Essay

AI orchestration, costed: what AI marketing work really costs

Published · Updated 20 min read

AI orchestration is the design of a workflow in which AI models and people each handle defined steps, with set handoffs, checks and escalation rules between them. The term comes from computing, where orchestration means the automated configuration, coordination, deployment and management of computer systems and software.[1] In marketing and operations it has come to mean something more practical: deciding which work goes to a model, which stays with a person, and what the whole arrangement costs.

This essay brings together four bodies of evidence. The first is controlled research on how much time AI tools save, and how well people judge that saving. The second is the public record of one large company, Klarna, that moved customer service to AI and then adjusted. The third is the published price list of the three largest model providers. The fourth is our own cost model, which prices seven common marketing and sales tasks in AI tokens and in human time at US Bureau of Labor Statistics wages. We close with a framework, the Cost Stack, and a step-by-step method any team can use to cost a workflow before it scales one.

Idea in brief

The problem. Most AI business cases compare a subscription fee with a salary. That comparison leaves out the line that decides the result: the time people spend checking, fixing and approving AI output.

Why it happens. Model prices are published and easy to compare. Review time is not, and people misjudge it. In a randomized trial, experienced developers using early-2025 AI tools took 19% longer on their tasks while believing AI had made them 20% faster.[2]

The gap (our data). We priced seven marketing tasks on 11 models from Anthropic, OpenAI and Google. A 1,500-word article draft costs $0.92 in tokens on the most expensive model we priced and less than one cent on the cheapest. One hour of review by a writer costs $52.83. In every task we modelled, people time made up between 77% and 98% of the cost of the AI workflow.

The answer. Cost the workflow, not the tool. Measure review minutes before you scale, set an escalation rule, and pick the model last.

Key findings

The essay in 3 minutes
  1. Finding 119%

    People misjudge AI time savings.

    In METR's 2025 trial, 16 experienced developers expected AI to make them 24% faster, took 19% longer, and afterwards still believed they had been 20% faster.[2]

  2. Finding 240%

    The measured gains are real but uneven.

    Professionals given ChatGPT finished writing tasks in 40% less time with 18% higher quality.[3] An AI assistant raised customer support output by 14% on average and by 34% for novice and low-skilled agents.[4]

  3. Finding 3100×

    Model prices vary a hundredfold.

    On 3 November 2026, published output prices across the 11 models we sampled ranged from $0.50 to $50 per million tokens.[5][6][7]

  4. Finding 4$0.92

    Tokens are the smallest line.

    In our model, a 1,500-word article draft costs $0.92 in tokens on the most expensive model and $53.75 in total once 60 minutes of review is added. Switching to the cheapest model saves $0.91. Doubling review time adds $52.83.

  5. Finding 570–78%

    Review time decides whether AI pays.

    Across seven tasks, the AI workflow cost 70% to 78% less than doing the work by hand under our base assumptions. If review takes as long as the original work, the saving disappears.

  6. Finding 6700

    A large company re-balanced after going all in.

    Klarna said in February 2024 that its AI assistant did the work of 700 full-time agents.[8] In May 2025 its CEO said a focus on cost had led to "lower quality", and the company began recruiting people for customer service again.[9]

Chapter 1 · 1 min

What AI orchestration is

AI orchestration sits one level above the tools. A team can own five AI subscriptions and still have no orchestration, because nobody has decided which steps go where, who checks what, and what happens when the model is wrong. For a small or mid-sized business, the unit of orchestration is one workflow: what goes in, what the model does, what a person checks, and what triggers a handoff.

Terms used in this essay

TermMeaning
AI orchestrationDesigning a workflow so that AI models and people each handle defined steps, with checks and handoffs between them.
TokenThe unit AI providers bill by. OpenAI's rule of thumb is that one token is about 0.75 words of English text.[6]
Input and output tokensInput is everything sent to the model (instructions, documents, the question). Output is what the model writes. Output is priced higher.
Reasoning tokensHidden working that some models produce before answering. Google states that thinking tokens are counted and billed.[7]
Loaded labour costWages plus benefits. In US private industry, wages were 70.0% of total compensation in the second quarter of 2026.[10]
Human in the loopA step where a person reviews, edits or approves AI output before it moves on.
Escalation ruleA rule that sends certain items straight to a person, such as refunds, complaints or legal terms.

