What content businesses should actually build when someone asks for the AI plan for revenue operations — and the order to build it in.
For publishers and media owners across web, mobile, CTV, broadcast, audio and video. Written to be forwarded internally.
Build the agent that accumulates proprietary state before the one that produces a document. Start in the lane where the record almost exists and the heat is lowest — for most publishers, delivery and ad operations or the deal desk — build the joined record for that lane, then put two agents on it: one that saves obvious time and one that compounds. Proposal drafting, the agent everyone asks for first, is the easiest thing on this page for a competitor to copy.
Ask a publisher's revenue leadership what agents they want and you get a version of the same five:
This is a good instinct and the wrong shape. It is good because every one of those five is real work that real people do badly at 6 p.m. on a Thursday. It is the wrong shape because all five sit inside one workflow — the direct-sales revenue cycle — and because four of the five produce a document rather than an asset.
An agent that produces a document gives you a cost saving your competitor will get in the same quarter. An agent that accumulates proprietary state gives you something they cannot copy.
Build the second kind first, even when the first kind is easier to demo.
What follows is the full surface area — eight lanes and thirty named agents — plus the two sorting questions, the format differences, the reason most of these fail on the first attempt, and a sequenced roadmap you can argue with. The agency-side catalog is The Agent Index — 120, and the publisher agents that face a buyer's agent across the table are in the Specialist Stack. This page is written for the sell side, about the whole business.
Before any build decision, run each candidate through two filters. They are independent, and most organizations only apply the second one.
| Commodity agent | Compounding agent | |
|---|---|---|
| Output | A document, a summary, a draft | A structured asset that persists |
| Second run | Same quality as the first | Better than the first |
| Copyable by a competitor | In weeks, with the same models | Only by re-running your history |
| Examples | Draft the RFP response, summarize the campaign, check creative specs | Advertiser relationship graph, campaign outcome library, rights ledger, buyer-preference model |
| What it buys you | Cost and cycle time | Cost, cycle time, and a widening information gap |
↔ scroll table
Both kinds are worth building. The error is building only the first kind and calling it a strategy. A proposal-drafting agent is genuinely useful and will pay for itself; it is also the single most replicable thing in this document. Every competitor of yours will have one within two quarters, because it requires nothing you own.
The compounding agents are slower to stand up, because they require a decision about where a fact lives and who is accountable for it. That friction is exactly why they are defensible.
Scott Brinker's Martec's Law holds that technology changes exponentially while organizations change logarithmically. Read from the other side, the same slowness that makes change hard also makes an advantage hard to copy: reproducing what you do would require a competitor to rewire their own priorities, incentives and commitments.
That is true, and publishers should read it with one caution attached. The protection works when pressure is symmetrical. In advertising it is not. The buy side — holding companies, planning teams, their own agent stacks — is changing considerably faster than the sell side. Organizational slowness only functions as a moat if someone is trying to storm the castle. If the buyer's agent simply never surfaces you because your product catalog is not machine-readable, slowness is not protection. It is absence.
The second filter is autonomy. Not “is this agent smart enough” but “what happens when it is wrong, and how fast can we undo it?” auxfirst scores that with the Action Heat Ladder: five dimensions — reversibility, blast radius, exposure, commitment, authority — and one rule. An action is as hot as its hottest dimension. No averaging.
| Heat | Characteristics | Publisher examples | Appropriate autonomy |
|---|---|---|---|
| LOW | Internal, reversible in minutes, no outside party sees it | Avails lookup, spec validation, log reconciliation, competitive research, meeting prep | Auto-run. Act freely, log the action |
| LOW-MED | Writes to an internal record, fully reversible, narrow scope | Segment building, forecast scenarios, relationship-graph updates, filing a finished flight | Sampled review. Act, and a named reviewer audits a slice |
| MEDIUM | Touches a live campaign or a shared system, reversible with effort | Package construction, pacing reallocation within a campaign, trafficking a campaign | Propose + approve. Nothing happens until a human approves |
| HIGH | Leaves the building, reversible only through a conversation | Proposal sent to a buyer, price quoted, make-good offered, discrepancy position stated, invoice issued | Named approver. Draft only; a named human sends |
| CRITICAL | Contractual, regulatory or public; one mistake is an incident | Signed IO terms, political or regulated creative approved, rights cleared for air, a filing to a regulator | Human executes. The agent assembles the evidence |
↔ scroll table
Two placements surprise people. An invoice is HIGH, not CRITICAL: it leaves the building and asks for money, but it can be reissued. Signed IO terms are CRITICAL, because a signature cannot be taken back.
