Scaling Video with AI Agents — montage ROI and Operations Design

The HUMAN_INPUT markers referenced in this article are placeholders that the AI writing skill leaves in the body text to indicate “a human should fill in the actual value here later” (example format: <!-- HUMAN_INPUT: insert number -->).

This article is written for executives, marketing leads, and product managers who need to think through the business case, cost structure, quality management, and mid-term ROI of scaling video content with AI agents. The architectural details of the agents themselves — the pipeline DAG, video component implementation — are covered in the companion technical article montage Pipeline: Technical Design Details (tech cluster sister pillar). This article contains no code.

The importance of video content is not something any marketing lead would dispute. Yet most organizations find themselves unable to produce it at scale even when they know they should. Production costs are high; outsourcing makes quality control difficult; in-house production drains human resources. A goal of “10 videos a month” stalls at 2 in practice — or never gets started at all.

AI-agent-driven video production has the potential to change this. But it is not a story about “AI automatically generating videos on your behalf.” It is a story about design — deciding which judgments stay with humans and which execution steps get delegated to agents.

This article uses Yakumo’s video production pipeline montage as a concrete example to lay out a comparative cost analysis, quality gate design, review structure, mid-term ROI estimation framework, and the key axes for making an adoption decision.

Context (as of writing, May 2026): Yakumo’s montage has not yet reached mass production — it is still in the testing phase. The pipeline foundation (agent division between researcher / scriptwriter / composer / reviewer) and video block templates are built, and we are running validation with 56+ tests under tests/ to produce output that satisfies the quality benchmarks in kpi.json. Meanwhile, publishing routes (automated upload to YouTube / SNS) are intentionally blocked by scripts/hooks/pre-publish-gate.sh (exit 2), and the output/ directory contains only test-named artifacts: *-e2etest-*, *-smoke-*, load-test-*, dev-test/, etc. The “cost structure,” “ROI estimation,” and “break-even calculation” frameworks in this article are designed to be validated with actual data once mass production begins — we do not yet have concrete measured figures. Read this as a framework for anyone in the same position (system built, but not yet at scale) trying to decide how to move toward mass production.

Three key takeaways:

1. AI-agent video production is not a technology challenge — it is a business design challenge. The technical machinery cannot run until you decide what to produce at scale, who handles reviews, and how quality is defined.
2. The real cost question is not “cost per video” but “cost per unit of scale.” The adoption decision hinges on projecting what changes at 50 or 100 videos per month, not on comparing individual unit prices.
3. The more you delegate to AI agents, the more the quality of human review judgment matters. The design of quality gates and the definition of where humans must stay involved determine the quality of what gets produced at scale.

What Does AI-Agent Video Production at Scale Actually Mean?

“Making videos with AI” currently means two different things depending on who is speaking.

One approach uses generative AI video tools (Sora, Runway, Kling, etc.) to have AI generate the video footage itself. The other uses agents to handle the production process — research, scripting, asset selection, editing, quality review — and renders the output using a structured rendering engine like Remotion.

montage is the second approach. Rather than AI “drawing the footage itself,” it is a system where “agents handle the process of creating video content.” This distinction matters for business decision-making, so it is worth establishing upfront.

Comparison with Traditional Video Production

Breaking down a traditional video production process, the key stages look roughly like this.

Research and Planning: Deciding what topic, structure, and audience a video is for. This involves market research, competitor video analysis, and gathering viewer insights.

Script and Narration Design: Writing the script that forms the backbone of the video. This means designing an engaging hook, the core message to deliver, and a closing call to action — while simultaneously accounting for read-time and captions.

Asset Production and Editing: Preparing text, graphics, charts, and narration audio, then assembling them into a video. Whether templates are used or everything is built from scratch, and which tools are selected, determines how labor-intensive this stage is.

Quality Review and Revision: Checking for brand guideline compliance, factual accuracy, and overall production quality, then applying revisions. This is often the most cost-intensive cycle in outsourced video production.

Distribution and Analysis: Exporting in formats suited to each platform, publishing, and feeding audience response back into the next round of planning.

Across this entire process, human labor concentrates most heavily at “research and planning” and “quality review and revision.” “Scripting” is also high-skill work that commands higher rates when outsourced. “Asset production and editing” has seen some automation through templates and tooling, but still demands significant labor.