Where AI orchestration helps, and where it does not

Type of workSend to AIKeep with peopleWhy
First drafts of routine emails, posts and summariesYesFinal checkFast to review, cheap to fix
Tagging, sorting, routing and data entryYesSpot checksHigh volume, clear right answer
Answers about policy, price or contract termsOnly with sources and limitsAnything unusualWrong answers carry legal and brand risk
Pricing, hiring, refunds above a limit, key accountsNoYesHigh stakes, needs judgment and accountability

The Hive's primer on what AI marketing is covers the marketing side of these choices.

Chapter 2 · 2 min

The evidence, part 1: what controlled studies measure

Most claims about AI productivity come from surveys or vendor case studies. Each study below compared people working with and without AI and measured the result.

Figure 1. People feel faster with AI than they are

Developers' forecast before the study
+24%
Developers' estimate after the study
+20%
Measured effect (early-2025 tools)
-19%

Effect on speed (%, positive = faster), -30 to +30

Experienced open-source developers working on 246 real tasks in their own repositories, randomly assigned to allow or disallow AI tools. The measured effect is the increase in time taken, shown as negative speed. Source: METR, "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity", July 2025.

METR, a non-profit research group, recruited 16 experienced developers from large open-source projects and had them complete 246 real issues, each averaging about two hours.[2] For each issue, AI use was randomly allowed or not. With AI, the developers took 19% longer, with a confidence interval of +2% to +39%.[11] Before the study they had expected a 24% speedup, and afterwards they estimated they had been sped up by 20%.[2]

The gap between the feeling and the measurement is the finding that matters for anyone costing a workflow. If skilled people cannot tell from experience whether a tool saves them time, a manager cannot either. Daniel Kahneman's work on confident but wrong judgment, summarised in the Hive's notes on Thinking, Fast and Slow, explains why a stopwatch beats a hunch.

The result also moved. In February 2026 METR published new data from 57 developers across more than 800 tasks.[11] For developers who had been in the original study, it estimated tasks took 18% less time with late-2025 tools, with a confidence interval from 38% less to 9% more. For newly recruited developers the estimate was 4% less time, with an interval from 15% less to 9% more. METR called the new data "only very weak evidence" of how much the speedup had grown, partly because 30% to 50% of developers said they were choosing not to submit some tasks they did not want to do without AI.[11]

Figure 2. What randomized and field studies of AI at work have measured

StudyWhoMeasureResult
Noy and Zhang, Science, 2023453 college-educated professionalsTime taken on writing tasks40% less
Noy and Zhang, Science, 2023453 college-educated professionalsOutput quality18% higher
Brynjolfsson, Li and Raymond, 2023 to 20255,179 customer support agentsIssues resolved per hour, all agents14% more
Brynjolfsson, Li and Raymond, 2023 to 20255,179 customer support agentsIssues resolved per hour, novice and low-skilled agents34% more
METR, July 202516 experienced developers, 246 tasksTime per task, early-2025 tools19% more
METR, February 202610 returning developersTime per task, late-2025 tools18% less (interval: 38% less to 9% more)
METR, February 202647 new developersTime per task, late-2025 tools4% less (interval: 15% less to 9% more)

Sources: Noy and Zhang, Science 381 (2023); Brynjolfsson, Li and Raymond, NBER Working Paper 31161, published in the Quarterly Journal of Economics 140(2), 2025; METR, July 2025 and February 2026.

Two patterns stand out. First, gains are largest on tasks with a clear format and a clear right answer, such as short professional writing and scripted support.[3][4] Second, gains are largest for less experienced workers. In the customer support study the tool raised output by 34% for novice and low-skilled agents, with minimal effect on experienced and highly skilled agents.[4] That matters for a marketing team: the person who saves the most time with AI is often the junior one, and the person who spends the most time reviewing is often the senior one.

Limits of the experimental evidence

Chapter 3 · 2 min

The evidence, part 2: Klarna, from 700 agents to "lower quality"

Klarna, the payments company, is the most quoted public example of AI replacing customer service work. Its record is useful because the company published numbers at each stage.

