AI gets smarter.
The messy human part
doesn't go away.
I run my own company on AI — not as a pilot, not as a concept, but as the actual operating model, every day. I've already made the mistakes, found the edge cases, and worked out what doesn't scale before my clients ever see it.
But even running that way, I was in the room when a senior manager who'd been using the same workflow for 15 years stopped using the new system the moment the field labels changed. Someone had to sit next to him for three days before it clicked. AI got smarter. That messy, human part didn't go away.
So the real question is: what do you hand to AI, and where does human judgment stay in the loop? Designing that boundary is what's needed most right now. For us this isn't a separate product — it's the deepest form of engagement in our system development work: deliver and hand off, take over operations, or coach your team until it runs in-house. That's where a sounding board session starts.
Shogo Harada — CEO, Linnoedge · Ho Chi Minh City. We build systems with a dedicated Vietnam-based team — including the 30× throughput deployment described below.
What separates companies that get results
from those that don't
AI moved from "something you use" to "something that works alongside you." Claude Code, Cursor, Microsoft Copilot — agents reading files, writing code, making decisions. The question is no longer whether you use AI. It's whether you've designed where the boundary sits.
The companies getting results aren't the ones with the best AI tools. They're the ones that designed what AI handles and what stays with humans — before the rollout, not during the cleanup after.
At one company I worked with, the mandate was simple: no work on Wednesdays that doesn't involve AI. That sounds blunt, but it worked — because it came after the harder work. The team had already mapped out which decisions could be handed off and which ones couldn't. The mandate didn't create the momentum. The prior clarity did.
At a manufacturing client, we rebuilt the estimate entry screen to look exactly like the one they'd been using for 12 years. A modern interface would have been unused within a week. There's nothing wrong with the people — a 55-year-old who's built their whole workflow around one screen doesn't need a redesign, they need their workflow respected.
Rebuilt it to look identical. Added AI to 3 of the 6.
Nobody needed retraining.
Then there's the flip side. At another company I worked with, AI made the team 30 times more efficient — and immediately broke something else. Procurement couldn't keep up with the volume. And as the AI automated more, the checkpoints where humans would normally catch errors became invisible. More throughput, less oversight. Nobody had designed for what happens when it works too well.
Nobody designed
what happens
when it works.
The patterns we see most often
- Tools distributed — adoption stays at 20%
- Stuck in vendor selection for months
- Training delivered, workflow unchanged
- Mapped to actual workflow — before a line of code is written
- Tested in production conditions, not just in a demo environment
- Accountable after go-live, not just at handoff
Honest note
This is not the right fit if:
- ×You want a polished roadmap document — not real change in how the work happens
- ×You need someone to own the AI strategy so your team doesn't have to think about it
- ×You're looking for validation of a decision you've already made internally
This is the right fit if:
You want to change how the work actually happens
- ✓AI is deployed, but adoption is stuck below 30% and you don't know why
- ✓You need someone to say out loud what the team can't say in the internal meeting
- ✓You want a vendor who stays accountable six months after go-live — not just at handoff
"I watched the AI hit 30× throughput in week three.
By week eight, there were 2,800 procurement tasks
with nobody assigned to review them.
Getting it to run was never the hard part.
Knowing what breaks next — that's the work."
— Shogo Harada, CEO, Linnoedge
What goes through your mind
when you're about to commit to AI
These come up in almost every conversation — regardless of company size or industry.
Will this actually move the numbers?
We don't have anyone internally who gets this.
What if the vendor delivers something unusable — or disappears?
What happens after go-live
Every deployment includes a 30-day stabilization period. During those 30 days our team in Ho Chi Minh City watches output quality, checks that the escalation path is actually being used, and catches drift in model outputs before it settles into a pattern.
After that, monitoring is optional. Companies that want us to keep watching move to a monthly contract. Companies that don't, don't, and the system stays theirs to run. That choice is deliberate. The thing we try not to leave behind is a maintenance contract you cannot get out of, wrapped around a system nobody inside your company can explain.
What the monthly contract covers
Performance monitoring, drift detection in model outputs, and escalation of the edge cases that need a human decision. The purpose is to keep the boundary between what the AI handles and what a person decides working the way it was designed, months after anyone last thought about it.
Where the optimization actually happens
Our team measures the accuracy of what the integration produces. When it drops, we go back into the model or the prompt and change it.
