AI in your business: George and Ollie show what's actually running in theirs

George Sanderson (Honest Dog Co.) and Ollie Groombridge (8hours) · 27 May 2026 · Venice All Hands

Full recording:

Overview

Two Venice members walked through exactly what’s running in their businesses. George built a self-serve dashboard that pulls every channel (Meta, Shopify, Amazon, TikTok Shop, 3PL, subscriptions) into a Google Sheet refreshed every fifteen minutes, plus a daily 8am Slack briefing with commentary. Ollie has a persistent agent named Jeff running on a Mac Mini at home — broad API access into Notion, Shopify, Gorgias, Slack, Gmail, and the company bank — that handles brief follow-ups, market setup, CS coaching, P&L tracking, and new-hire onboarding. The session’s strongest signal: this isn’t theory. Both are in production today, built in hours not months, costing hundreds of dollars not thousands.

Five takeaways:

1. You can replace a Triple Whale–class analytics stack with a Google Sheet, Railway, and Claude Code in about four hours. Stack cost: ~$205/month, all in.
2. The daily 8am Slack briefing — numbers plus commentary — is what shifts a dashboard from “I check it” to “it tells me what changed.”
3. You can trust the Dashboard 99%, but you apply some skepticism to the LLM commentary. Railway pulls directly from APIs into the sheet, advice generated on top of it can sometimes be funky.
4. An always-on agent on separate hardware unlocks the “remote employee” use case. You can send your agent a plain-text SMS from a plane, scope the API permissions, no laptops overheating during calls.
5. Use agents as a private coaching layer for your human CS rep — top coaching points each week, in DM.

Part 1 — George’s Honest Dog dashboard

One sheet pulling every channel

George’s reporting starts from a single Google Sheet that updates every fifteen minutes. It pulls revenue, net profit, ad spend, new-customer CAC, AOV, margin, blended ROAS, new-customer ROAS, and new-customer subscription uptake — blended across Meta, Shopify, Amazon (via SellerBoard), TikTok Shop, Mintsoft (his 3PL), and Loop Subscriptions. COGS and shipping rates live in the sheet. Before this, he was juggling SellerBoard, Shopify reports, and platform logins daily; now everything sits in one place that he checks “pretty religiously throughout the day.”

The 8am Slack briefing

The piece George calls a game-changer is a daily Slack message at 8am — a “CMO/CFO briefing from the previous day.” Net profit across channels, total value created (including future subscription LTV), CAC vs. ceiling, subscription uptake vs. target, and product-level commentary on the previous day’s movers. In the example he showed, chicken bone broth was up 46% D2C, goat milk had a 51% D2C bump, and subscription rate had dropped six percentage points — flagged as “potentially due to lower acquisition costs.” The brief also surfaces churn reasons (top one: “already having enough stock”) and supply-chain warnings.

“It’s just been such a game changer for me in terms of pulling data from various places and having actual commentary on it as well.”

— George Sanderson, 00:03:57

Why off-the-shelf tools didn’t fit

George had tried many existing analytics platforms first and the consistent gap was blending Amazon and Shopify into a single view. Channel-specific tools could go deep on one, but couldn’t tell him how the two played together. Building his own meant 10–15 minutes in Claude Code to add a new metric, or to drop in COGS for a new SKU.

The Amazon halo from Meta spend

For Honest Dog “around 40%” of Meta spend flows over to Amazon revenue. Amazon is roughly 40% of revenue overall, so accounting for that halo materially changes what he can afford to spend on Meta. No third-party tool he’d used surfaced that relationship cleanly.

Hallucination risk on the numbers

George’s split is important. He has “99% confidence” in the dashboard numbers because Claude isn’t in the loop after build — Railway pulls directly from APIs into the sheet. Where to apply skepticism is the daily commentary, “when it’s actually given advice” — that side can sometimes be “a little bit funky.”

The build

George has zero coding background. He built everything in Claude Code (he uses the highest tier, ~$200/month) in about four hours one afternoon. Railway runs the fifteen-minute cron that pulls from each platform’s API into the Google Sheet. He took the chat he’d used to build it and asked Claude to write a reusable prompt that other members could start from — you can see that here.

Part 2 — Ollie’s agent, Jeff

Agents as the world’s smartest intern

Ollie’s frame: Jeff is “another employee on your team — the world’s smartest intern” who can take action across every tool in the stack.

“The single biggest limitation to the use of agents is my little human brain, and figuring out ways to use it.”

