Summary: Most thought leadership on AI in marketing is written by people who don’t actually run a marketing team. This isn’t that. We run upGrowth Digital, a 32-person agency serving 150+ clients across SaaS, fintech, D2C, and healthcare. Over the last 18 months we rebuilt our internal workflow three times as AI tools matured. This article walks through what actually changed operationally, which metrics moved, one experiment we abandoned, and what clients now ask us that they didn’t ask two years ago. If you run an agency, lead in-house marketing, or evaluate agencies as a CMO, this is the uncensored version.
When Claude Code launched in 2024 and Claude Design shipped this April, most of the public commentary came from people outside the day-to-day of running a marketing team. The takes sound confident and usually miss what actually happens when you try to bolt AI onto a live services business with payroll to meet and clients expecting deliverables on Monday. At upGrowth Digital, we’ve been in the middle of that rewiring for 18 months. This is what actually happened, with the numbers we can share.
Context first. We’re a 32-person agency. We serve over 150 active clients. Our verticals are SaaS, fintech, D2C, healthcare, and edtech, with a significant presence in Dubai and the broader GCC. Our largest case studies are Lendingkart (5.7x lead volume, 30 percent CPL reduction, 4x spend scaling), Delicut in Dubai (20K AED per month to 2M AED per month), and a stack of mid-market fintech and SaaS brands most people haven’t heard of because we don’t publish every win. We started the AI rebuild in late 2024 and we’re still rebuilding.
What follows is what a competitor agency would want to know and what a CMO evaluating us would want to hear before signing a contract. I’ll cover what changed in the work, what changed in the team, what changed in the pricing conversation, one thing we tried and abandoned, and what’s still hard. If you want the sanitized case study version, this isn’t it.
What Actually Changed in the Work
Three things rewired the production side of the agency. None of them were the obvious “let’s use ChatGPT for content” move that most agencies started with and then stopped at.
The first was the first-draft layer. Starting in late 2024, every piece of content, every ad variant, every reporting narrative passes through an AI-first draft before a human touches it. This sounds trivial. It’s not. Getting the prompts right, the reference material curated, the brand voice training dialed in, and the QA checklist disciplined took about six months of iteration. The first version produced generic content that our senior editors rewrote from scratch, which was slower than pure human work. The current version produces drafts that land at roughly 70 percent of the final quality, and human editors spend their time on the 30 percent that matters (the angle, the nuance, the client-specific insights). Turnaround on a long-form blog article dropped from 8 working days to 3.
The second was the measurement and reporting layer. We built internal agents that pull data from GA4, Google Search Console, Meta, LinkedIn, and client CRMs and produce first-pass performance narratives. The first version was wrong more often than it was right, which taught us that data quality is the binding constraint, not AI capability. We spent four months cleaning data pipelines before the AI layer produced anything useful. Once the pipelines were clean, a weekly client report that used to take an account manager 3 hours to prepare now takes 25 minutes. The manager’s time shifted from data pulling to insight generation.
The third was the client-facing research layer. When a prospect fills out our contact form, our system now runs a structured research pipeline before anyone on our team sees the inquiry. Company research, founder background, recent news, likely pain points, and a preliminary fit score are all compiled in under a minute. The lead response email that reaches the prospect is written with this context baked in, which changes the response rate significantly. We moved from a model where a senior person spent 15 minutes per lead deciding whether to respond to one where that same person spends 2 minutes reviewing an already-drafted response. Our inbound close rate approximately doubled over 12 months, though I can’t fully attribute the lift because we changed other variables at the same time.
We didn’t reduce headcount. We’re still 32 people. What changed is the ratio of what they do.
Two years ago, roughly 70 percent of our team’s time went to production work (writing, designing, building reports, running campaigns) and 30 percent went to what we call “non-production” work (strategy, client calls, internal tooling, relationship work, hiring). Today that ratio is closer to 40 percent production and 60 percent non-production. The production work is still happening, but AI does a larger share of the keystrokes and our team spends more time on judgment, client conversations, and building internal systems that compound.
We also restructured who does what. We used to have a clear hierarchy: junior writers, senior writers, content leads, account managers, strategy leads. That hierarchy still exists on paper, but the work distribution is different. Junior team members now act more like editors and QA operators on AI output than as first-draft producers. Senior team members spend more time on client strategy and less on reviewing junior work. The middle layer (content leads, performance leads) got more strategic because the operational load on them dropped.