First, most of the value on this map is in the cooler bands. 21 of the 30 agents below read LOW to MEDIUM. The demoable agents read HIGH. Organizations consistently build the hot ones first because they are visible to executives, then discover that the review burden eats the saving.
Second, the binding constraint on an agent program is not model capability. It is verification throughput — how many agent outputs a qualified human can meaningfully check per day. If your agent produces forty proposals and one person can genuinely review six, you have built a bottleneck with better branding. Design the verification step at the same time as the agent, or the program stalls at pilot.
6 of 40 checked
The other 34 either go out unchecked or wait in a queue that grows every day. Neither is a saving. Measure verification capacity per lane before you set an agent's volume.
Eight lanes cover the commercial life of a publisher: five that run the revenue cycle from brief to wrap, one that sits underneath all of them, one that faces outward to the buyers' agents, and one that bridges editorial and programming into product.
How to read the agents below. Each carries the band of the hottest action it takes in its default design, and the posture that band earns. Where an action is born hot, the reading names what keeps the agent itself cool. Six agents carry a compounding mark: they accumulate a record that makes the next run better. The same method, applied to a different industry, is the Retail Agent Gallery.
The pitch agent everyone wants is reactive. It waits for a brief. The higher-value siblings do not.
Publishers respond to nearly every brief and win a minority of them. An agent that scores inbound briefs on fit, expected margin, displacement cost and historical win rate with that specific agency team lets you invest properly in three instead of thinly in nine. Worked example: a news publisher receives a Q4 brief for a category it has never converted in, requiring inventory that would displace a committed annual advertiser. The triage agent surfaces both facts within an hour of the brief landing, with the evidence attached. That is a decision a sales director can make in five minutes and would otherwise make in two weeks by accident.
Auto-run. It scores and surfaces the evidence; the bid decision stays with the sales director.
Watches signals — product launches, agency roster moves, category commentary in earnings calls, competitor campaign activity, seasonal patterns from your own history — and drafts the pitch nobody asked for. This is how a publisher escapes the queue. It is also, notably, a compounding agent: every proposal outcome teaches it which signals actually precede spend.
Named approver. The pitch leaves the building, so a named seller sends it. Draft-only output (D1) keeps the agent itself cool.
Who moved from which holdco to which, who owns which brand now, who was in the room for the last three wins, which planner has never bought your CTV inventory. No one maintains this and everyone needs it. It is unglamorous, low-heat, and one of the two or three most valuable assets on this page.
Sampled review. It writes to an internal record; a human audits a slice of the updates.
For a given advertiser: where is the money going that is not coming to you, what does the gap look like, what would you have to prove to close it.
Auto-run. Analysis for the account team; nothing leaves the building.
For publishers with a content studio: feasibility, production cost bands, talent availability, editorial approval risk — before sales promises something the newsroom will refuse to make.
Auto-run. A feasibility read before anyone promises anything; the newsroom keeps its veto.
Not “do we have inventory” but “if we take this, what do we turn away, and is the trade positive?” Most avails tooling answers the first question, which is why sales teams sell the wrong inventory cheaply.
Sampled review. Forecast scenarios the sales team will lean on; a planner audits them on cadence.
Category exclusivity, competitive separation rules, adjacency restrictions, sponsor conflicts. Critical in broadcast and audio, handled everywhere by someone's memory and a spreadsheet.
Auto-run. It checks and flags; resolving a conflict stays human.
