JaFicht

COO - Springland Pets

AI - antiques - growing businesses

[JaFicht @JaFicht]

Most businesses aren’t in a position to deploy ai agents.

First step -> get a server.

A small always-on computer that becomes the operational brain of the company.

Right now, most SMBs run on scattered systems:

* Shopify for orders
* QuickBooks for accounting
* Gmail for approvals
* Spreadsheets for inventory
* Slack texts and sticky notes for everything else

The data exists.
But the business has no memory.

So every week becomes:
“Can someone pull that report?”
“Who updated this spreadsheet?”
“Why are these numbers different?”
“Did we reorder that already?”

What I’m building is essentially an AI operating layer for my business

A private server that:

* pulls data from every system
* stores it in one database
* keeps historical memory
* lets your team talk to the business in plain English
* and gradually replaces manual operational work

And what’s so different now with ai coding tools vs. before - the software system is flexible and can adapt to the needs to the business rather than the business adapting to the needs of the software tool.

Once the company has a clean operational memory:

* inventory planning changes
* forecasting changes
* approvals change
* reporting changes
* onboarding changes
* decision making changes

You stop asking:
“Where is the information?”

And start asking:
“What should happen next?”

That’s the transition I think a lot of people are missing.

AI is not just another SaaS tool category.

It’s the first time non-software businesses can realistically build their own internal software systems without needing an engineering team.

[JaFicht @JaFicht]

For years, software companies promised to “replace the spreadsheet.”

Honestly, most of them didn’t.

At least not for me.

The spreadsheet always came back.

Inventory planning.
Forecasting.
Open POs.
Launch calendars.
Margin scenarios.
Retail tracking.
Ad hoc operational logic.

The reason spreadsheets survive is simple:

Businesses are messy.

The real workflow never fully matches the software.

So eventually you end up back in a spreadsheet because it’s the fastest way to adapt reality.

Most SaaS tools were trying to force the business into fixed workflows.

What changed for me wasn’t better dashboards.

It was AI + MCP servers + centralized business data.

That combination finally got me to the point where the spreadsheet actually started disappearing.

Because the system became MORE flexible than the spreadsheet.

Now instead of:
- manually updating tabs
- exporting CSVs
- stitching reports together
- maintaining fragile formulas
- rebuilding the same analysis every week

…the AI can interact directly with the operational layer of the business.

The key shift:
The business context is no longer trapped inside spreadsheets.

It lives in a shared data layer:
- orders
- inventory
- lead times
- supplier history
- marketing data
- financials
- operational logic

Then MCP servers expose that context to AI systems securely.

So instead of building another spreadsheet, I can ask:
- “What inventory risks do we have for Q4?”
- “Which SKUs are becoming inventory traps?”
- “Build a reorder recommendation based on current velocity.”
- “Compare current forecast to last year adjusted for Meta spend.”

And the answers are generated dynamically.

Ironically, this is the first thing that has actually felt more flexible than spreadsheets.

That’s the important part.

The spreadsheet wasn’t winning because it was ideal.

It was winning because it was adaptable.

AI-native operational systems are the first real alternative I’ve seen that preserve flexibility while removing a massive amount of manual work.

[JaFicht @JaFicht]

Wispr flow is the most fun tool I've been using these past couple months.

It's completely changed how I work on a computer. Instead of typing long emails, I'm now dictating and Claude cleans up my words into something that's really concise and actionable.

When I'm doing work and building workflows I just riff on the topic and Wispr Flow listens and I drop that into Claude Code.

Claude then takes that context and uses our internal systems with all of our data connections to our key business tools to accomplish complex tasks quickly.

The gap between "we should automate this" or "we should do this" used to be a quarter. Now it's minutes to a couple hours.

[JaFicht @JaFicht]

Stop giving your team ChatGPT.
Give them an MCP server.

Most companies "rolling out AI" hand employees a chatbot and hope for the best.

The result: 50 people pasting screenshots of revenue data into a generic LLM, getting different answers, and making decisions on hallucinated numbers.

 There's a better pattern. It has three pieces:

 1. A shared MCP server
MCP (Model Context Protocol) lets Claude connect to your actual systems — your warehouse, your CRM, your analytics, your internal docs. You stand up one
server. Every employee's Claude points to it. No one is copy-pasting CSVs anymore.