Figure 3. Klarna's AI assistant, in the company's own numbers

DateMetricValue
Feb 2024Conversations handled in the first month2.3 million
Feb 2024Share of customer service chatsTwo-thirds
Feb 2024Equivalent full-time agents700
Feb 2024Time to resolve an errand, before11 minutes
Feb 2024Time to resolve an errand, with the assistantLess than 2 minutes
Feb 2024Drop in repeat inquiries25%
Feb 2024Estimated 2024 profit improvement$40 million
May 2025Improvement in response times since launch82%
Q2 2025Revenue per employee (two years earlier: $0.4 million)$1.0 million
Q4 2025Revenue per employee$1.24 million
Q4 2025Headcount change since 2022Down 49%

Sources: Klarna press release, 27 February 2024; CX Dive, 9 May 2025, quoting a Klarna spokesperson; Klarna Q2 2025 and Q4 2025 earnings releases.

In February 2024 Klarna said its assistant had held 2.3 million conversations in its first month, two-thirds of its customer service chats, and was "doing the equivalent work of 700 full-time agents".[8] It said customers resolved errands in less than 2 minutes against 11 minutes before, that repeat inquiries had fallen 25%, and that it expected a $40 million profit improvement in 2024.[8]

Fifteen months later, chief executive Sebastian Siemiatkowski told Bloomberg that "cost unfortunately seems to have been a too predominant evaluation factor" and that "what you end up having is lower quality", as reported by CX Dive.[9] Klarna began recruiting people for customer service again. A company spokesperson told CX Dive the chatbot still handled two-thirds of inquiries, that response times had improved by 82% since launch, and that repeat issues had dropped 25%.[9]

Klarna did not abandon AI. Its later earnings releases report revenue per employee of $1.0 million in the second quarter of 2025, up from $0.4 million two years earlier, and $1.24 million in the fourth quarter of 2025, with headcount down 49% since 2022.[12][13] What changed was the split between machine and person. That is an orchestration decision, and it is the one this essay is about.

Limits of the Klarna evidence

A second public case shows that cost of error. In Moffatt v. Air Canada, decided in 2024, a British Columbia tribunal found the airline liable for wrong bereavement-fare information its website chatbot gave a customer, and rejected the argument that the chatbot was "a separate legal entity that is responsible for its own actions". It awarded CA$812 in damages, plus interest and fees.[14] The award was small. The international coverage was not, and that is the reputation risk line that never appears on a software invoice.

If you want an outside read on where AI already speaks for your business, and what a mistake there would cost, the Brave New audit is where we start.

Chapter 4 · 4 min

Original research: what seven marketing tasks cost in tokens and in people time

The studies above measure time. We wanted to put a price on both sides of the comparison using only published data. So on 3 November 2026 we built a transparent cost model. It prices seven common marketing and sales tasks on 11 AI models, using each provider's published per-token rates, and compares the result with the cost of the same task done by hand at US median wages. The full method is at the end of this essay, and the scripts and data are published with it.

Finding 1: model prices vary a hundredfold

Figure 4. Published API prices for 11 AI models, 3 November 2026

Output priceInput price
Claude Fable 5.1 (Anthropic)
$50
$10
gpt-6-astra (OpenAI)
$50
$10
Claude Opus 5.5 (Anthropic)
$20
$4
Claude Sonnet 5.5 (Anthropic)
$10
$2
gpt-6.1-sol (OpenAI)
$10
$2
Gemini 4 Argon (Google)
$10
$2
gpt-6.1-terra (OpenAI)
$5
$1
Gemini 3.8 Flash (Google)
$3.75
$0.75
Gemini 3.5 Flash-Lite (Google)
$2.5
$0.3
Claude Haiku 5.5 (Anthropic)
$0.5
$0.1
gpt-6-luna (OpenAI)
$0.5
$0.1

US dollars per million tokens (log scale)

Lime marks the priciest and cheapest output prices: $50 against $0.50, a 100x spread.

Standard pay-as-you-go prices for prompts under each provider's long-context threshold, without caching or batch discounts. Source: Anthropic, OpenAI and Google pricing pages, read 3 November 2026; Brave New Media analysis.