A continuous inventory of what could be sold and is not: newsletters, app notification slots, CTV pause states, mid-rolls in the back catalog, transcript and archive pages, live event overlays, second-screen moments. Run monthly, this produces a pipeline of new products rather than a one-off audit.
Auto-run. A standing inventory that grows every month; nothing is sold until a human packages it.
ads.txt and sellers.json hygiene, spoofing exposure, made-for-advertising adjacency inside your own reseller chain, fee transparency by exchange. Low heat, high embarrassment avoidance.
Auto-run. It audits and flags; editing the public files stays with a named owner.
Rate card exceptions, floor approvals, margin guardrails, package construction against policy. This work is rule-bound, high-frequency, and currently sits in one senior person's inbox as a queue. It is the clearest early win in the entire map.
Propose + approve. It clears the queue against written policy; the senior approver signs off in batches instead of one request at a time.
The real yield question, asked continuously per segment and daypart rather than quarterly in a meeting.
Propose + approve. Allocation moves touch live inventory; a yield manager approves the shift.
Turns unsold remnant into named, sellable products with a rationale attached — which is mostly a naming and storytelling problem, and therefore well suited to an agent with access to performance history.
Sampled review. It drafts products into the internal catalog; sales decides what goes to market.
For hybrid publishers: per cohort, is this reader worth more behind the wall or in front of it, and what does the answer do to inventory forecasts?
Auto-run. A per-cohort analysis; changing the paywall is a separate, human decision.
This lane is missing from almost every publisher's agent wish list, and it is where the toil actually lives.
Creative specs, tag QA, macro validation, VAST wrapper checks, third-party tracker verification. Enormous volume, high error rate, entirely rule-bound.
Propose + approve. Validation runs on its own; a trafficker approves before anything goes live.
Not a dashboard that turns red on day 26 of a 30-day flight, but an agent that flags the risk on day 4 with three reallocation options and the revenue consequence of each.
Propose + approve. It flags early and lays out the options; a human picks the reallocation.
File weight, safe zones, audio loudness standards such as EBU R 128, caption presence, accessibility requirements, format-specific technical rules. This is the highest-volume, lowest-judgment work in the building.
Auto-run. Checks against the spec; a rejection goes back with the reason attached.
Publisher counts versus agency counts versus verification vendor. The three-way argument currently resolved by someone exporting CSVs at month end and losing.
Named approver. The reconciliation itself is cold; the position you state to the client is not.
What is owed, what can be offered, what it costs us to offer it, what the client accepted last time.
Named approver. It assembles what is owed and what it costs; a named person makes the offer.
The post-buy agent is right, but its output should feed Lane 1 rather than an archive folder.
The report and the recommendation in the same artifact. A wrap report that ends without a proposal is a missed sale.
Named approver. The report leaves the building; a named account lead sends it.
Turns a win into a sales asset — sanitized, approved, tagged by category, format and objective — which becomes ammunition for the proposal agent. This loop, wrap → case study → pitch, is the single most compounding structure in the map, because it makes every campaign make the next pitch better.
Sampled review. It writes into an internal library; using a case study outside the building needs the client's approval.
Structures every completed flight into a reusable record: what ran, where, at what price, against which objective, and how it performed. Slow-burning and unglamorous — and the input to everything in Lanes 1 and 5 later.
Sampled review. An internal record written after the fact; a human audits a sample.
Reconciling different measurement currencies and your own first-party numbers into one defensible story. Live, genuinely hard, and increasingly a purchase criterion in video and CTV.
Auto-run. It reconciles the numbers; the story told to the client travels in the wrap, under its approver.
Which advertisers are pacing below plan, which are quietly rotating budget out, which have stopped asking questions.
Auto-run. Internal alerts to the account owner.
Format-heavy, under-served, and the lane where money leaks silently.
Music licensing, talent usage rights, territory windows, archive re-use permissions. Video, audio and broadcast publishers lose real revenue to “we could not confirm in time that we were allowed.”
Human executes. The agent assembles the evidence from the rights record; a person clears the use.
Political, gambling, alcohol, pharmaceutical, financial promotions — rules by market, by format, by daypart.