Hiring is the hardest part of this transition and I’ll be honest about what we got wrong. We tried hiring “AI-fluent” mid-level marketers for about six months in 2025 and most of them underperformed. The signal we were screening for (comfort with prompting, familiarity with tools) turned out to be a weak predictor of actual performance. What mattered more was taste, judgment, and the ability to critique AI output rigorously. We shifted our hiring to prioritize those traits and deprioritize “AI tool fluency” because the latter is now easy to teach and the former isn’t.
The people who thrive on our current team are ones who are opinionated about quality, willing to disagree with AI output rather than just accept it, and genuinely interested in client business outcomes rather than in craft for its own sake. That profile hasn’t changed much from the pre-AI era, interestingly. Good marketers are still good marketers. The tooling around them changed; the underlying capability didn’t.
What Actually Changed in the Pricing Conversation
Our average retainer size went up, not down, over the last 18 months. That surprises most people.
The reason is that we restructured what the retainer buys. Two years ago a typical mid-market fintech retainer with us was around 4 lakh per month and covered a defined scope of deliverables: a certain number of articles, a certain amount of paid media management, a certain reporting cadence. The client bought outputs. Today a comparable engagement runs closer to 5.5 to 7 lakh per month and the scope is framed around a growth motion rather than outputs. “Scale organic demand capture for our enterprise segment” or “drive qualified SQL volume from LinkedIn” is the scope. The outputs underneath are variable and we optimize them.
The retainer increased because the work underneath got more valuable. Strategy, relationships, measurement, and accountability are all priced higher than content production used to be. We can serve more clients per senior person because AI handles execution leverage, but we charge more per engagement because the work is upstream of production. The clients who’ve made this switch with us (most of our existing roster) haven’t pushed back because they’re getting outcomes they weren’t getting at the old scope. The ones who want to buy outputs at the old rates either went elsewhere or we transitioned them out. That’s fine. Not every client is the right fit for the new model.
We also added performance-tied compensation on about 40 percent of our engagements. Typically 10 to 20 percent of the fee is tied to an outcome metric (ROAS, CPL, organic revenue, qualified pipeline) with a clear definition and a cap. This created internal discipline that flat retainers didn’t. When a client asks us to do work we think won’t move the outcome, we push back harder because our margin is on the line. That friction produces better decisions on both sides.
For about four months in 2025, we tried to build a fully autonomous lead response agent that would handle inbound inquiries end-to-end without a human in the loop. The pitch to ourselves was that we could scale inbound handling to infinity with no marginal cost per lead. We built it, deployed it, and watched it fail in ways that taught us more than the successes did.
The first failure was tone. Even with extensive brand voice training, the fully-autonomous responses had a sameness that prospects picked up on. We started getting feedback like “felt automated” and “didn’t seem to really read my brief.” The responses were technically correct. They just didn’t land the way a human response would.
The second failure was context. The agent made accurate surface-level observations about the prospect’s company but missed the judgment calls that matter in a sales conversation. When to push back on a brief. When to offer something unexpected. When to say no and refer them elsewhere. These are judgment calls AI didn’t handle well in that specific context, and the result was a response pipeline that converted at about 30 percent lower rate than our human-in-the-loop version.
We pulled it after four months and rebuilt the current version, which uses AI for research and drafting but keeps a senior person in the loop for review and personalization. That hybrid model converts better than either pure-human or pure-AI did. The lesson we took from this was specific: AI is great at augmentation and bad at autonomy when the work requires judgment under incomplete information. We now default to augmentation everywhere and only consider autonomy for tasks with clear right answers.
What Clients Now Ask That They Didn’t Ask Two Years Ago
The nature of client conversations shifted in predictable ways as the market caught up to what AI can do.
The first new question we get is “how are you using AI in our workflow.” Two years ago this question was theoretical and clients trusted us to figure it out. Now it’s specific and they want concrete examples. We walk them through the first-draft layer, the measurement layer, the research layer, and we name the tools. Transparency here matters more than it used to, because clients are calibrating their spend based on how much AI leverage they’re paying for. Vague answers cost us deals.