AI-agent-driven production is an attempt to redistribute the ownership of this process.

The Business Meaning of Agent Division (researcher / scriptwriter / composer / reviewer)

In montage, the production process is divided among four primary agents: researcher, scriptwriter, composer, and reviewer.

researcher: Collects information on a specified topic or company and organizes the data, facts, and context that will become the raw material for the video. It compresses what would be substantial research time for a human into an initial draft.

scriptwriter: Takes the researcher’s output and generates the script. Designs narration and captions suited to the video’s length, target audience, and tone.

composer: Takes the script and assembles the video structure — which scenes in which order, in which design — then generates Remotion-based video data.

reviewer: Checks whether the generated video meets quality criteria. Detects misinformation, brand violations, and structural problems, then either issues correction instructions or approves the video.

What is the business significance of this four-stage division? Each stage these agents handle was previously the work of a human specialist.

Hiring an external researcher costs time and money. The scriptwriter’s quality determines content quality. The video creator needs design skills. Quality review is typically handled by a director. The shift here is not about replacing these roles with agents, but about restructuring so that agents produce the first draft and humans handle approval. That is the real substance of “delegation.”

An Overview of the montage Production Flow

In rough terms, the montage production flow aims for a pipeline where “you pass in an input (a ticker or company name), and the system automatically generates video drafts until the reviewer approves.” At this point, this automated chain runs only partially in the test environment, and the publishing route (automated upload to YouTube / SNS) is still intentionally blocked — we are currently validating at the smoke test level whether “the reviewer’s approval criteria pass consistently” before opening the pipeline to production.

By design, human judgment will be concentrated at two points.

The first is specifying what to create (input): setting the topic, target channel, and tone in advance, or specifying them per run.

The second is deciding whether to publish (approval): the reviewer agent automatically runs quality checks, and anything that meets the criteria gets routed to distribution. Anything that falls short is escalated to a human.

When this structure runs at production scale, human capacity should not grow proportionally with output volume. The design intent of montage is that setting up inputs and handling exceptional approval cases are the only human tasks.

Understood as a business proposition, the claim that “output can increase without changing headcount” depends entirely on the quality of inputs and the design of the review structure. As the team building montage, the difficulty we actually feel in the testing phase is here — the work of articulating “templates,” “guidelines,” and “KPIs” in writing has consumed more time than the code itself.

Comparative Cost Analysis

When evaluating AI-agent video production, the first question is always: “How much does the cost change?”

This section lays out the cost structures for human production versus montage. Since Yakumo has not yet reached mass production, this article does not contain actual measured figures for in-house costs, outsourcing rates, or API unit costs. The plan is to collect real data when production begins and update this article. Read what follows as a framework for understanding the structure, not as concrete figures.

The Cost Structure of Human Production

The cost of a single video produced by humans varies considerably depending on the production setup. Three primary variables drive it.

Labor cost (in-house): The human time cost invested across planning, research, scripting, editing, and review. For example, if a video requires 5 hours total and the responsible person’s time is valued at the equivalent of 3,000 JPY per hour, that is 15,000 JPY per video — or 300,000 JPY/month at 20 videos per month.

Outsourcing cost (external creators): This varies widely depending on video type, length, and quality level. Standard marketing videos (30–90 seconds with captions) fluctuate significantly with the creator’s level and number of revision rounds. Substitute your own recent estimates here, and include the cost of revision communication in the total.

Production lead time: When outsourcing, a single delivery typically takes one to two weeks. This becomes the bottleneck for scaling. Placing simultaneous orders for 10 videos a month often exceeds what a production company can handle.

Whether in-house or outsourced, a fundamental characteristic is shared: costs scale linearly with volume. Going from a budget for 10 videos to 50 videos requires roughly five times the cost. This linear cost structure is the fundamental constraint of any volume production strategy.

The Cost Structure of montage

The cost of using montage breaks down into three components.

API costs: The LLM API costs incurred when the researcher / scriptwriter / composer / reviewer agents run. Per-video token consumption varies with video length, complexity, and the number of retries. At Yakumo, we are only observing this at the smoke test level in the testing phase; actual figures at 50 or 100 videos per month have not yet been collected (we plan to aggregate them once production begins).