Human executes. It checks the rules for the market, format and daypart; a person approves the creative.
Political-ad labels and transparency notices — in the EU, required since October 2025 under Regulation (EU) 2024/900, which also provides for a European repository of online political ads — and the marking of AI-generated creative under Article 50 of the EU AI Act. Any publisher accepting synthetic creative needs a position on this, and most do not have one.
Human executes. What goes to a regulator or a public repository is prepared by the agent and submitted by a person.
The most defensible thing on this list and the least built.
A machine-readable product catalog, structured avails, an agent-legible description of what your formats do and who they reach, documented access, and a stated policy on what a buy-side agent may and may not do without a human. This is an AI-readiness audit pointed at your ad inventory instead of your content marketing. The plumbing already has names — OpenDirect for guaranteed inventory, AdCP for agent-to-agent media buying.
Named approver, until a written policy says otherwise. A provisional answer to a buy-side agent is a quote; decide in writing which answers may go out unattended.
Per-crawler access rules under the Robots Exclusion Protocol, separation of search crawling from training crawling — Google-Extended and OpenAI's separate crawlers make that split possible — unauthorized-use detection, and a tracked position across licensing conversations. Publishers are negotiating these blind and inconsistently across their own properties.
Named approver. Detection runs on its own; every change to crawler rules is signed off by a named owner.
A topic cluster is surging; turn it into a sellable segment while it is still hot, with the editorial calendar attached and the newsroom's boundaries respected.
Sampled review. Segments stay internal until sold; the newsroom's boundaries are rules, not suggestions.
Match advertisers against the upcoming slate — for broadcast, live audio and events, where the calendar is the product.
Auto-run. Recommendations against the slate; sales makes the call.
Buyer experience belongs in this lane too — not as one more agent, but as the promise the other thirty make affordable. It is argued for in section 07.
Put the two filters on one grid and the argument draws itself. The usual wish list — dashed — sits in the commodity row, mostly in HIGH. The agents that are both cool and compounding sit top left, and almost nobody arrives asking for them.
↔ scroll the grid
Two cells deserve a name. Top left — cool and compounding — is where a program starts: the relationship graph, the unmonetized-surface inventory, the case study library and the campaign outcome library. The HIGH column of the commodity row is where most wish lists start: review-heavy, easy to demo, and the first thing a competitor copies.
The roadmap below is this grid read from the left: pair one cool commodity agent with one cool compounding agent, then move right as the records earn it.
The same lane behaves differently depending on what you publish. This table is the short version of a conversation worth having internally.
| Lane | Web / mobile | CTV / video | Broadcast TV | Audio / podcast |
|---|---|---|---|---|
| Forecasting | High volume, elastic, statistical | Hybrid: fixed slates plus programmatic | Fixed, scarce, conflict-heavy | Fixed slots; back-catalog inventory is a live asset |
| Conflict rules | Light | Moderate | Heavy — separation and exclusivity | Heavy — host-read and sponsor conflicts |
| Creative compliance | Weight, tags, viewability | Transcoding, captions, safe zones | Broadcast technical standards, clearance | Loudness, host-read scripting, disclosure |
| Measurement pain | Attribution and identity | Currency fragmentation | Panel versus big-data reconciliation | Downloads versus listens, attribution gaps |
| Rights exposure | Low to moderate | High — music, talent, windows | High | High — music licensing especially |
| Biggest agent win | Trafficking and yield | Currency translation and avails | Conflict, clearance and packaging | Back-catalog monetization and host-read ops |
↔ scroll table
Publishers with more than one of these — most large media owners — should expect that the agent that works brilliantly for the web business needs different rules, not just different data, for the broadcast business.
Almost every agent above fails for the same reason, and it is not the model.