The second new question is “why are you charging more than last year if AI is making this cheaper.” This one required us to reframe how we explain the work. The answer is that the production cost dropped but the strategy and judgment cost didn’t, and we’ve restructured what the fee buys. Clients who understand the distinction stay. The ones who expect a pricing cut because “AI should be cheaper” usually don’t, and we don’t try hard to keep them. The economics don’t work.
The third new question is “can you guarantee revenue outcomes.” This one we still decline. We can guarantee what we control (ad spend, uptime, reporting quality, response time) and we’re willing to tie fees to outcomes we can influence (CPL, ROAS, qualified leads). We don’t guarantee revenue because too much of it sits with the client’s sales team, product, pricing, and timing. Agencies that guarantee revenue outcomes in 2026 are either overpromising or running a different business model than we do. We’d rather lose the occasional deal than commit to something we can’t deliver.
The fourth new question, and this one is interesting, is “who on your team is actually making the decisions.” Clients are increasingly aware that the human judgment layer is where the value sits, and they want to know the specific people whose thinking they’re renting. This has made senior team visibility more important in the sales process. The partner or strategy lead who shows up on the first call needs to be the same person who’ll be present through the engagement. Bait-and-switch staffing (senior person sells, junior person delivers) doesn’t work anymore because clients can tell the difference.
What’s Still Hard
I want to be honest about what hasn’t been solved yet, because the triumphant “we rebuilt the agency with AI” narrative has a lot of survivorship bias in it.
Onboarding new clients is still hard. Getting a new client’s data, brand voice, positioning, and operational context loaded into our AI pipelines takes between four and eight weeks, and during that window the AI leverage doesn’t work well. We’ve shortened it from six months originally, but onboarding drag is real and we haven’t found a clean solution. The first month with a new client is still mostly human-intensive.
Data quality is a persistent issue. About 30 percent of the time we onboard a client, their analytics are misconfigured in ways that make AI-layered measurement less reliable. Cleaning this takes time that clients don’t want to pay for but the measurement work depends on. We’ve started including a paid data audit as a prerequisite for most engagements, which some prospects reject. That’s a trade we’ve made peace with.
Model and tool changes break things constantly. When a major AI tool updates its API or shifts its quality, some percentage of our internal workflows break or degrade. We’ve built a small internal QA team whose job is to monitor for these shifts and update the affected workflows. This is ongoing cost that doesn’t show up in the marketing narrative about AI efficiency.
Hiring senior talent is harder than it was. The pool of people who have both strategic judgment and comfort with AI-augmented workflows is small and getting competed over by in-house teams and well-funded startups. We’re paying more for senior hires than we were, which shows up in our margin. The alternative is not hiring, which shows up in our capacity. We chose the higher cost.
From conversations with other agency founders and operators, there are three common patterns I’d call out.
The first is layering AI on top of unchanged workflows. An agency buys ChatGPT licenses for the team, runs a few prompt engineering workshops, and expects efficiency gains. The gains don’t materialize because the underlying workflow wasn’t designed around AI. This produces the worst of both worlds: you pay for the tools and your team is slower than before because they’re context-switching between old processes and new. The only way this works is a from-scratch workflow rebuild, which is disruptive and expensive, and most agencies aren’t willing to do it.
The second is pricing the output lower instead of pricing the work differently. Agencies in this bucket reduce their rates because they think AI makes the work cheaper, so they cut prices 20 percent to stay competitive. This doesn’t work because it misdiagnoses the value. The production cost dropped but the strategy cost didn’t. Price cuts commoditize the agency and compress margin to unsustainable levels. The move is to restructure what the fee buys, not to cut the fee for the same thing.
The third is ignoring the measurement and accountability layer entirely. Most agencies we see pitching AI integration talk about content production and maybe creative production. Very few talk about how AI has changed the measurement and reporting cadence, or how they’re using AI to tie work to outcomes. This is where the most durable value lives and where most of the industry is still behind. The agency that gets measurement and accountability right has a moat the content production agencies don’t.
Five Common Questions About Running an AI-Augmented Agency
Q: How long did the workflow rebuild take?
A: About 18 months from start to current state, and we’re still rebuilding. The first three months produced no visible gains because we were redesigning from scratch. Months four through twelve produced inconsistent gains as we iterated. Months thirteen through eighteen is when the compounding started. Most agencies stop at month three because the gains haven’t shown up yet, which is the exact wrong time to stop.
