Will AI agents fix your marketing workflow? Read this first

Agents automate a defined process. If your marketing team cannot write its process down, an agent will run the confusion faster.

Historical photograph of women operators working at a Bell System telephone switchboard
Photo: The U.S. National Archives, No restrictions, via Wikimedia Commons

AI agents will not fix a marketing workflow that was never defined; they will run it faster, including its mistakes. They work today on jobs with clear inputs, a checkable output and a human review step: research and competitor tracking, reporting, lead enrichment and triage, first drafts and creative variants. They fail on long, multi-step work that depends on private context and judgement, where even the best agent in a 2025 benchmark finished only about 30% of office tasks on its own. Define and measure the process first, automate the boring steps next, and keep a person accountable for the outcome.

Key takeaways
  • 30.3% of 175 workplace tasks were fully completed by the best agent in TheAgentCompany benchmark.
  • Only about 5% of generative AI pilots produced rapid revenue gains, per MIT NANDA (2025).
  • 14% average, 34% for novices: the productivity lift from an AI assistant for 5,179 support agents.
  • 65% of Indian marketers using AI for content still rely on standalone tools (MMA-EY, 2026).
  • Write it, measure it, automate one step, review, then chain: the order that avoids failed pilots.

The argument behind the question is a common one in marketing forums. One camp says marketing teams expect agentifying to solve problems that are really missing processes. The other points to tiny growth teams at AI companies and asks whether the traditional marketing org is finished.

Both are partly right. The research gives a fairly precise picture of where agents earn their keep and where they quietly fail.

What can AI agents actually do for a marketing team today?

They do well on tasks a competent junior could do with a checklist, and badly on tasks that need context nobody wrote down.

The clearest test is TheAgentCompany, a Carnegie Mellon-led benchmark that built a simulated small software company, with internal chat, documents, project tools and colleagues. It gave AI agents 175 real workplace tasks across engineering, product, HR, admin, finance and data science.

30.3%tasks fully completed by the best agent (Gemini 2.5 Pro)
39.3%its score with partial credit
175workplace tasks in the benchmark
$4.20average model cost per task
Carnegie Mellon University campus buildings in Pittsburgh seen from Centre Avenue, with a clock tower and wooded hillside behind
Carnegie Mellon University in Pittsburgh, home of the team behind TheAgentCompany. Its simulated firm gave agents 175 tasks; the best finished about three in ten.Photo: Cbaile19, CC0, via Wikimedia Commons
AI tool: one step, you stay in controlYou askdraft this emailTool does one stepcopy appearsYou checkand decide what nextAI agent: many steps on its own, so it needs a process and a gateGoalweekly reportPlanlist the stepsAct in appspull, write, sendCheck itselfdid it work?loops until it thinks it is doneHuman review gate: a named person verifies the output against the source before it ships
A tool does one step when you ask; an agent chains many steps and loops. That loop is why it needs a written process and a human gate.

How the agents failed

The pattern of failures is the useful part. Agents did best on coding tasks and worst on admin, finance and data science work, which the authors note are easier for humans but depend on private company data and messy web interfaces. Agents:

  • got stuck on a closable pop-up;
  • failed to follow up when a colleague referred them to someone else;
  • in one case "solved" a task by renaming a chat user to the person they could not find.

Marketing operations is full of exactly these jobs: logging into five dashboards, chasing approvals, reconciling numbers that do not match.

TheAgentCompany: Benchmarking LLM Agents on Consequential Real World TasksFrank F. Xu, Graham Neubig et al. · Carnegie Mellon University and others · arXiv 2412.14161, v3 September 2025

The authors built a self-contained company environment with internal websites, files and simulated colleagues, and tested agents on 175 tasks spanning software development, project management, HR, admin, finance and data science. The best model completed 30.3% of tasks end to end (39.3% with partial credit) at about 27 steps and $4.20 per task. Simple tasks were often solved; long-horizon tasks, social coordination and complex web interfaces were not, and some agents faked progress when stuck. The limit: a simulated software company, and models improve every few months, so treat the numbers as a floor that is rising.

Read the paper (free)

Why do so many AI projects in companies fail to pay off?

Because most are bolted onto workflows that were not redesigned, and the money goes where the demos are rather than where the returns are.

MIT's NANDA initiative, in The GenAI Divide: State of AI in Business 2025, drew on about 150 leader interviews, a survey of 350 employees and 300 public deployments. It found only about 5% of generative AI pilots produced rapid revenue gains; the rest stalled with little measurable effect on profit and loss.

Two findings for marketing leaders

  • Budgets go to the wrong place. More than half of generative AI budgets went to sales and marketing tools, yet the biggest returns came from back-office automation.
  • Buying beats building. Projects bought from specialist vendors or built with partners succeeded about two-thirds of the time, against roughly a third of that rate for internal builds.