The facts are not joined. Availability lives in the ad server. Performance lives in the reporting stack. Advertisers and contacts live in the CRM, partially. Rights live in a spreadsheet owned by one person in legal. Creative specs live in a PDF from 2023. Editorial plans live in a calendar tool the commercial team cannot see. Pricing policy lives in a deck and in the sales director's head. Historical outcomes live in an email thread.
| Agent | Systems it must join | Typical blocker |
|---|---|---|
| Pitch / RFP | Ad server, CRM, reporting, case study library, rate card | No structured record of what was actually sold, at what price, and how it performed |
| Displacement forecast | Ad server, order management, CRM pipeline | Pipeline probability is not maintained honestly |
| Conflict and exclusivity | Order management, contracts, programming schedule | Exclusivity terms live in PDFs, not fields |
| Rights and clearance | Media asset management, contracts, licensing records | No canonical rights record exists at all |
| Buyer readiness | Product catalog, pricing, ad server, site | No product catalog exists in machine-readable form |
↔ scroll table
The joined context layer for one lane, plus two agents on top of it that prove the join was worth building.
Sold that way, the first engagement is defensible and the second is obvious. Sold as five agents, the program delivers five demos and one disappointed CFO.
A useful internal test. Pick any agent on the map and ask three questions. Where does each fact it needs live today? Who is accountable for that fact being correct? What happens when it is wrong?
If you cannot answer all three for every input, you are not ready to build that agent — you are ready to build the record it depends on. Run formally against one workflow, that test is an AI agent readiness assessment.
Here is the part worth arguing about in the leadership meeting.
Publishers talk about audience experience constantly and buyer experience almost never. Yet the buyer experience most media owners deliver is poor: multi-week turnaround on briefs, proposals assembled by hand from stale decks, wrap reports that arrive after the next planning cycle has started, monthly discrepancy arguments, and proactive service reserved for the top twenty accounts because that is all the team can carry.
Everyone knows this. Nobody fixes it, because fixing it means giving hundreds of accounts the treatment currently reserved for twenty, and no one can hire for that.
Agents change the economics of exactly that problem. Not by replacing sellers, but by making premium responsiveness affordable across the whole account base. These are not AI features. They are operational promises that were previously unaffordable:
| The promise to the buyer | What makes it affordable |
|---|---|
| A 48-hour proposal turnaround on every brief | Bid/no-bid triage decides which briefs deserve it; the campaign outcome library supplies the evidence |
| A wrap report the week the campaign ends, with the next proposal attached | Wrap plus next-buy, feeding the case study generator |
| A reconciliation position before the client's finance team asks | Discrepancy reconciliation |
| Proactive service beyond the top twenty accounts | Renewal risk and relationship graph maintenance |
| Rights questions answered before they block a deal | Rights and windowing, on a canonical rights record |
↔ scroll table
And they are hard to copy, for the reason Martec's Law points at: delivering them consistently requires an organization to change how it works, not to buy a model. A competitor can access the same tools. Committing to a service standard across an entire book of business, and rebuilding the record-keeping that makes it possible, is a multi-year organizational choice.
So the framing for the board is not “AI will make our sales team faster”. It is this:
AI finally makes it affordable to keep the service promise we have been making in pitches for a decade.
That is a moat. The proposal-drafting agent, on its own, is not.
Tiered as Nowstart in the next quarter, Nextthe following two, and Testrun as a genuine experiment — do not promise outcomes.
| Action | Output | Owner |
|---|---|---|
| Map the eight lanes against your actual workflows; delete what does not apply | One-page lane map with owners named | Revenue ops |
| Run the two filters on every candidate | Scored shortlist: compounding or commodity, heat band | Revenue ops + one commercial lead |
| Audit where the facts live | The systems-and-blockers table above, filled in for your top five candidates | Data / BI |
| Grade autonomy | A publisher-specific heat ladder — every workflow, its band, its named accountable human | Revenue ops + legal |
| Set the verification budget | How many agent outputs per day a qualified human can genuinely check, per lane | Line managers |
↔ scroll table
The deliverable at day 30 is a decision document, not a pilot. Organizations that skip this stage build the proposal agent, because it is the one everybody can picture.
Exit criteria for day 90: one lane's record is canonical and maintained, two agents are in daily use by named people, verification load is measured rather than assumed, and you can state in one sentence what the compounding agent knows now that it did not know in month one.