Q: What’s your margin now versus two years ago?
A: Margin went up modestly, roughly 5 to 8 percentage points, despite higher senior compensation. The lift came from serving more clients per senior person (leverage on the senior layer) and from performance-tied compensation kicking in on well-executed engagements. The increase wasn’t from reducing headcount or cutting costs aggressively.
Q: What AI tools do you actually use day-to-day?
A: The stack changes quarterly. Current anchors are Claude for most drafting and reasoning, custom Apps Script pipelines for lead handling and reporting, internal skill libraries that orchestrate Claude for specific workflows, and standard measurement tools like GA4, GSC, and Meta that we’ve wrapped with AI-layered reporting. Tool selection is less important than workflow design. An agency using best-in-class tools with a bad workflow will lose to an agency using adequate tools with a great workflow.
Q: Would you recommend in-house marketing teams try to replicate this?
A: For companies above about 30M USD in revenue with a real marketing team, yes, partially. The first-draft layer and the reporting layer are feasible internally with 6 to 12 months of focused effort. The research layer is harder because it requires external data pipelines. For companies below that size, the ROI on building internal AI infrastructure usually doesn’t work and an AI-native operator agency is the better path. The math changes at scale.
Q: What’s the single biggest mistake other agencies make when trying to rebuild around AI?
A: Treating it as a tool problem rather than a workflow problem. Most agencies buy AI tools and keep their workflow, then are disappointed when gains are modest. The tools are commoditized. The workflow is the moat. Rebuilding the workflow is disruptive, takes 12 to 18 months, requires pissing off some clients and team members along the way, and produces outsized gains if you stay the course. Most don’t stay the course.
Your Next Move: Decide Where You Sit and What to Do About It
If you run an agency, the question is whether to start the workflow rebuild now or keep pretending AI is just a tool to add to your stack. The rebuild is hard and the payoff is 12 to 18 months out. Not rebuilding means slowly losing competitive position to agencies that did. There’s no middle path that actually works, as far as I’ve seen.
If you’re a CMO or founder evaluating agencies, the question is whether your current agency has done this rebuild or is still layering AI on top of unchanged workflows. The diagnostic questions in the “Three Signs Your Agency Contract Is Obsolete” section of our previous article are a useful filter. The Agency Fit Score diagnostic is a faster one.
If you want to have a specific conversation about what rebuilding a marketing function around AI looks like for your company, we run a 30-minute operator conversation for qualifying mid-market companies. No deck, no pitch, just a working session on where your marketing function sits and where it should go. Book your operator conversation here.
About the Author: I’m Amol Ghemud, Chief Growth Officer at upGrowth Digital. We help SaaS, fintech, and D2C companies shift from traditional SEO to Generative Engine Optimization. This shift has generated 5.7x lead volume increases for clients like Lendingkart and 287% revenue growth for Vance.
For Curious Minds
Operational rewiring means systematically embedding AI into core processes, not just using it as a standalone tool. This shifts human effort from manual execution to strategic oversight and quality control. For upGrowth Digital, this involved creating three distinct AI-powered layers that transformed how the team works.
This new structure includes:
An AI-first draft layer for all content and ad copy, which reduced the turnaround time for a long-form article from 8 working days to 3.
A measurement and reporting layer that automates data pulling and initial analysis, freeing up account managers for higher-level insight generation.
A client-facing research layer that instantly vets and enriches new leads before human interaction.
Instead of writing a blog post from scratch, a team member now refines a 70% complete draft. This focus on strategic refinement rather than manual creation is the core change, allowing senior talent to apply their expertise where it matters most. You can explore the full details of this transformation in the complete article.
The strategic value of an AI reporting layer is its ability to convert raw data into actionable narratives at scale, shifting an account manager's time from assembly to analysis. A report that previously took 3 hours to prepare now takes 25 minutes at upGrowth Digital, which directly translates to more time spent on strategy. However, this is only possible if the underlying data is clean and reliable.
The agency learned that data quality is the absolute binding constraint. An AI agent is only as good as the data it processes; feeding it inconsistent or inaccurate information from GA4, Meta, or client CRMs produces flawed and unusable narratives. They spent four months dedicated to cleaning their data pipelines before the AI layer became effective. This foundational work ensures that the generated reports are trustworthy, making the shift from data pulling to insight generation a reality. The full story details the challenges of this data-first approach.