The other axis is scope. When your workflow changes, the line between what the AI handles and what a person decides has to move with it. An integration that still runs correctly but has stopped being used produces nothing, so we widen the scope to the next workflow when the numbers hold.
When quality drops, one team owns the answer
We build the data pipeline and deploy the model ourselves, in that order. That matters most when output quality drifts, because someone has to say whether the cause is the data or the model, instead of two vendors pointing at each other.
Where you are now shapes
what comes naturally after the session
Depending on where your organization is stuck, the sounding board session will surface a different next step. It might be one of the paths below — or it might be "not right now" or "a different vendor would be a better fit." Either way, you'll leave with a clearer picture than you came in with.
ChatGPT, Copilot, Gemini — licenses distributed, and maybe 20% of the team actually uses them. This is the most common pattern we see right now.
The first thing we do is spend a week observing how the tools are actually being used, and more importantly, why they're not. Almost every time, the answer isn't the tool — it's the workflow. Nobody mapped out which decisions can go to AI and which ones need a human. Until that's clear, more training just adds noise.
We redefine the scope for one workflow, make the boundary explicit, and run it there first before spreading it further.
Before getting internal sign-off, there's more to sort out than success metrics. Who's making decisions? Who's checking the outputs? Who stops the process when something's off?
Without that role design in place first, even a well-defined success metric doesn't move the team. We set the exit threshold before anything is built — so if it doesn't work, the course correction happens while the damage is still manageable.
We help you go into budget approval with numbers you can actually report on — not "we'll see how it goes."
With AI agents in the mix, a system that would have cost $130,000 to build two years ago can now come in around $65,000. That's real — but it's the development cost, not the full picture.
This becomes especially critical with agentic AI systems — RAG pipelines, LLM-orchestrated workflows, and multi-step automation — where the model makes sequential decisions without a human in the loop for each step. The boundary between autonomous execution and human escalation must be explicit before deployment, not discovered during a production incident.
Getting the system to stick in the actual workflow takes more time than building it. Before any code is written, someone needs to go on-site, watch the work happen for a day, and understand what's on which screen, in what order, and what decisions get made where. That's where the boundary gets drawn: which tasks go to AI, which ones stay with the person.
Skip that step, and you get a technically correct system that nobody uses.
Wherever you enter, the destination is the same — AI and humans each have a defined role, and the work is actually different. The path there depends on where you're starting from.
What this costs — the numbers, before you ask for them
A page that makes you request a quote before it will tell you anything never makes the shortlist. If I were the one buying, I would leave. So here is what we can put in writing before we have spoken.
For reference: with AI agents in the build, a system that would have cost around $130,000 two years ago now comes in around $65,000. That is the development cost. Getting it to stick in the actual workflow takes time the development budget does not cover — where the saved money goes is worth talking through in the session.
Measure first
Before you reach out, measure where you stand
Reaching out cold can feel like a big step. So before that, two free tools to see — right here — whether now is the right time for you, and how much your cost would change. Bring the result to a 30-minute session, and the conversation moves a lot faster.
Is now the right time for you?
Answer eight questions and we score your readiness, then send back three specific recommendations from your answers. Not “you should do this” — “here’s what to do, in your situation.”
Check your readiness → Cost calculator / instantHow much would your cost change?
Enter your team size and budget, and see how the next 12 months change — right here, with the reasoning behind the numbers laid out.
Calculate your cost →
In 30 minutes, we'll find where
your AI implementation is stuck.
You invested in AI. The tools are there. But the results aren't showing up — or nobody's using them. Most of the time, the cause isn't the technology.
If anything on this page felt like it was describing your situation, that's enough. Use the 30 minutes to say it out loud.
I run my own company the way I tell clients to run theirs — AI in every workflow, not as a side project, but as the actual operating model. Almost no hour of my workday happens without AI in some part of it.
Which means I bring the session what most consultants don't: I've already made the mistakes, hit the edge cases, and worked out what doesn't scale. Not a slide deck about AI potential — what I've actually seen break.
Based in Ho Chi Minh City. Working with companies in Japan, Southeast Asia, and globally.
Common questions about AI integration
consulting with a Vietnam team
Linnoedge is LINNOEDGE JOINT STOCK COMPANY (Enterprise Registration Certificate 0318272580), founded in 2024 and based in Ho Chi Minh City, Vietnam. The team is 10 to 49 people — engineers, data engineers and project managers — working with companies in Japan and Southeast Asia. AI integration and development engagements start from US$5,000 per project — a per-project floor that does not apply to our separately bookable training sessions.