Review cost (human time): The human time cost for videos that the reviewer agent cannot automatically approve, or that require final human sign-off. Higher approval rates mean lower costs; stricter quality criteria mean higher costs. We also plan to measure average review time per video once we enter the production phase — it has not been captured yet at the current testing stage.

Setup cost (initial investment): The up-front design cost of incorporating brand guidelines, theme configurations, channel settings, and pipeline definitions into montage. Once configured, this does not scale with additional volume — it behaves like a fixed cost. At Yakumo, the “setup” itself is still ongoing; rather than a fixed figure, it is running as an “accumulating investment that grows each time we redesign something” (we plan to treat the start of mass production as the point at which setup is finalized and produce a total).

The structural property of montage is that variable costs (API costs + review time) tend not to scale linearly as volume increases. The more standardized the input configuration becomes, and the higher the reviewer approval rate climbs, the lower the per-video human cost tends to be.

Projecting Cost Changes at Production Scale

When comparing these two cost structures, a per-unit price comparison is not the right lens. What matters is total cost at N videos per month.

The following shows the structure of the comparison framework (substitute your own organization’s actual figures — Yakumo does not yet have measured values):

The critical variable in this framework is the approval rate. As the share of videos automatically approved by the reviewer agent increases, human cost at scale approaches a plateau — the marginal human effort between 100 and 200 videos per month can approach near-zero.

This plateau characteristic is precisely why AI-agent video production offers “economies of scale.” The point where the linear cost curve of human production intersects the logarithmic curve of AI production — the break-even point — becomes the axis for the adoption decision.

Quality Gate and Quality Management Design

After cost, the next question is: “Can quality be ensured?”

Can videos generated by AI agents be released to viewers as-is? Will they reflect well on the brand? Are there factual errors? These concerns are reasonable, and if they cannot be resolved, production at scale will never start.

Quality management design should be considered in three layers.

Quality Challenges in AI-Generated Video (Consistency / Brand Alignment / Information Accuracy)

AI-agent-generated videos have quality challenges that differ from those in human-produced work. Three of the most important:

Consistency problems: Even within the same brand’s videos, when agents generate each one independently, font, color, pacing, and tone can drift subtly. “This one feels a bit too formal” or “this one feels different from the last” — these issues occur more frequently when the inputs passed to agents (brand guidelines, theme settings) are vague.

Brand alignment problems: Word choice, strength of claims, and compliance with prohibited expressions. Agents may use language the organization has classified as off-limits — for example, superlatives like “revolutionary” or “industry-first,” or phrasing that creates misimpressions about the business model.

Information accuracy problems: When the researcher’s collected and organized information contains errors or is out of date, those errors carry through directly into the video. Accuracy verification is especially critical for videos dealing with financial figures, legal information, or specific company data.

Designing quality gates requires recognizing these challenges first.

Division of Responsibility Between the Reviewer Agent and Human Review

The practical quality management design is to separate “quality problems that agents can detect” from “quality problems that require human judgment.”

Quality checks handled by the reviewer agent are those that can be defined with rules:

• Presence of prohibited expressions (scanning against a specific word list)
• Conformance of brand colors and fonts to specified values
• Video structure (balance of opening, body, and closing)
• Whether text length and duration fall within specified ranges
• Caption-to-audio sync drift

These can be detected mechanically. The reviewer agent holds a checklist, automatically approves anything that passes, and routes anything that fails into a correction flow.

Reviews that humans must handle are those requiring contextual judgment:

• Factual accuracy (especially verification of numbers and citation sources)
• Subtle tone drift (“technically OK by the rules, but not how we sound”)
• Timing appropriateness (is this the right moment to publish a time-sensitive video?)
• Consistency of brand message (is this appropriate not just as a standalone but in the context of a series of videos?)

Keeping these two separate — not conflating them — allows human review effort to be compressed to “only the parts that machines cannot judge.”

Defining and Managing Quality Standards

Telling someone to “ensure quality” without defining quality standards gives neither agents nor humans anything to act on. The prerequisite for quality management is articulating quality criteria in writing.

Concretely, three things need to be prepared in advance.

Prohibited expression list: Words and phrasings that must never be used. Superlatives, prohibitions on competitor mentions, and specific banned formulations. The reviewer agent scans against this list.

Brand guidelines: Font, color palette, logo usage rules, caption formatting. The unified visual standard for the “look” of videos.