Output prices run from $0.50 to $50 per million tokens, a hundredfold spread. Prices also move. Google's page lists Gemini 4 Argon at $2 input and $10 output "through November 30, 2026", rising to $4 and $20 from 1 December 2026, and Gemini 3.8 Flash doubling from 1 January 2027.[7] Anthropic notes that its newer models use a tokenizer that "produces approximately 30% more tokens for the same text", so the same page can cost more on a model with a similar headline price.[5] We adjusted for that in the model.

Finding 2: people time is priced in tens of dollars an hour

Figure 5. Loaded hourly cost of the people who do marketing work, United States

Loaded hourly costMedian hourly wage
Customer service representatives
$30.76
$21.53
Interpreters and translators
$41.33
$28.93
Graphic designers
$43.24
$30.27
Sales reps, wholesale and manufacturing
$49.50
$34.65
Public relations specialists
$51.34
$35.94
Writers and authors
$52.83
$36.98
Editors
$53.51
$37.46
Market research analysts and marketing specialists
$54.10
$37.87
Marketing managers
$114.56
$80.19

US dollars per hour

Median hourly wage from BLS Occupational Employment and Wage Statistics, May 2025, national. Loaded cost divides the wage by 0.700, the wage share of total compensation for private industry workers in the BLS Employer Costs for Employee Compensation, Q2 2026. Source: US Bureau of Labor Statistics; Brave New Media analysis.

A marketing specialist at the national median costs about $54 an hour once benefits are included, and a marketing manager about $115. Those rates, not token prices, set the scale of every comparison that follows.

Finding 3: tokens are the smallest line in the AI workflow

Figure 6. Cost per task: done by hand versus an AI workflow with human review

By handAI workflow (with human review)
Tag and route an inbound RFQ email
$1.54
$0.33
Reply to a routine customer email
$3.08
$0.91
Summarise a 45-minute sales call into CRM notes
$16.50
$3.72
Write a 200-word product description from a spec sheet
$40.57
$9.22
Turn one article into five LinkedIn posts
$54.10
$13.99
Localise a 1,000-word product page into one new language
$123.99
$31.64
Draft a 1,500-word blog article
$211.31
$53.75

US dollars per task (log scale)

AI workflow split, model tokens versus human review: RFQ tagging $0.08 and $0.25; routine reply $0.14 and $0.77; call summary $0.42 and $3.30; product description $0.21 and $9.01; five LinkedIn posts $0.46 and $13.53; localised page $0.65 and $30.99; blog article $0.92 and $52.83.

Token cost uses Claude Fable 5.1 at $10 input and $50 output per million tokens, with two attempts per task, 1,000 words of instructions per call and two hidden reasoning tokens per visible output token. Task sizes and minutes are Brave New Media assumptions, listed in the Methodology. Source: Brave New Media cost model, 3 November 2026.

Even on the most expensive model we priced, and with generous allowances for instructions, reasoning and a second attempt, tokens never cost more than $0.92 per task. The cheapest model brought the same article draft under one cent ($0.0071 on gpt-6-luna). People time made up 77% of the AI workflow's cost for the smallest task and 98% for the largest.

Under our base assumptions, the AI workflow cost 70% to 78% less than doing each task by hand. That saving comes almost entirely from the assumption that review takes about a quarter of the original time. Which is why the next figure matters more than the last one.

Finding 4: review time, not model choice, moves the total

Figure 7. What changes the cost of one 1,500-word article draft

Review time halves (30 min)
$27.33
Cheapest model in sample (gpt-6-luna)
$52.84
Base case (Claude Fable 5.1, 2 attempts, 60 min review)
$53.75
Model prices double
$54.67
Agent loop: 10x the tokens
$62.02
Review time doubles (120 min)
$106.58
Agent loop: 100x the tokens
$144.70
Human-only baseline (240 min)
$211.31
Review time equals writing time (240 min)
$212.23

Base case: $53.75

Human-only baseline: $211.31

US dollars per article

One article drafted by a model and reviewed by a writer at a loaded $52.83 an hour. Source: Brave New Media cost model, 3 November 2026.

Moving from the most expensive model to the cheapest saves $0.91 per article, under 2%. A doubling of every model price adds $0.92. A doubling of review time adds $52.83, almost doubling the total. Only one model-side change came close: an agent that loops 100 times over the same material, re-reading and rewriting, which pushed the article to $144.70. Agent workflows can do that, and it is one reason to cap how many steps an agent may take.