Marketing is spending the most and getting the least, largely because it automates the visible output (more posts, more ads) instead of the process underneath.

The story

In January 2023 the tech site Futurism reported that CNET had been quietly publishing personal-finance explainers written by an AI tool and credited to its CNET Money team. CNET audited all 77 articles from the trial of its internally built AI engine, and on 25 January it said 41 needed corrections. A small number needed substantial fixes; the rest had minor errors such as incomplete company names and vague language. Some also had phrases that were not entirely original, which CNET's plagiarism checker either was not used properly by the editor or failed to flag.

Editor-in-chief Connie Guglielmo paused the tool until CNET felt confident that the tool and its editorial process could prevent errors. The machine had done the writing quickly. What failed was the human review step, the part that makes the jobs in the table below safe to hand over.

Engadget: CNET corrected 41 of its 77 AI-written articles (25 January 2023)

Where does AI help marketing people most?

On structured, repeated work, and most for the least experienced people doing it. The strongest field evidence is from customer support, the closest cousin of marketing operations: Erik Brynjolfsson, Danielle Li and Lindsey Raymond tracked 5,179 support agents as a generative AI assistant was rolled out in stages.

Generative AI at WorkErik Brynjolfsson, Danielle Li, Lindsey R. Raymond · NBER Working Paper 31161 (Quarterly Journal of Economics, 2025)

Using the staggered rollout of an AI conversational assistant to 5,179 customer support agents, the authors found issues resolved per hour rose about 14% on average and about 34% for novice and low-skilled workers, with minimal gains for the most experienced. Customer sentiment and employee retention also improved, and the evidence suggests the tool spread top performers' practices to everyone else. The limit: one company and one assistant-style tool that suggested responses, not an autonomous agent acting alone.

Read the paper (free)
5,179support agents tracked in the field study
14%average productivity lift
34%lift for the least experienced staff

The jagged frontier

A second experiment shows the boundary. Fabrizio Dell'Acqua, Ethan Mollick and colleagues gave 758 Boston Consulting Group consultants GPT-4 for 18 realistic tasks.

  • Inside what they call the jagged frontier of AI ability, the AI users completed 12.2% more tasks, 25.1% faster and at higher quality.
  • Outside it, on one complex task chosen for that purpose, they were 19% less likely to get the right answer than colleagues without AI.
The jagged frontier (Dell'Acqua et al., 758 BCG consultants)Inside the frontier+12.2% tasks, 25.1% fasterOutside the frontier19% less likely to be righttask looks easytask looks similar, AI failsTasks below the line suit AI; tasks above it do not, and the line is not where you expect.
AI ability is uneven: two tasks that look alike can sit on opposite sides of the line. Map tasks one by one before automating.

The same tool helped and hurt within one workflow, which is why task-by-task mapping matters more than a blanket "use AI" mandate.

758BCG consultants in the experiment
+12.2%more tasks done inside the frontier
25.1%faster inside the frontier
-19%less likely to be right outside it

Why the support assistant worked

The mechanism is the lesson. The assistant worked because the company had years of resolved tickets to learn from, which is to say a documented process with known good answers. A marketing team with written briefs, past campaign results and a style guide has the equivalent.

A team that runs on the founder's head has no equivalent, and no agent can learn from what was never recorded.

Which marketing tasks should you give to agents, and which to keep?

Give agents tasks with defined inputs and outputs that a person can check in minutes; keep strategy, positioning, final approval and anything that commits money or speaks for the brand. This is my working map, built from the research above.

Marketing taskAgent fit todayHuman check needed
Competitor and market research summariesGoodVerify every number and link before it is used
Weekly performance reports from ad and CRM dataGood, once metrics are definedSense-check anomalies and the "why"
Lead enrichment, scoring and routingGood with clear rulesAudit a sample weekly; watch for bias in scoring
First drafts: emails, ad variants, social copyGoodEdit for voice, claims and accuracy
Creative resizing and versioningGoodSpot-check brand and legal elements
Campaign setup in ad platformsPartialApprove budgets, audiences and launch
Multi-step cross-tool workflows with approvalsWeak (long-horizon, messy UIs)Keep a person as owner
Positioning, pricing, brand strategyPoor; useful only as a sparring partnerHuman decision
Public replies in a crisis or complaintPoorHuman only

Use this to check what you just read: plug in a month of agent tasks and see what the benchmark completion rate leaves for people to finish.

What an agent leaves for your team

How mature is AI use among Indian marketers?

Widely adopted, but mostly as standalone tools rather than integrated workflows. The first MMA India and EY AI Marketing Maturity Study, published in May 2026 from interviews with more than 60 senior marketing leaders in 11 sectors, found 84 to 88% of AI adopters reporting medium to high impact, and over 90% saying so for ad creation, video editing and copy.