The difference lies in a systemic approach versus a tactical one. Simply using a chatbot for a blog post often results in generic content, whereas upGrowth Digital built a disciplined, multi-step workflow that treats AI as a component within a larger quality-controlled process. This elevates the output from a simple draft to a strategic asset.
The key differentiators in their process include:
Curated Reference Material: The AI is fed with specific, high-quality inputs, not just a generic prompt.
Brand Voice Training: Extensive work is done to dial in the AI's output to match each client's unique tone.
Disciplined QA Checklists: Human editors use structured checklists to refine the final 30% of the work, focusing on nuance and client-specific insights.
This systemic integration versus tool-based usage leads to consistently higher quality work that performs better, directly impacting metrics like lead volume and cost per lead. Learn more about their six-month iterative journey to perfect this system in the main article.
AI-powered workflows directly accelerate the core loop of digital advertising: testing, learning, and optimizing. The impressive results for Lendingkart, including the 30 percent CPL reduction, were achieved by using AI to dramatically increase the scale and speed of this process. It is not about a single AI tool but a combination of integrated systems.
The first-draft layer allows the team to generate a massive volume of ad variants, exploring different angles, headlines, and calls-to-action far faster than a human team could alone. Simultaneously, the measurement and reporting layer continuously analyzes performance data from these variants, quickly identifying winners and losers. This enables a strategy of rapid iteration and optimization, where the budget is reallocated to top-performing ads almost in real-time. This combination of scaled creation and accelerated analysis is what drives significant metric improvements. The full piece explores this link between operational speed and client results.
Supporting the explosive growth of a D2C brand like Delicut requires an operational capacity that can keep pace with rapidly increasing ad spend. AI-driven workflows are essential for maintaining quality and agility during such a scale-up, particularly in a competitive market like Dubai. The agency's success hinges on two key areas enhanced by AI.
First, the client-facing research layer can be adapted for market intelligence, quickly analyzing competitor strategies and identifying new audience segments. Second, the first-draft layer for content and ads is crucial for scaling creative production to match media spend. As the budget grew from 20K to 2M AED, the demand for fresh ad creatives, landing pages, and social content would have increased exponentially. AI allows the agency to meet this demand without a proportional increase in headcount, ensuring campaigns stay fresh and effective. The full article provides more context on managing high-growth accounts.
This efficiency gain stems from a custom-built system that automates the entire data-to-narrative process, fundamentally changing the account manager's role. The time saved is not just about cost but about reallocating senior talent to what humans do best: strategic thinking. This system is composed of several key parts.
The workflow begins with internal agents that automatically pull data from multiple sources like GA4, Google Search Console, Meta, and client CRMs into a centralized location. A critical second step involves cleaning and standardizing this data to ensure accuracy. Finally, a proprietary AI model processes the clean data to produce a first-pass performance narrative. The account manager's job transforms from data pulling to strategic analysis and insight generation, as they now start with a nearly complete report and spend their time interpreting the results and planning next steps. The full post explains how this shift improves client relationships and outcomes.
An in-house SaaS marketing team can implement an AI-first draft layer by starting small and building a rigorous human-in-the-loop process. The goal is to augment your existing team, not replace it, which requires a structured rollout to maintain quality and secure buy-in from your writers and editors. A practical plan follows a pilot, train, and validate model.
Here are three actionable steps:
Step 1: Pilot with a Single Content Type. Choose one specific format, like blog posts or ad copy, to test the workflow. This limits the scope and allows you to refine the process before a full rollout.
Step 2: Develop Rigorous Prompts and Brand Guides. Create detailed documentation on your brand voice, audience personas, and desired content structure for the AI. This is the most critical step for achieving high-quality drafts.
Step 3: Implement a Human-Led QA Checklist. Arm your editors with a specific checklist that guides them to focus on adding nuance, client-specific insights, and strategic angles—the 30% that AI struggles with.
This methodical approach helps you build a reliable system over time. Discover more implementation details in the complete article.
Building an automated research pipeline gives your sales team a significant advantage by equipping them with crucial context moments after a lead comes in. This requires stitching together a few different technologies into a cohesive workflow that turns a simple contact inquiry into a rich strategic brief. The process is about creating a structured research pipeline.