The last two rows are the METR lesson in money. If review and repair take as long as writing from scratch, the AI workflow costs slightly more than the human one, and nobody notices unless someone measured both.

Finding 5: what this looks like over a month

Figure 8. A month of marketing and sales tasks for a hypothetical industrial supplier

By handAI workflow
Tag and route an inbound RFQ email
$924
$198
Reply to a routine customer email
$1,232
$364
Summarise a 45-minute sales call into CRM notes
$1,320
$298
Write a 200-word product description from a spec sheet
$811
$184
Turn one article into five LinkedIn posts
$433
$112
Localise a 1,000-word product page into one new language
$992
$253
Draft a 1,500-word blog article
$845
$215

US dollars per month

This company is made up, but the volumes are typical of a mid-sized B2B supplier. Same assumptions as Figure 6. Source: Brave New Media cost model, 3 November 2026.

For this made-up supplier, the work takes about 160 hours a month by hand and costs $6,557. The AI workflow needs about 36 hours of review and costs $1,624, of which $153 is model spend on the most expensive model, or $31 on a mid-tier model such as Claude Sonnet 5.5. The freed time is worth something only if it moves to other paid work, such as calling the accounts the summaries flagged or opening a new market.

Limits of our cost model

Chapter 5 · 1 min

A framework: the Cost Stack

The research points to a fixed order of cost lines, from the most visible to the most decisive. Most business cases stop at the first two rows. The money is in the last three.

LayerWhat it coversEvidenceHow to measure it
1. TokensModel usage per taskUnder $1 per task on the priciest model we sampled (Figure 6)Provider usage dashboard, cost per task
2. Seats and toolsSubscriptions, connectors, hostingFixed monthly fees, often overlappingInvoice review
3. Setup and upkeepPrompts, testing, integrations, re-testing when models changePrices and models changed within months (Figure 4)Hours logged, spread over the workflow's life
4. ReviewReading, correcting and approving output77% to 98% of AI workflow cost in our model (Figure 6)Stopwatch on a sample of items
5. ErrorsWhat a mistake costs when it gets throughKlarna's "lower quality"; Air Canada's liability rulingError rate times average cost to fix

The order matters because each layer is larger and harder to see than the one above it. A team that negotiates a 20% discount on tokens and ignores review has optimised the smallest number on the page.

Chapter 6 · 1 min

How to cost an AI workflow in six steps

You can cost any AI workflow with a spreadsheet and two weeks of honest timing.

Step 1: Time the work as it is today

Pick one workflow. Count the items handled a month, time a sample of ten by hand, and multiply by the loaded hourly cost of the person doing it. Use your own payroll figures, or the BLS rates in Figure 5 as a starting point. Without this baseline you cannot prove anything later.

Step 2: Design the split and the escalation rule

Write down which steps the model does, which a person does, and which items go straight to a person. Decide the escalation rule before the pilot, for example "anything about refunds, complaints, pricing or contract terms goes to a human".

Step 3: Price the tokens

Count the words that go in and come out of a typical item, convert at about 0.75 words per token, and multiply by the provider's published rate.[6] Add your instructions to every call, and allow for reasoning tokens and retries. It will almost always be the smallest number in the table, which is the point of checking it.

Step 4: Time the review

During a short pilot, time how long a person takes to read, fix and approve each item. Do not ask people to estimate it. The METR study shows that estimates can point the wrong way.[2]

Step 5: Count the errors

Track what share of items needed a real correction after approval, and what each one cost to fix. Multiply the two.

Step 6: Run the sensitivity check, then re-cost at 30 and 90 days

Change review time and error rate and watch the total. If a small change turns a saving into a loss, the workflow is fragile. Re-time the work after 30 and 90 days, because models, prices and your team's habits all change.

Tip: Count only time that moves to other paid work. If a saved hour becomes a longer lunch, it is a nicer week, not a cost saving.

Which workflows to cost first

Workflow typeVolumeCost of an errorReview timeStart here?
Tagging and routing inbound requestsHighLowSecondsYes
Call summaries into the CRMMediumLow to mediumA few minutesYes
Routine customer replies with an escalation ruleHighMediumA minute or twoYes, with a pilot
Product pages and localisation for new marketsLow to mediumMediumLarge share of the originalAfter a pilot
Long-form thought leadershipLowHigh for the brandClose to the originalLast

Chapter 7 · 1 min

How Brave New Media builds AI flows

At Brave New Media, an AI flow is AI orchestration pointed at growth: a designed workflow that helps a business take one clear idea to more channels, more markets and more buyers, without adding headcount at the same rate. The Cost Stack applies to every flow we build.

