If we’re going to ask clients to invest in AI capability, we should be at least as far along as we’re asking them to go. So here’s what AI looks like inside our own operations — the honest version, including what’s working and what isn’t.
The setup
Frogslayer is a team on the order of a few dozen people. A mix of senior strategists, engineers, build leads, designers, and operations staff, distributed across Texas with our HQ in Bryan-College Station. Revenue is growing materially with the AI Office product line; profitability is stable.
We’ve been seriously deploying AI inside our own operations for about 18 months. This is what it looks like today.
What AI does for us today
Going through our actual workflow, AI is meaningfully embedded in seven places.
1. Proposal and Discovery Letter drafting. When a prospect asks for a proposal, AI drafts the Discovery Letter — pulling from the discovery call notes, the prospect’s public information, our case study library, and our standard pricing. A senior strategist reviews and edits. That’s typically 20-30 minutes of editing on a draft that would have taken 2-3 hours to build from scratch.
2. Client briefings. Every Monday, our senior strategists get an AI-generated briefing on their client portfolio: what shipped last week, what’s at risk, what decisions are pending, what’s on the agenda for the week’s standing calls. The strategist reviews it in 10 minutes and walks into the week prepared.
3. Engineering documentation. Our build engineers use AI heavily during builds — for code generation, for documentation, for explaining their work to non-engineers. Documentation that used to lag the build now keeps pace, because AI makes it cheap.
4. Knowledge base management. Our internal knowledge base — case studies, patterns, playbooks, technical guides — is AI-searchable and AI-citable. New hires get to productive contribution faster because they can find precedent without interrupting a senior person.
5. Quarterly business review prep. I get an AI-drafted version of our quarterly review every quarter: financial summary, pipeline summary, client portfolio health, operational metrics, narrative. I review and edit, and the leadership team gets a better deck faster than we used to produce.
6. Operations and finance. Our CFO uses AI heavily for variance analysis, forecasting, and the narrative work around month-end and quarter-end. The close is faster. The narrative is sharper.
7. Marketing and content. A lot of what’s published on frogslayer.com is AI-drafted by our marketing team and edited by senior people who know the substance. The cadence we maintain — 4-6 pieces of substantial content per month — wouldn’t be possible without AI.
What we tried that didn’t work
In the interest of honest reporting, here’s what we got wrong.
We tried an autonomous lead-qualification agent. The idea: an AI agent reads inbound inquiries and decides whether to disqualify or route to a strategist. It produced too many false negatives — disqualifying real prospects on signals that didn’t actually predict fit. We pulled it after 6 weeks and rebuilt it as a workflow where AI drafts a fit assessment and a senior strategist makes the routing decision in 30 seconds. Better outcome.
We tried AI-drafted client weekly briefings. The idea: clients get an AI-drafted summary of their account every Friday. We launched it. Some clients loved it; some hated the format; some felt the substance was thin. We pulled it after 4 months because the variance was too high. We replaced it with strategist-written briefings that use AI as a tool for the strategist, not as a customer-facing AI workflow. Better outcome.
We tried to automate the discovery call summary. The idea: AI listens to the call recording, drafts the summary, and builds the candidate workflows. The summary part worked. The candidate-workflows part didn’t — the AI surfaced surface-level patterns that missed the texture senior strategists pick up on. We kept the summary and abandoned the workflow drafting.
These are the kinds of failures every serious AI operator has. We share them because the success stories are more impressive when the failures are honest.
What this actually costs us
The figures below are illustrative, rounded estimates of our own internal spend — directional, not audited line items. Our internal AI spend in a given year lands roughly in these ranges:
- Tools and licenses: on the order of $40K-$50K (Claude for Work, ChatGPT Enterprise, a few specialty tools)
- Internal build effort: roughly $75K-$85K equivalent (our own engineers building our own workflows, partially documented as opportunity cost)
- External tooling for marketing/SEO/AEO infrastructure: on the order of $25K-$35K
Total: somewhere around $150K-$160K in direct cost, give or take. Indirect cost: roughly 12-15% of one operations leader’s time, plus meaningful working-session time from senior people.
What it’s produced (again, our own internal estimates, not audited):
- On the order of several hundred thousand dollars of internal labor savings — work senior people don’t have to do
- Materially better content output, which drives marketing pipeline
- A faster, more accurate discovery-to-proposal cycle
- A more knowledgeable team, because the knowledge base is alive instead of dead
Our rough internal estimate of return: on the order of 2-3X in year 1 of our own serious deployment, compounding as we extend. Treat that as a directional figure, not a guarantee.
The honest meta-point
Eating our own cooking isn’t just marketing. It’s how we know what works.
When a client asks “how should we handle X workflow,” the answer is shaped by having tried similar workflows internally. When a client says “we want to try an autonomous agent for Y,” we can speak from the experience of having tried autonomous agents and learned where they break.
If we weren’t running this in our own operations, our advice would be theoretical. Theory is fine. Experience is better.
What we’d do differently if we started today
In honest hindsight, three things we’d change if we were starting our own AI deployment today.
We’d start the internal knowledge base earlier. We started it in mid-2025; it should have been early 2024. The compounding value of a structured internal knowledge base is so high that the cost of starting late is meaningful.
We’d hire the internal AI Operations Lead earlier. We promoted ours in late 2025; it should have been 2024. The job exists at scale earlier than we recognized.
We’d be more disciplined about retiring workflows. We let a couple of marginal workflows linger longer than they should have. The discipline of killing workflows that aren’t earning their keep should be installed from day one.
These aren’t fatal mistakes. They cost us 6-12 months of compounding value on a few specific dimensions. The cumulative impact is real.
Why this matters for you
If you’re considering working with us, I want you to know that we operate the way we ask you to operate. The cadences, the discipline, the measurement, the workflows — we run them inside our own business before we ask you to run them inside yours.
That doesn’t make us right about everything. We’ve gotten things wrong. We’ll get more things wrong. But the gap between “we say to do this” and “we do this ourselves” is small. That’s a fair test for any AI partner.
If you’re interviewing AI partners and you want to ask whether they’re using AI seriously in their own operations — ask. You’ll learn a lot from the answers.
— Ross