The key steps are:
Trigger and Enrich: The workflow is triggered by a new form submission. The system then uses APIs to pull data from public sources like the company's website, LinkedIn for founder backgrounds, and recent news articles.
Synthesize with AI: The collected raw data is fed into a large language model with a specific prompt instructing it to synthesize the information into a structured summary. This summary should include likely pain points based on their industry and a preliminary fit score based on your ideal client profile.
Deliver the Brief: The final, synthesized brief is delivered directly to the appropriate team member's inbox or CRM, allowing for a highly informed and rapid first contact.
This system transforms your lead response from reactive to proactive. The full article touches on how this front-end efficiency connects to later client success.
The role of creative professionals will pivot from pure production to strategic direction and curation. As generative AI tools like Claude Design become more capable, the value of raw output will decrease, while the value of high-level judgment, taste, and brand understanding will increase dramatically. Creatives will become more like creative directors for AI systems.
To stay ahead, they must prioritize developing a new set of skills beyond their core craft. The most critical capabilities will be:
Creative Strategy: Understanding the client's business goals and translating them into a creative brief that an AI can execute.
Advanced Prompt Engineering: Mastering the art of communicating with AI to elicit nuanced, on-brand, and effective outputs.
Critical Curation and Editing: Possessing the discerning eye to select the best AI-generated options and perform the crucial final 10-30% of refinement that elevates the work.
Their future value lies in guiding and refining AI, not competing with it. The full text offers more perspective on the future of the agency team structure.
The conversation is shifting away from hours and deliverables towards performance and outcomes. As clients understand that AI can produce a first draft of a blog post in minutes, they will be less willing to pay hourly rates for tasks they perceive as automated. Agencies must now justify their fees based on the strategic value they add.
This means the value proposition must evolve to emphasize what humans, augmented by AI, can do better. For example, instead of selling "four blog posts per month," an agency should sell "a content strategy proven to reduce CPL by 30%," citing examples like the Lendingkart case study. The new client expectation is for value-based pricing and outcome-driven partnerships, where the agency is compensated for its strategic impact, not its labor. Agencies that cannot demonstrate a clear link between their work and client business goals will struggle to compete. The full article explores this shift in client dynamics.
The primary pitfall is treating AI as a magic box that works perfectly out of the gate, when in reality it is a powerful tool that requires a sophisticated operator and a well-defined process. upGrowth Digital encountered common failures stemming from weak prompts, a lack of contextual brand information, and, most importantly, poor data quality. These issues led to generic content and inaccurate reports.
They overcame these challenges by focusing on process over prompts. The solutions were not found in a better AI model, but in the human-led systems built around it. They implemented several key QA measures:
Meticulous Data Cleaning: They spent four months cleaning data pipelines before their reporting AI produced anything useful.
Curated Reference Materials: The AI was trained on specific brand voice documents and high-quality examples.
Disciplined QA Checklists: Human editors were given a structured framework to guide their review, ensuring the final 30% of the work added critical nuance and insight.
This demonstrates that successful AI integration is a matter of operational discipline. The full piece shares more on the experiments they had to abandon.
A mature operational approach integrates AI into a repeatable, multi-step workflow with strong human oversight, rather than using a tool like ChatGPT on an ad-hoc basis. The goal is not just to create a single piece of content but to build a reliable system that consistently produces high-quality assets at scale. This requires significant upfront investment in process design.
The six-month iteration period for upGrowth Digital was necessary to solve the complex challenge of achieving consistent quality. During this time, the team focused on dialing in prompts, curating reference materials, and refining the human QA process. The initial versions produced generic content that was slower to fix than writing from scratch. It took months of trial and error to build a system that could reliably produce drafts at 70% of final quality, which is the threshold where true efficiency gains are realized. Learn about this iterative journey in the full article.
Growth Strategy and Planning
Inbound Growth
Growth Hacking
Fractional CMO
Search Engine Optimization
Paid and Performance Marketing
Social Media Marketing
Demand Generation
AI-Driven Growth Strategy
AI-Native Workflow Automation
Generative Engine Optimization
Generative AI Content Creation
AI Automation
Framework
Case Study
Tools
Growth Tools
Decision Tree
Compare
Custom GPT's
Guides
Offer
Webstories
Quizzes
Blog
AI-powered Apps
Social Media Tools