2. A semantic layer in front of your data
This is the part most companies skip and then wonder why their AI is wrong. Raw tables don't know what "active customer" means at your company. Or "Revenue." Or "fulfilled order." A semantic layer (Cube, dbt, etc.) defines those concepts once, in code, with the business logic baked in.

3. Tool access scoped by role
Marketing gets analytics + the CMS. Ops gets inventory + shipping. Finance gets the GL. Same Claude, different toolbelts. Permissions live on the server,
not in a prompt.

 What this unlocks:
 → A CSM asks "which of my accounts are at risk?" and Claude pulls usage data, support tickets, and renewal dates — through governed metrics, not guesses.

 → A marketer asks "what did we spend on Meta last week vs. pipeline generated?" and gets a real answer in 10 seconds.

 → A founder stops being the human API between the data team and everyone else.

 The companies winning with AI right now aren't the ones with the fanciest models. They're the ones who built the plumbing so their team can actually USE
 the model on real work.

 MCP server + semantic layer + scoped tools. That's the stack.

[JaFicht @JaFicht]

@briancheong Great point - just read about Dunai - seems like you guys are solving a lot of the issues I’ve been building around - will check it out

[JaFicht @JaFicht]

Most AI agent failures aren’t model failures. They’re data failures.

You can hand an agent the smartest model on earth, but if it’s pulling raw rows straight from your database, it’ll invent metrics, miscount revenue, and confidently give you the wrong answer.

The fix is something most teams skip: a semantic layer.

At Springland, we use Cube (free and open source) as the layer between our database and everything that touches data — dashboards, agents, analysts. It’s where we encode what each metric actually means.

Here’s the difference:

Without a semantic layer, every dashboard re-implements its own logic. What counts as an “active customer”? Is “net revenue” before or after returns? Each dashboard answers slightly differently. Multiply that across teams and agents, and you get drift.

With a semantic layer, that logic lives in one place. Define “net revenue” once. Every dashboard, every agent, every person pulling data sees the same number.

The payoff for AI agents specifically:

→ They query business concepts, not raw tables
→ Their answers stay consistent with what your team sees in dashboards
→ Building a new dashboard for anyone on the team becomes trivial
→ You can actually trust the analysis

If you’re building toward an AI-driven business, the semantic layer isn’t optional. It’s the piece that makes the agents actually useful.

[JaFicht @JaFicht]

At Springland, inventory is the most analytically complicated and nuanced part of running our business. It's not just "forecast demand, cut a PO." It's:
 - Cash flow (always cash flow)                                                               
 - Unit pricing and volume tier pricing                                                           
 - Balances across multiple warehouses                                                            
 - Multiple sales channels
 - Kitting and pack configurations                                                              
 - Big one-time wholesale orders we know are coming
 - Manufacturing constraints and variable lead times                                                     
 - Holidays and seasonality
 - Different suppliers with different terms 
- COGS, tariffs, shipping costs — which means it determines whether we're actually profitable or just think we are.

This used to be all managed by a ton of manual spreadsheet based processes and a lot of knowledge that only existed in my head. It was super time consuming and always felt like we were serving the system and that the system wasn't serving us.          

So I rebuilt our factory ordering workflow as an MCP server tool with Claude code. Now I plan production orders by talking to Claude. But the conversation only works because of the layers underneath:

 1. Data flows. All the data affects our inventory decisions lands in one place, kept fresh.                              
 2. Document the data. We have really clear definitions of what the data means and how to interpret different fields.
 3. Document the business logic on top of it. The formulas you'd otherwise carry in your head — velocity, lead times, reorder points, cost rollups — written
  down and available to call.                                                                
 4. Encode your thought process. The judgment calls. How you weight recent vs long-term sales. When to skip a SKU. How you think about volume pricing
 tradeoffs. What "covered" actually means.

Once all of that is in code and docs, Claude can reason on top of it.

And what's even cooler is that the analysis is flexible to adapt to new scenarios and ways of thinking. "Big retailer order coming in March." "This SKU is being discontinued, don't reorder." "Factory is moving locations next quarter, add two weeks of lead time." Those used to require manual override on every report. Now they're facts in the system the model factors in automatically.              

The tool stops being a static report and starts feeling like an analyst sitting next to you.

[JaFicht @JaFicht]

When I started building internal tools with AI, security was my biggest worry.