Quality score definition: What evaluation criteria does the reviewer use, and what score threshold triggers automatic approval? This is the work of translating the feeling of “high quality” into a measurable score.

This preparation must happen before montage is deployed. If it already exists, it gets incorporated into montage’s configuration. If it does not, creating it — at minimum a “list of words we never use” and “caption font and color rules” in writing — is the first significant investment.

Review Design — Where to Keep Humans Involved

After cost and quality, the next question is: “How should human involvement be designed?”

The instinct to hand everything to agents is understandable. But if quality failures in mass-produced videos could create a brand crisis, some human judgment needs to stay in the loop. Getting this wrong in either direction — too much involvement loses the benefits of scaling; too little causes quality failures.

Decision Criteria for Full Automation vs. Human-in-the-Loop Hybrid

Two axes determine whether “full automation” or “keeping humans involved” is the right call.

Magnitude of risk: The scope of impact if a quality problem occurs. Is this a video published externally or used for internal operations? Is brand damage risk high or low? Does it handle regulated information — stock prices, personal data, medical information? Higher risk means more human involvement checkpoints.

Degree of repeatability: Are you producing the same type of video repeatedly, or is each one unique? High-volume templated work (weekly earnings recap videos, standard promotional videos) is readily automatable. Creative videos that begin with fresh planning each time have fewer stages agents can reliably handle.

Mapping these two axes, montage delivers the most value for mass production of videos with moderate or lower risk and high repeatability.

For example:

• Low risk, high repeatability (easy to fully automate): Standard product introductions, facility walkthroughs, company overview videos. Templated structure with information swapped in.
• Medium risk, high repeatability (human for final approval only): Regularly published news explainers, market summaries, earnings recaps. The format is set but information accuracy verification is required.
• High risk, low repeatability (agents as support only): Campaign advertising, brand message videos, executive commentary footage.

montage’s current design is optimized for the “medium risk, high repeatability” category.

Designing the Approval Flow (When in the Process Do Humans Check?)

The timing of human review significantly affects production efficiency depending on how the flow is designed.

Upstream check: A human reviews the data the researcher has collected before it is passed to the scriptwriter. Accuracy is ensured at the earliest point in the pipeline. This reduces rework but increases human labor.

Downstream check (finished product review): A human does a final review after the composer has assembled the video. Labor concentrates at the end, but judgment can be based on seeing the finished product. If revisions are required, the cost of going back upstream is high.

Parallel check: The reviewer agent runs automated checks and only passes to a human anything it cannot automatically approve. Humans only see the cases “the agent could not auto-approve.” This design avoids labor concentration while preserving human judgment for the cases that need it.

montage is fundamentally designed around the “parallel check” approach. The reviewer agent handles the first gate; humans focus on exception handling and sampled verification.

A practical threshold for designing the approval flow: can the reviewer agent achieve a sustained approval rate above a certain level (say, 80%)? That is, can it auto-approve roughly 8 out of 10 videos? If the approval rate is low, the priority is adjusting the agent’s quality criteria or refining the brand guidelines. Once the approval rate is high, human review labor compresses in practice.

How Brand Guideline Compliance Is Verified

A prerequisite for running the approval flow is that brand guidelines exist in a form that can be mechanically checked.

Even if you have a strong intuitive sense of “what a video looks and sounds like for our brand,” if that is not written down, it cannot be passed to agents or verified by the reviewer.

Three minimum requirements for incorporating brand guidelines into montage:

Prohibited word list: A list of words that must never be used. In addition to superlatives (“revolutionary,” “industry-first,” “fully automated”), include organization-specific prohibitions (competitor names, matters under dispute, expressions that implicate personal information).

Tone definition: “Formal or casual?”, “declarative or politely suggestive?”, “warmth-first or expertise-first?” Without this definition, the scriptwriter will produce a different tone every time.

Visual rules: Color palette, permitted fonts, logo placement rules, forbidden layouts. The rules the composer follows when assembling the video.

With these in place, the reviewer agent’s automated check accuracy improves, and the cost of human verification decreases. Put another way: organizations with underdeveloped brand guidelines will take longer to see results from montage adoption.

Mid-Term ROI Estimation

With quality management design understood, the next question is: “Is the investment worthwhile?”