Every flow has five parts: signals (what a buyer already did, such as an RFQ or a spec sheet download), segmentation, the flow of messages itself, the automation that connects your existing tools, and tracking to pipeline and revenue. In each part a person holds a defined place. Nothing sends without a person approving the messages the flow can send, and a person reads the numbers and decides what runs next. Human approval is review time, so we budget it, and tracking to revenue gives the baseline and the 90-day review.

Chapter 8 · 1 min

What this means for leaders

For a manufacturing or supply-chain business, the case for AI orchestration is rarely about cutting a team. It is about what the same team can reach. In our made-up supplier, the largest cost lines by hand were call summaries, customer replies and RFQ handling: the daily work that keeps existing accounts served. Moving that work to a reviewed AI flow frees time for the tasks that grow a customer base, such as following up on every signal from a new region or localising product pages for a market the company has never sold into.

That expansion angle is where token prices start to matter in a good way. Localising a 1,000-word product page cost $0.65 in tokens on the most expensive model we sampled, against $124 for a translator working from scratch. Review by a fluent person is still needed, and in regulated or technical categories it should be. But in our model a reviewed AI workflow localised a page for $31.64, about a quarter of the $124 cost by hand, so a mid-sized manufacturer can test several new markets for the price of one and keep the ones that answer. We see that pattern across supply chain and manufacturing: long sales cycles, few buyers per market, and a strong product that nobody abroad has heard of yet.

The Klarna record gives the warning. Scale without quality control turns into "lower quality", and a brand pays for that in trust, which the Hive's entry on trust architecture explains in more depth.[9] Your brand and your reputation are built on the parts that still sound like you.

If you want an AI flow that takes a proven offer to more buyers and more markets, costed line by line before it scales, start with the Brave New audit.

Chapter 9 · 1 min

Common mistakes

Chapter 10 · 1 min

Methodology

Prices. On 3 November 2026 we downloaded the public pricing pages of Anthropic (docs.claude.com), OpenAI (developers.openai.com, Markdown version of the pricing page) and Google (ai.google.dev, Gemini API pricing) with a script that identified itself as "BraveNewResearch/1.0" with a link to brave-new.com. OpenAI's marketing pricing page returned HTTP 403 to that script; its developer documentation page did not. We used standard pay-as-you-go rates for short-context prompts, without caching, batch or priority discounts. We chose 11 models covering three tiers from each provider. For Anthropic models we multiplied token counts by 1.3, following Anthropic's note that its newer tokenizer produces approximately 30% more tokens for the same text.

Wages. We pulled national median hourly wages for nine occupations from the BLS Occupational Employment and Wage Statistics, May 2025, through the public BLS API (series OEUN000000000000 plus the occupation code and data type 08). We converted wages to loaded cost by dividing by 0.700, the share of wages and salaries in total compensation for private industry workers in the BLS Employer Costs for Employee Compensation, second quarter of 2026 (series CMU2020000000000P).

Tasks. Seven tasks, each with a stated input size, output size, human minutes, review minutes and occupation. These are our assumptions:

TaskOccupationInput wordsOutput wordsMinutes by handReview minutes
Tag and route an inbound RFQ emailCustomer service representatives5005030.5
Reply to a routine customer emailCustomer service representatives80015061.5
Summarise a 45-minute sales callSales reps, wholesale and manufacturing6,500300204
200-word product descriptionMarketing specialists2,0002004510
Five LinkedIn posts from an articleMarketing specialists1,8007006015
Localise a 1,000-word product pageInterpreters and translators1,2001,10018045
1,500-word blog article draftWriters and authors3,0001,50024060

Token cost. For each task: two attempts, each sending the task input plus 1,000 words of instructions, and each producing the visible output plus two hidden reasoning tokens per visible output token. Words convert to tokens at 0.75 words per token, OpenAI's stated rule of thumb for English.

