. Insertion orders and contract terms
How it works now. The same 2018 AdExchanger column described direct sales as historically "a mostly manual process that involved scheduling deals, running custom reports, handling insertion orders and negotiating price points". An IO can incorporate the 4A's/IAB Standard Terms and Conditions (Version 3.0), which cover delivery, measurement and makegoods.
What an agent can take on. Drafting the order from the agreed terms and checking it against the standard terms. IAB Tech Lab's reference buyer agent books deals through OpenDirect 2.1, and AdCP's accept_proposal task covers "new-buy, amendment, or cancellation terms".
What stays with a person. Signing, any non-standard clause and any commitment above the approval threshold.
4. Deal ID setup and activation
How it works now. Once a deal is agreed, its terms have to reach both the selling system and the buyer's platform. IAB Tech Lab's Deal Sync API was written to "decrease manual entry of deal information across systems".
What an agent can take on. Pushing deal terms and confirming the deal ID reached the buying platform. IAB Tech Lab's reference seller agent can "push deals to buyers via IAB Deals API v1.0", and the reference buyer agent obtains deal IDs for activation.
What stays with a person. Deciding what to do when a deal does not spend; how deal IDs work in agentic advertising covers the detail.
5. Launch checks and creative QA
How it works now. After launch, traders usually run checklists by hand, according to the May 2026 AdExchanger report: is the campaign running in the right geographies, and are the pixels placed correctly?
What an agent can take on. Running the checklist a few hours after launch and reporting exceptions. AdCP includes sync_creatives to "upload and manage creative assets" and get_media_buys to retrieve "media buy status, creative approvals, and delivery snapshots".
What stays with a person. Creative approval and judgement about brand context.
6. Pacing checks and budget moves
How it works now. "Is it pacing properly?" sits on the same hand-run checklist in the AdExchanger report.
What an agent can take on. Watching delivery against the plan, flagging lines at risk and proposing where to move budget. AdCP's control_media_buy task applies "operational delivery controls" to a live buy.
What stays with a person. Any move outside agreed caps, pausing a campaign and trade-offs between a client's priorities.
7. Discrepancy investigation and reconciliation
How it works now. Under the 4A's/IAB terms, a measurement gap of more than 10% over the invoice period, with the controlling figure lower, triggers "a reconciliation effort". A 2015 AdExchanger column put it plainly: "While programmatic buying happens in real time, reconciliation and billing are hardly programmatic."
What an agent can take on. The first pass: comparing both sides' records line by line and explaining the gap. AdCP includes sync_reporting_receipts to record "revision and adjustment reconciliation". One published simulation, Agree. Transact. Verify., found mismatched records in 95.3% of simulated deals when each side kept its own record, against 0.19% with a shared record.
What stays with a person. Disputes, write-offs and final sign-off on the invoice.
8. Make-goods
How it works now. The 4A's/IAB terms set out makegoods as a remedy, for editorial adjacency breaches for example, with a credit if none can be agreed. In TV, Digiday described make-goods in 2020 as "inventory owed to advertisers in order to make up for the viewership shortfalls".
What an agent can take on. Spotting under-delivery early against the agreed terms, assembling the evidence and drafting the make-good request with replacement options.
What stays with a person. Accepting a make-good or a credit, and the seller relationship; make-goods in agent-negotiated deals goes further.
9. Reporting and client updates
How it works now. Half of the campaign reports assembled by programmatic traders between November 2018 and January 2019 were created manually, according to 60% of the 93 respondents surveyed by CtrlShift, as Digiday reported in 2020.
What an agent can take on. Pulling delivery data and drafting the report. AdCP's get_media_buy_delivery task retrieves "performance and delivery data", and the May 2026 AdExchanger report describes agents that draft replies to client emails, "with a human required to approve it before it's sent".
What stays with a person. Interpreting the results and the conversation with the client.
How much should an agent be trusted with each task?
Trust should follow evidence, task by task. IAB Tech Lab's reference seller agent lists "human-in-the-loop approval gates", and AgenticAdvertising.org calls its governance principle Embedded Human Judgment, "the principle that humans stay in the loop on decisions with real consequences". Launch checks can move towards Bounded automatic action sooner because errors are cheap and visible, while IOs and make-goods usually stay at Draft or Human-approved action longer; how AI agents can help with media buying has a stage-by-stage table.
Does handing these tasks to agents replace media buyers?
The evidence points to a changed job, not a removed one. "It's important to acknowledge that a lot of roles in programmatic have changed, but if anything there are more of them now, not less," Amir Malik, then a digital marketing expert at Accenture Interactive, told Digiday in 2020. Handing these chores to agents leaves people with the brief, the relationships and the approvals.
How can a team test an agent before it takes on real work?
The safest route is to measure an agent on realistic briefs before it touches live budgets, then widen its authority one task at a time. Alkimi is neutral on standards: it can work with agents built on AdCP or AAMP, and with or without a DSP. Concourse benchmarks any buying agent without asking which standard it uses, the Gym trains and spars any agent regardless of protocol, and the neutral Marketplace keeps human approval at the deal level, with a DealSheet both parties hold as the same timestamped record of what was agreed.
Frequently asked questions
Which parts of media buying can AI agents automate?
Agents can take on repetitive coordination work: RFPs, proposal comparison, IO drafting, deal ID setup, launch and pacing checks, reconciliation, make-good requests and reporting. Each can be checked against a written record of what was agreed.
Which media buying tasks should stay with people?
The objective, approval of commitments above a threshold, disputes and client relationships should stay with people. These involve trade-offs and accountability that a record cannot settle alone.
Do agents need a particular standard to take on this work?
No single standard is required; both AAMP and AdCP define structured steps for proposals and buying. What matters more is a clear record against which the agent's work can be checked.
Where should a team start?
Start with a frequent, low-risk chore such as launch checks or report drafting, at the Observe or Recommend stage. Widen the agent's authority only when its record supports it.