This section presents a framework for estimating ROI. Yakumo has not yet entered mass production, and the actual measured values needed to calculate break-even (monthly cost savings, initial setup costs) are not finalized. Substitute your own organization’s figures. Yakumo’s numbers will be updated in this article once mass production begins.

ROI Calculation Framework

Three variables drive the ROI of AI-agent video production.

Production cost reduction (cost axis): How much does producing the same number of videos cost before versus after montage adoption?

Formula: (Human production cost/month) − (montage cost/month) = monthly cost savings

Scale increase (volume axis): The increase in videos producible after montage adoption. Even at the same cost, higher output means higher ROI.

Formula: Videos/month after adoption ÷ Videos/month before adoption = scale multiplier

Quality maintenance cost (quality axis): The human involvement cost required to maintain the quality of mass-produced content (review time, revision handling).

Formula: Review time per video × monthly video count × hourly rate = monthly quality maintenance cost

ROI is captured in this simple formula:

ROI = (Cost savings + Additional value from scale increase) / (Implementation cost + Monthly quality maintenance cost)

The challenge is quantifying “additional value from scale increase.” If monthly video output goes from 10 to 50, what is that worth in monetary terms? This requires prior measurement of how video content contributes to business results (leads generated, site traffic, brand awareness, etc.). Organizations that lack this measurement will find the “value side” of ROI difficult to estimate.

In that case, it is fine to calculate ROI based on cost savings alone and measure additional value separately afterward — a two-stage evaluation approach.

Break-Even Point (How Many Videos Before the Investment Pays Off?)

Adopting montage requires up-front costs: the labor of developing brand guidelines, theme configurations, and pipeline design. How many months of monthly cost savings it takes to recover that initial investment is the break-even point.

Formula:

Break-even (months) = Initial setup cost / Monthly cost savings

If initial cost is S and monthly savings are R, recovery takes S / R months — a simple formula. Plugging your organization’s figures into S and R reveals a realistic break-even timeline.

A general decision framework: break-even within 3 months = “early GO”; within 6 months = “worth considering”; beyond 12 months = “build the right use case and scale first.”

The Value of Video Content as a Long-Term Asset

In addition to cost savings and scale projections, accounting for the “asset value” of video content connects directly to mid-term ROI.

Video content continues to be viewed after publication. Videos accumulated on YouTube and social media attract viewers without additional production investment. This is the compounding effect of content assets.

Producing 50 videos per month for a year yields a library of 600 videos. If each video averages 100 views per month, by year-end the accumulated asset is generating 60,000 monthly views. This compounding structure is a form of mid-term value that monthly cost comparisons alone cannot capture.

Capturing this compounding effect requires a structure that can sustain production at scale. If cost, quality, or review capacity becomes a bottleneck and production stops, the asset accumulation stops too. montage’s role can be understood as “the mechanism that makes sustained production possible.”

Adoption Decision Framework

With cost, quality, and ROI understood, here is a framework for deciding: “Is montage the right fit for our organization?”

Conditions for Organizations That Should Adopt montage

The more of the following conditions an organization meets, the more likely montage is to deliver results.

Condition 1: There is intent to consistently produce video content at scale

If you are producing 5 or fewer videos per month, on an irregular schedule, without a clear purpose, the overhead of building the infrastructure often outweighs the value. The setup only pays off when the state is “we want to consistently publish 20 or more videos per month” and “we have a strategy for growing a channel.”

Condition 2: There are repeatable video formats

If the output is predominantly one-of-a-kind creative videos, there is limited territory for agents to cover reliably. Content like regular earnings recaps, product introduction series, and weekly market reviews — where “a template repeats” — is well-suited to production at scale.

Condition 3: Brand guidelines are articulated (or can be developed)

If the only brand identity that exists is a vague intuition of “something like this,” that intuition must be written down first. What cannot be articulated cannot be communicated to agents.

Condition 4: Someone in-house can handle reviews

A structure must exist for humans to process cases the reviewer agent cannot decide automatically. Organizations without “a human to do a final check on what the AI produces” will find quality management does not function.

Condition 5: The effect of video content is measured (or can be measured)

Accurately measuring ROI requires a mechanism for measuring how video content contributes to the business. Without this measurement, the results of scaling cannot be made visible, and decisions about whether to continue rest on gut feel.