Score each question 0–2. Below 12 out of 20, the honest first project is a record, not an agent.
For a read on how far your organization already trusts the agents it runs — authority, oversight, traceability, failure design and ownership — the Agent Trust Index scores it in eight minutes, with no email.
Proposal generation is hot, commodity and review-heavy. It looks best in a board meeting and delivers the least durable advantage.
It produces eight context layers, none maintained, and a maintenance burden that outlives the enthusiasm.
Frame it that way and you lose the people whose tacit knowledge the agents need as input. The correct frame is coverage: the same team, an order of magnitude more accounts served properly.
Every HIGH and CRITICAL output needs a named human, not a queue. Accountability that belongs to “the team” belongs to nobody. Writing that human down, agent by agent, is the AI agent operating model.
A chat box on top of an unjoined data estate is a faster route to a wrong answer.
It protects you from imitators. It does not protect you from a buy-side agent that cannot find you.
A publisher that has done this properly does not have twenty agents. It has:
The observable differences are mundane and hard to copy. Briefs are answered in days. Wrap reports arrive with the next proposal attached. Rights questions are answered before they become deal blockers. The long tail of accounts gets the service that used to be reserved for the top twenty. And the proposal that goes out on Friday is materially better than the one that went out in January, because everything that happened in between was captured.
Getting better rather than staying fast is the only property your competitors cannot prompt into existence.
Six to eight agents on one record is an estate, and an estate needs versions, owners, evaluation and a way to retire what stops earning its keep. That operating discipline is the AI agent development lifecycle.
Start in the lane where the record almost exists and the heat is lowest — for most publishers, delivery and ad operations, or the deal desk. Build the joined record for that lane, then put two agents on it: one that saves obvious time, such as a creative compliance agent, and one that starts accumulating state, such as a campaign outcome library. Proposal drafting, the agent most teams ask for first, is the easiest for a competitor to copy.
A commodity agent produces a document, and its second run is no better than its first, so a competitor with the same models can copy it in weeks. A compounding agent accumulates a structured asset that persists, such as an advertiser relationship graph or a campaign outcome library, so each run improves the next and a competitor could only copy it by re-running your history.
Decide action by action, not agent by agent. Score each action on the Action Heat Ladder — reversibility, blast radius, exposure, commitment and authority — and let the hottest dimension set the band. An avails lookup can run on its own; a proposal sent to a buyer or a make-good offer needs a named approver; signed IO terms and regulated creative stay with a human, with the agent assembling the evidence.
The number of agent outputs a qualified person can genuinely check in a day. It is the binding constraint on an agent program: an agent that produces forty proposals a day for a reviewer who can check six has built a bottleneck. Design the verification step at the same time as the agent.
Because the facts are not joined. Availability lives in the ad server, performance in the reporting stack, advertisers in the CRM, rights in a spreadsheet and pricing policy in a deck. The honest first project is the joined context layer for one lane, plus two agents on top that prove the join was worth building.
Whether a buy-side AI agent can discover what a publisher sells, understand its pricing logic and get a provisional answer without a human on the publisher's side. It takes a machine-readable product catalog, structured avails, an agent-legible description of formats and audiences, documented access, and a written policy on what a buy-side agent may do unattended.
The map sequences it over eighteen months: thirty days to decide, days 30 to 90 to build one joined record and two agents, months four to nine to extend along the compounding chain, and months nine to eighteen for the machine-facing layer. The end state is one joined commercial record with six to eight agents on it, not twenty agents.
Standards and regulation cited on this page
The frameworks underneath
From the ad industry hub
Method · No market statistics appear on this page. Counts are the agents named here. Heat readings are auxfirst's default reading of each agent's hottest action, before any design move that cools it — your systems can move any of them. The forty-and-six example is an illustration, not a measurement. Standards and regulations link to their primary sources. The map, the two filters and the roadmap are auxfirst's own frameworks, developed through client work rather than derived from a published standard.
The ten questions are the starting point. The full version — your lanes, your records, your heat ladder, your first join — is a working session with auxfirst.