Limits. This is a model, not a measurement of any company's work. It leaves out errors, setup, seats and upkeep. It uses US national wages, and the human and review minutes are illustrative. Its value is that every input is visible and can be swapped for your own.

The scripts, the raw pricing and wage downloads, and the data behind every figure are published with this essay as CSV files.

Chapter 11 · 1 min

Frequently asked questions

What is AI orchestration in simple terms?

AI orchestration means planning how AI models and people share a piece of work. You decide which steps the model handles, which steps a person checks or does, and which items go straight to a person. For most businesses it is less about software and more about designing one workflow carefully, then measuring whether it saves money.

How much does it cost to run an AI model on a marketing task?

Less than most people expect. In our cost model, using published prices from 3 November 2026, a 1,500-word article draft cost $0.92 in tokens on the most expensive model we priced and under one cent on the cheapest. The larger cost is the person who reviews the draft, about $53 for an hour of a writer's time.

Does AI workflow automation save money?

It can, if review is fast. In our model the AI workflow cost 70% to 78% less than doing seven marketing tasks by hand, assuming review takes about a quarter of the original time. If review takes as long as the original work, the saving disappears. Time a sample of real items before you decide.

How can you tell if an AI tool is saving your team time?

Time the task by hand first, then time it with the tool, including checking and fixing the output. Do not rely on how fast it feels. In METR's 2025 trial, experienced developers using AI took 19% longer while believing they had been 20% faster. A short pilot with a stopwatch settles the question.

Which AI model should a small business choose to keep costs down?

Choose on quality and review time first, price second. Output prices across the 11 models we sampled ranged from $0.50 to $50 per million tokens, yet switching from the priciest to the cheapest model changed the total cost of an article draft by under 2%. A model that needs less fixing usually saves more than a cheaper one.

Your next step

This week, pick one workflow you have automated, or want to, and time ten items by hand with a stopwatch. Then run the same ten through your AI tool and time the review, including every fix. Put both numbers into the Cost Stack. That one table will tell you more than any vendor demo.

If you want a team to build and cost an AI flow that takes a proven offer into more markets, start with the Brave New audit.

Brave New Media is a marketing agency that builds AI flows to help companies scale their ideas, grow their customer base and expand into new markets, with experience across supply chain and manufacturing.

See also

References

1. "Orchestration (computing)." Wikipedia, read 3 November 2026. en.wikipedia.org/wiki/Orchestration_(computing))

2. METR. "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity." 10 July 2025. metr.org. Paper: Becker, J., Rush, N., Barnes, E. and Rein, D. arxiv.org/abs/2507.09089

3. Noy, S. and Zhang, W. "Experimental evidence on the productivity effects of generative artificial intelligence." Science 381(6654), 187 to 192, 2023. pubmed.ncbi.nlm.nih.gov/37440646

4. Brynjolfsson, E., Li, D. and Raymond, L. "Generative AI at Work." NBER Working Paper 31161, 2023; Quarterly Journal of Economics 140(2), 2025. nber.org/papers/w31161

5. Anthropic. "Pricing." Claude documentation, read 3 November 2026. docs.claude.com

6. OpenAI. "Pricing" and "Key concepts." API documentation, read 3 November 2026. developers.openai.com

7. Google. "Gemini Developer API pricing." Read 3 November 2026. ai.google.dev

8. Klarna. "Klarna AI assistant handles two-thirds of customer service chats in its first month." 27 February 2024. klarna.com

9. Doerer, K. "Klarna changes its AI tune and again recruits humans for customer service." CX Dive, 9 May 2025. customerexperiencedive.com

10. US Bureau of Labor Statistics. Employer Costs for Employee Compensation, private industry, Q2 2026, and Occupational Employment and Wage Statistics, May 2025. bls.gov/oes

11. METR. "We are Changing our Developer Productivity Experiment Design." 24 February 2026. metr.org

12. Klarna Group plc. Second Quarter 2025 Results. s205.q4cdn.com

13. Klarna Group plc. Fourth Quarter 2025 Results. s205.q4cdn.com

14. Moffatt v. Air Canada, 2024 BCCRT 149. Summary: en.wikipedia.org/wiki/Moffatt_v._Air_Canada

15. Brave New Media. Cost model of seven marketing tasks on 11 AI models, 3 November 2026. Data published with this essay.