How One SaaS Company Replaced $300K/Yr in Agency Fees With AI

The Quiet Shift in the Outsourcing Economy

For the past decade, the standard playbook for mid-market SaaS companies was simple: hire agencies. Copywriting agencies for landing pages and blog content. Design agencies for brand refreshes. SEO agencies to handle keyword research and link building. Marketing automation agencies to wire up email sequences and retargeting funnels. Each one billed $20K to $60K per year, and together they added up to a line item most CFOs had learned to treat as unavoidable overhead.

At Northbeam Labs, a B2B SaaS company selling a customer analytics platform out of Austin, that line item had grown to roughly $300K annually across four agencies by 2023. The company had 47 employees, $12M in ARR, and a marketing budget that felt permanently on fire. The CEO, a former engineer named Dana Kowalski, kept circling the same question in leadership meetings: "Why are we paying external experts to do work our software should be able to handle?"

That question, framed differently, was being asked in boardrooms across the SaaS landscape in 2023 and 2024. Not as a theoretical exercise, but as an urgent cost-optimization strategy. The answer, for companies willing to invest in internal capability rather than external headcount, has been surprisingly fast.

What the $300K Was Actually Buying

Before dissecting the replacement, it matters to understand what those agency fees were covering. At Northbeam, the four agencies handled:

• Content & Copywriting ($85K/yr): Blog posts, case studies, email copy, landing page rewrites, and customer-facing documentation updates.

• SEO & Technical Optimization ($70K/yr): Keyword research, on-page optimization, technical audits, and content clustering strategy.

• Design & Brand ($65K/yr): UI microcopy reviews, template design, social media graphics, and annual brand refresh elements.

• Marketing Operations & Automation ($80K/yr): HubSpot workflow management, A/B test design, attribution modeling, and campaign performance reporting.

The total represented roughly 12% of Northbeam's revenue. In the SaaS world, healthy marketing spend typically lands between 10% and 20% of revenue, so the company was at the upper end of the range — a sign they were over-leveraged on external help relative to their internal sophistication.

The Internal Build: A 90-Day Plan

Dana and her COO, Raj Mehta, mapped the work agencies were doing and identified which tasks were genuinely creative or required judgment versus which were pattern-based and automatable. The split was more lopsided than expected. Roughly 70% of the agency work was pattern-based: producing content from templates, optimizing pages against known rules, generating design variations from brand guidelines, and configuring automation workflows.

The remaining 30% required a human-in-the-loop approach — senior-level editorial judgment on flagship pieces, strategic brand decisions, and complex attribution modeling. That 30% was manageable with one experienced in-house marketing lead, a role Northbeam already had.

The 70% was the AI opportunity.

The 90-day build looked like this:

Days 1–30: Prompt Engineering & Knowledge Base Construction

Northbeam's marketing ops team fed their entire brand voice guide, past content library, product documentation, and customer interview transcripts into a fine-tuned language model. They built a set of 34 specialized prompts covering everything from "write a case study in our voice using these interview notes" to "optimize this landing page headline against these 12 conversion-tested patterns." They versioned the prompts, A/B tested outputs internally, and iterated weekly.

Days 31–60: Automation Pipeline Integration

The prompts were embedded into Northbeam's existing tech stack. Content briefs triggered in their project management tool now auto-generated first drafts in Notion. SEO briefs from their keyword tool auto-populated structured data suggestions and internal linking recommendations. Design requests in Figma triggered a custom plugin that generated on-brand variations. Marketing automation workflows in HubSpot were rebuilt using AI-generated copy blocks that marketing ops could approve or tweak in one click.

Days 61–90: Human-in-the-Loop Review Layer

A single marketing ops manager (FTE cost: ~$115K/year, or about 30% less than the agency spend) became the quality gate. She reviewed AI outputs against a rubric, caught hallucinations or brand-voice drift, and approved the final package for publication. The average turnaround for a blog post went from 14 days (agency brief → draft → revision → approval → publish) to 4 days (internal brief → AI draft → 2-hour review → publish).

The Numbers, Eighteen Months In

By month 18 after the transition, Northbeam's marketing spend looked dramatically different:

The net savings were $150K annually — not the full $300K, because Northbeam hired a dedicated in-house ops manager to own the AI pipeline. But the output volume had actually increased. Blog content output went from 8 posts/month to 22. Landing page A/B tests doubled. Email sequence coverage went from 60% of user journeys to 94%.

The quality debate is the one worth having. Northbeam's content engagement metrics (average scroll depth, time on page, CTR) held steady or improved slightly. Their blog's organic traffic grew 34% year-over-year, outpacing the agency era's 18%. The design output was rated "9/10 vs. agency 9.5/10" in a blind internal review — a negligible gap for a B2B product where design is functional rather than aspirational.

What Made It Work (And What Almost Broke It)

Three factors determined success more than the AI model choice:

1. Brand Voice as a System, Not a Vibe.

Northbeam spent two weeks in the build phase writing a 12-page "voice constitution" — specific diction rules, sentence length targets, banned phrases, and 40 annotated examples of "exactly right" vs. "close but off." Without that artifact, AI output drifted into generic SaaS marketing paste within two weeks.

2. The Review Layer Was Never Optional.

Every AI-generated asset passed through a human reviewer. Dana's instinct to "just trust the model" was overridden by Raj's pushback. The reviewer caught a product claim in month 3 that would have been a legal liability, and a pricing error in month 5 that would have cost them a deal. The model was confident both times. The reviewer was not.

3. Prompt Infrastructure Treated as Code.

The prompts were version-controlled in Git, tested against a regression suite of 200 edge-case inputs, and deployed through CI/CD. When the underlying model was updated by the provider, outputs were re-tested before the change went live. This engineering discipline, borrowed from Northbeam's own product team, was the difference between a reliable system and a toy.

What almost broke it: In month 4, the team tried to scale the pipeline to handle 5x the content volume by reducing review steps. Output quality dropped, two pieces were pulled post-publication for factual errors, and the brand took a small hit in a key industry newsletter. They walked it back. The 90-day build, they learned, wasn't the finish line — it was the minimum viable system that required ongoing tuning.

The Broader Pattern

Northbeam isn't an outlier. Across mid-market SaaS companies that have made the shift — from 100-employee startups to $200M ARR public companies — the pattern is consistent:

• Content production costs drop 60–80% when AI handles first-draft generation and pattern-based optimization.

• Output volume increases 2–4x at the same headcount, not less.

• The human role shifts from "producer" to "editor and strategist," which changes the seniority profile of the marketing team.

• The savings rarely equal 100% of the agency spend because some quality gate and strategic layer must remain human. Net savings of 40–60% of prior agency spend is the typical landing zone.

• The build takes 60–120 days for a mid-market company with a tech-savvy ops team. Companies without that internal capability often hire a fractional AI consultant for the initial build, a cost of $25K–$40K that pays for itself within two quarters.

The agencies weren't doing bad work. The economics of their model — billable hours multiplied by a human knowledge premium — simply collided with a technology that commoditized the pattern-recognition layer of the work. The companies that adapted fastest weren't the ones that fired their agencies overnight. They were the ones who spent 90 days building a system good enough to make the agency relationship optional, then let the data decide.

At Northbeam, Dana put it bluntly in an all-hands: "We didn't replace the agencies with robots. We replaced the agencies with a process that happens to be 80% automated. The humans are still the ones making the calls. They just stopped typing the third draft of a blog post at 2 PM on a Tuesday."

That Tuesday-afternoon typing, multiplied across four agencies and three years, is where the $300K went.