— Ollie Groombridge, 00:21:36

The setup: OpenClaw on a Mac Mini at home

The agentic layer is OpenClaw. Jeff lives on a screenless Mac Mini at Ollie’s house, plugged into a 150-Mbps home connection, always on. The Claude LLM behind it is Sonnet 4.6 (more on cost below). Three reasons for the separate device:

1. Safety. The agent can take destructive actions; you don’t want your passport on the same disk. API-key permissions let you scope read-only vs. write (Gmail read vs. send).
2. Background load. Running an agent on the laptop you’re also using competes with calls and overheats the machine.
3. Always-on remote access. On a recent ten-hour flight to the US, with just enough Wi-Fi for a plain-text message, Ollie was queuing tasks to Jeff — who was running them at home speed from the Mac Mini.

Memory: how Jeff doesn’t start fresh every conversation

The technical piece Ollie spent the most time on is memory architecture. The problem he’s solving: each conversation with a stock LLM starts fresh. Jeff has five layers:

• memory.md — long-term curated. Business context, key decisions, authoritative data, rules (ICP, unit economics, internal access rules).
• Daily memory logs — the nitty-gritty of what’s happened day-to-day.
• soul.md — who you are, how to behave. Ollie says Jeff “came pretty well-behaved and polite already,” so this stays light.
• tools.md — every API key so Jeff never forgets a connection.
• Agent operating rules — group-chat behavior, who he can DM, etc.

Coaching the CS rep instead of replacing her

Jeff reads the last three days of Gorgias tickets and writes a private coaching brief for Patricia, who runs CS from the Philippines. Five coaching points each week, every one tied to a real ticket from her inbox.

“There is no email sent by AI in our customer service. It’s purely the coaching layer.”

— Ollie Groombridge, 00:44:00

Onboarding a new COO via Jeff

A new COO recently joined 8hours and was given Slack access to Jeff. Within days, the COO was pulling COGS, the VAT number, shipping locations, 3PL addresses, and most of the questions that would have come straight to Ollie. The framing he uses: “There’s not much human connection in a question about VAT.”

Pick the model that fits the task

Ollie runs Jeff on Sonnet 4.6. Opus 4.7 would be roughly 4× the cost, and most of the tasks Jeff handles — pulling Notion records, drafting follow-up messages, reading Gorgias tickets — don’t need the strongest model. You can also set hard spend limits in Anthropic’s console; the agent will stop and ask before it overshoots.

Q&A highlights

Q (Gaia, Founder of Tootilab): How confident are you the numbers aren’t hallucinated?

George puts 99% confidence on the dashboard figures because Claude isn’t in the loop post-build — Railway pulls directly from APIs into the sheet. The place to be suspicious is the daily Slack commentary layer; when the model is generating advice on top of the numbers, it can be funky.

Q (Nicolas, Co-Founder of Wegter Living): Does AI ever take over from a human CS rep?

No customer-facing email is sent by AI at 8hours. Ollie tried other platforms and CS-specific AI services (not OpenClaw itself); the nuance in customer service kept producing wrong answers. So Jeff coaches Patricia privately each week and she sends every reply herself.

Q (Maarten, Founder of Maison 365): Why a Mac Mini at home instead of a cloud VM?

Ollie’s honest answer: contained hardware at home feels safer to him personally. The agent can only access files on that machine, and he tightly controls API scopes.

Q (Maarten, Founder of Maison 365): What does Jeff cost you per month?

Mac Mini up front: $800–$900. API tokens: $300–$400/month. The bigger lesson Ollie added: he ran into days where costs spiked because he hadn’t picked the right model — switching to Sonnet 4.6 from a heavier default brought the run-rate back in line.

Q (Davey, Founder of Zippit): How long did it take to train Jeff?

Ollie added Jeff to Slack and told him to read every channel and Gmail thread from the previous three months. “If you were to go through and read the last three months of messages on your Slack and Gmail, you’d get a pretty solid idea about what’s going on in the business.” From there, Ollie coaches as edge cases come up; Jeff does the rest of the discovery himself.

Resources George and Ollie referenced

• OpenClaw — the agentic layer Ollie runs on the Mac Mini.
• Claude Code — what George used to build the dashboard; what Ollie uses to interact with Jeff.
• Railway — the fifteen-minute cron host that pulls APIs into George’s Google Sheet.
• SellerBoard — Amazon analytics; George pipes its API into the sheet instead of going direct to Amazon.
• Mintsoft — 3PL inventory management; SKU-level stock feeds the dashboard.
• Loop Subscriptions — subscription platform for Shopify; powers the new-customer subscription metrics.
• TripleWhale — Ollie’s previous analytics stack at 8hours. “Fucking expensive” is his on-record summary.
• Gorgias — the CS platform Jeff reads to generate Patricia’s weekly coaching brief.
• Honest Dog Company — George’s brand.
• 8hours — Ollie’s brand.

About George Sanderson

George is the founder of Honest Dog Company, a UK natural dog-supplements brand best known for its bone-broth range. Connect with George in Slack here.

About Ollie Groombridge

Ollie is co-founder of 8hours, an Amsterdam-based sleep and recovery supplement brand built around the circadian rhythm. Connect with Ollie in Slack here.