Environments to Develop Before Adoption (Brand Guidelines / Review Structure)

Organizations that assume montage is “plug it in and it runs” tend to fail. There are environments that need to be built before the technology can function.

Brand guideline development: Prohibited expressions, tone definition, visual rules. Without these at minimum, quality gates cannot function. If an existing brand book exists, use it as a base and translate the relevant sections into montage’s configuration. If nothing exists, start by documenting at minimum “a list of words we never use” and “caption font and color rules.”

Review structure design: Who handles reviews, when, and for how many videos? If the production volume is set above what can be reviewed within, say, two hours per week, reviews become a bottleneck and videos cannot be published. Initially, calibrate the production scale to match review capacity, then scale up as the approval rate improves.

Video purpose and KPI definition: What is the mass-produced video content supposed to achieve? Define measurable KPIs first — channel subscribers, view counts, site traffic, inquiry volume. “Just increase the number of videos” provides no basis for verifying investment returns.

A Phased Adoption Roadmap

montage does not need to be fully operational from the start. The following three phases are realistic. Yakumo is currently in the middle of Phase 1, running brand guideline development in parallel with smoke testing while finalizing quality criteria for pilot output.

Phase 1 (months 1–2): Brand guideline development + pilot production ← Yakumo is here

Start by documenting the prohibited expression list, tone definition, and visual rules. Then run a pilot production of 5–10 videos per month and measure automatic approval rate, review labor, and video quality. The goal here is to develop a feel for “the template is taking shape.”

Phase 2 (months 3–6): Scale-up and quality tuning

Use data from the pilot to increase monthly output. Adjust the reviewer agent’s criteria to raise the automatic approval rate. Monitor reviewer labor monthly and address bottlenecks as they appear.

Phase 3 (month 6+): Sustained operation and ROI verification

Measure the impact of the accumulated video content library, verify ROI, and confirm whether break-even has been achieved and whether contributions to view counts and inquiries are materializing. Use findings to make management decisions about continuing, expanding, or changing course.

One note: the biggest blocker to Phase 2 in Yakumo’s montage is building out the distribution pipeline. Automated upload to YouTube / TikTok / X is intentionally blocked via exit 2 in scripts/hooks/pre-publish-gate.sh, and we are separately designing the checklist needed to safely open that gate (pre-publication final review, rollback procedures for erroneous posts, thumbnail quality criteria). The approach we are taking: keep the distribution path blocked in the implementation while running internal pilot production, then open publication all at once once quality has stabilized.

Summary — Decision Checklist for AI-Agent Video Production Adoption

Whether AI-agent video production succeeds or fails comes down to “is it designed as a business?” — not “is it technically possible?”

Use the following checklist to assess your organization’s readiness.

Clarity of Strategy

• Is it clear how many videos you want to produce per month, and for which channels?
• Do you have repeatable video formats with a defined template?
• Is there a defined KPI — what video content is a means of achieving?

Brand Guideline Readiness

• Does a prohibited expression list exist?
• Is a tone definition (formality level, speaking style) documented?
• Are visual rules (colors, fonts, logo) explicitly specified?

Review Structure Design

• Is there someone in-house to handle final approvals?
• Can the weekly review time required be accommodated?
• Are review criteria articulated (what counts as “OK”)?

ROI Calculation Readiness

• Is the current video production cost (in-house and outsourcing actuals) known?
• Can API costs and initial setup costs be estimated?
• Has a break-even calculation been run?

Readiness for Phased Adoption

• Is there willingness to start with a Phase 1 pilot?
• Can this be approached not as “full automation from day one” but as “designing the division of labor with humans”?

The more “yes” answers in this checklist, the more likely montage-based video production is to deliver results. Items with “no” answers can be read as a prioritized list of environments to develop before adoption.

The technical design of montage — the pipeline DAG, inter-agent data flow, video component structure — is covered in detail in montage Pipeline: Technical Design Details (tech cluster sister pillar). Engineers and implementation leads should refer to that article.

Scaling video content is a “high-priority but perpetually deferred” challenge for most organizations. AI-agent-driven production is one path toward solving it. But actually walking that path requires designing it as a business before the technology can run: what to produce at scale, who handles reviews, how quality is defined — this design comes first.

montage is the technical mechanism for executing that design. Without the design, it cannot run. With the design, it is not particularly hard to operate. If this article helps move the first step of that design forward, it has done its job.