As I started deploying tools and centralizing our data on a server for our team to access - I needed to be very confident in our security posture.

I’d run security checks on the code, but I couldn’t say every vibe-coded app was airtight. And every new app expanded my attack surface.

Then I set up Cloudflare Zero Trust (free up to 50 users).

Now my server isn’t on the public internet at all. The only way in is through a Cloudflare-managed tunnel: sign in with Google, your email gets checked against a whitelist, and only my team gets through.

The real win was the mental shift. Instead of securing a dozen apps individually, I secure one gate. Outbound API calls still work fine. Inbound traffic from anyone not on my team simply doesn’t reach the server.

I trust Cloudflare + Google a lot more than I trust code from Claude.

If you’re shipping internal tools with AI, this pattern changes everything

[JaFicht @JaFicht]

When I started building internal tools with AI, security was my biggest worry.

I’d run security checks on the code, but I couldn’t honestly say every vibe-coded app was airtight. And every new app expanded my attack surface.

Then I set up Cloudflare Zero Trust (free up to 50 users).

Now my server isn’t on the public internet at all. The only way in is through a Cloudflare-managed tunnel: sign in with Google, your email gets checked against a whitelist, and only my team gets through.

The real win was the mental shift. Instead of securing a dozen apps individually, I secure one gate. Outbound API calls still work fine. Inbound traffic from anyone not on my team simply doesn’t reach the server.

I trust Cloudflare + Google a lot more than I trust my vibe coded web apps.

If you’re shipping internal tools with AI, this pattern changes everything.

[JaFicht @JaFicht]

I built our own CRM at Springland.

~20 hours of part-time work over two weeks.
The reason wasn’t to save on SaaS subscriptions — though that’s a nice byproduct.

The real reason: we own the tools and the data. Which means we can shape the exact experience we need, instead of bending our process around someone else’s product roadmap.

Here’s what ours does:
→ Pulls every inbox into one view per customer
→ Consolidates wholesale orders from every platform we sell on into one place
→ Imports tradeshow contacts and the notes we take there in seconds
→ Powers super-customized email flows using full customer context
→ Surfaces the next action and current bottleneck for every account
→ Summarizes meeting notes and ties them to the right accounts
→ Debriefs me on the highest-value next steps across the entire pipeline, with top priorities clearly spelled out
→ Researches prospects at target accounts, and suggests new accounts based on patterns in our existing customer list

But the part I didn’t expect to love this much:

I can ask it anything in Claude Code and get a real answer in a few minutes.

“Which wholesale accounts haven’t reordered in 90 days but were trending up before that?”

“What’s the common thread across our top 10 stores this year?”

“Draft a re-engagement email for everyone who met us at Global Pet Expo but never placed an order.”

It’s like having a sales analyst and a copilot sitting next to me at all times.

This is trivial when you own your data and your product experience.

It’s nearly impossible when you’re locked inside someone else’s CRM.

What used to take an engineering team months or years can now be built by one person in a couple of weekends — if you know how to direct AI well.

Owning your CRM used to be a liability — the juice of customization wasn’t worth the squeeze. That math has totally flipped. Now it’s an edge.

Sales ops time is down 80%+, and we’re staying on top of 3–4x more deals with way more clarity across the pipeline.

And because the whole thing is our own, I keep encoding more of the sales process into it — vendor onboarding, pricing, inventory allocation. Every week it gets a little sharper.

[JaFicht @JaFicht]

@mikefutia VAULT

[JaFicht @JaFicht]

@NanFlorin2 122k

[JaFicht @JaFicht]

@NanFlorin2 122k

[JaFicht @JaFicht]

18th century Dutch onion bottle

[JaFicht @JaFicht]

Neolithic Chinese Jar from Majaiyao Culture - 4000 years old

[JaFicht @JaFicht]

Calligraphy scissors from the Ottoman Empire - 17th century

[JaFicht @JaFicht]

Medieval Bronze Measure from Continental Europe, beautiful condition

[JaFicht @JaFicht]

Dutch 17th century Heemskerk Candlestick

[JaFicht @JaFicht]

800-900 year old Temmoku Tea Bowl From the Southern Song Dynasty

[JaFicht @JaFicht]

@oliverbrocato Which analytics tools have you gotten the most value out of and why?