Indian marketers using AI who have fully integrated it, by area
See the data as a table
Fully integrated (%)
Propensity modelling57
Customer and sales journeys54
Measurement and insights43
Content and creative35

Source: MMA India and EY, AI Marketing Maturity Study, via Social Samosa (May 2026) · Share of AI users in each area. In content and creative, 65% still rely on standalone tools.

60+senior marketing leaders interviewed
84-88%of AI adopters report medium to high impact
65%still use standalone tools for content
~20%building in-house capability

The irony is in the bottom bar. Content and creative is where Indian marketers report the most impact and the least integration: 65% still use standalone tools, and only about 20% are building in-house capability.

The biggest barrier respondents named was measuring ROI, followed by integrating AI into existing systems. Their top 2026 priorities were:

These are process and measurement problems before they are AI problems.

In what order should a marketing team automate?

Document, measure, then automate the highest-volume checkable step, and only then chain steps into an agent. Skipping to the last step is how teams end up in the 95%.

1. Write itthe process2. Measuretime and quality3. Automateone step4. Reviewhuman check5. Chaininto an agentMost failed projects start at step 5.
The automation staircase. Each step needs the one below it; an agent is the last step, not the first.
  1. Write the workflow down. Inputs, steps, owner, output, and what "good" looks like, for one process such as the weekly performance report or inbound lead handling.
  2. Measure the baseline. Hours per week, error rate and turnaround. Without it you cannot prove the agent helped, which is the ROI problem the MMA-EY respondents named.
  3. Automate the highest-volume checkable step. Data pulls, enrichment, first drafts, resizing. Buy specialist tools where they exist; MIT NANDA found bought or partnered projects succeeded far more often than internal builds.
  4. Add a review gate. A named person signs off outputs against sources, at least until error rates are known and low.
  5. Chain steps into an agent last. Only after each step works alone, and with logs a person can audit.

In my own work, AI workflows cut production cycle time by about 60%. What made that possible was the boring part: briefs, templates and review steps that existed before any model touched them. The same idea, that small competent teams with strong systems beat large teams with weak ones, runs through my essay on the one-person marketing team. More on how I use AI in practice is on the applied AI page.

Does this mean marketing teams will shrink?

Some will, mostly in production roles, but the evidence points to fewer hands and more need for judgement, not to marketing without people. The Brynjolfsson study found the largest gains for less experienced staff, which means AI lets a small team perform like a larger, more experienced one. TheAgentCompany results show why someone still has to own the outcome: a third of tasks done autonomously is useful, but two-thirds is not.

The practical consequence for a founder is a different hiring mix. You need fewer people producing assets and more people who can define processes, judge quality and connect marketing to revenue. Two concrete places to start: lead handling over WhatsApp Business API, where rules are clear and volume is high, and deciding whether you need a fractional CMO to design the system before you automate it.

Questions people ask

What is the difference between an AI agent and an AI tool?

A tool does one step when you ask, such as drafting copy. An agent plans and carries out several steps on its own across apps, such as pulling data, writing a report and sending it, which is why it needs a defined process and audit logs.

Which marketing tasks are easiest to automate with agents?

Research summaries, recurring reports, lead enrichment and routing, first drafts and creative versioning. They share clear inputs, frequent repetition and outputs a person can check quickly.

Why do AI agent projects fail?

Mostly because the underlying workflow was undefined, success was never measured, or the agent was asked to handle long multi-step work across messy tools. Benchmarks show agents struggle with exactly those long-horizon tasks.

Should we build our own marketing agents or buy them?

Buy or partner first. The MIT NANDA research found externally bought or partnered AI projects succeeded about 67% of the time, against roughly a third of that rate for internal builds.

Can one person run marketing with AI agents?

For a narrow product with a defined funnel, a very small team can cover a lot, because AI raises the output of less experienced people most. Someone still has to own strategy, review outputs and answer for results.

Sources

  1. Xu et al.: TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks, arXiv (2025)
  2. Brynjolfsson, Li and Raymond: Generative AI at Work, NBER Working Paper 31161 (2023)
  3. Fortune: MIT report finds 95% of generative AI pilots at companies are failing (2025)
  4. Social Samosa: MMA-EY AI Marketing Maturity Study (2026)
  5. Engadget: CNET corrected 41 of its 77 AI-written articles (January 2023, source for the story)
  6. Dell'Acqua, McFowland, Mollick et al.: Navigating the Jagged Technological Frontier, Organization Science (2026)
AI agentsmarketing operationsautomationgenerative AImarketing team
Sapta

Sapta (Saptarshi Sarkar) is a mechanical engineer turned marketer in Bengaluru and a solo touring rider. He writes the seven Saptaverse blogs. Work with him or read more of Marketblog.

Keep reading in Marketblog

From the other blogs