We embed with the team, not the org chart

Teaching you the tool is where most projects stop.
We stay until the AI is running your work.

Which number has to move, for whose decision, by when. We settle that first — then build it with the two-to-five person team that actually owns the problem. Company size is not the point. A division inside a global firm and a company of five are the same unit to us. After three months what remains is not a slide deck; it is an agent that has already finished its work before Monday morning.

  • 11weeks, 8 sessions
    for a global electronics maker
  • 2client members now build
    their own agents
  • 0client systems or internal
    data touched
quota-intelligence · AI diagnosis SYNTHETIC

Weekly analysis effort

15hrs 5min

99.4%of the weekly roll-up removed

Every client sorted into four quadrants, automatically

  • Structural risk
  • Growth driver
  • Behind pace
  • Holding
  • 2wksearlier warning on a miss
  • +$7.6Mof revenue surfaced from field notes

Open the one closest to your work — all six are live and clickable.

or
An artefact we actually built, running on entirely synthetic data. Real engagements use public data only — we do not touch client systems or internal data.
Trial slot · one-shot This, built on your own data. One working dashboard, ¥1M flat. A slot for trying it once — full projects are quoted separately.

The problem

Why training alone never changes anything

01

“I understand it” is not “I built it”

Seminars end at comprehension. But a workflow only changes when someone has actually built the thing that replaces it — with their own hands, on their own data.

02

Generic prompt libraries don’t know your job

Every company’s real work is held together by undocumented judgement calls. A generic template cannot encode the exception that your team handles every Thursday.

03

Outsourced systems stop when the owner leaves

A tool built entirely by a vendor is a tool nobody inside can extend. Six months later it is quietly abandoned — and the budget is gone.

So we work the other way around: your people build it, we sit next to them until it runs.

Why us

Engineers already embed with clients.
We are the business-side version of that.

Software has a role called the Forward Deployed Engineer — a strong engineer who embeds inside the client’s workflow and rebuilds it from within, so that one person’s judgement leverages the whole organisation. What AI adoption is actually short of is the business-side equivalent.

Carried the number

We have owned a P&L, not observed one

Sales leadership at a Western tech company, with the revenue and the loss on the same sheet. So we do not close a report with “we removed some hours.” We push it to which number moved, and by how much.

Close to the data

We can tell when an output is not decision-grade

Careers built next to the analytics and the digital stack. The hard part of automation is not generating an answer — it is knowing when the answer is not good enough to act on. That judgement only comes from having had to act on it.

Across industries

Not one function in one industry

Business lead roles across several industries rather than a single vertical. It means that on an unfamiliar workflow we already have a hypothesis about where the waste collects — usually at the handoffs nobody owns.

So our work does not stop at swapping a tool in. We rebuild the workflow from the marketing and business layer, and then reconnect the hours freed and the accuracy gained to the investment decision they were supposed to serve. That last connection is the part that usually goes missing.

How we work

Four steps. We leave when you no longer need us.

  1. 01

    Take the work apart

    We sit in on the actual task — the one people call “a knack” or “something only they can do” — and translate it into steps an AI can hold.

  2. 02

    Build the first one together

    Not a lecture. Your person drives the keyboard; we sit beside them. The goal of week one is a single small thing that visibly runs.

  3. 03

    Run it weekly on real data

    This is where we differ from training. Every week we hunt the reason it broke, and fix it. Demos survive one day; this survives Mondays.

  4. 04

    Hand over, and let go

    We finish when your team can add the next agent without us. Templates, prompt libraries and the decision log stay with you.

Selected work

Before “how much”, here is what stays behind.

Client names are withheld under confidentiality. Everything below was actually built and delivered. Figures under “scale of the saving” are the projections agreed with the client when the work was scoped, not audited results. Prices live in one place only — see “how we work together” below.

Descriptions only go so far. These six actually run — open whichever is closest to your work.

Overseas subsidiary · Consumer electronics · Marketing

An agent that has already finished the Monday analysis before Monday

BEFORE MONTUEWED THUFRI Manual collation Misses the meeting; decisions slip a week AFTER MON 08:30 — already generated MONTUEWED THUFRI All five days go to execution
The problem
Weekly competitor pricing and review trends were compiled by hand. The analysis was never ready in time for the Monday morning meeting, so decisions slipped a week.
What we built
A set of agents that collect public marketplace data, analyse price movement and review signals, and generate a prioritised action list before Monday 08:30 local time.
What this was
A workshop, and a business transformation project. The client ran it on the same understanding.
How it ran
11 weeks. Eight 90-minute working sessions and three asynchronous preparation weeks, across a time zone, in Japanese and English.
What stayed behind
Two members of the client’s own team can now design and add new agents without us. Public data only — no internal systems, accounts or company data were used at any point.
Scale of the saving
Agreed with the client at design time: roughly 700 hours of manual work removed over two years (about 400h on the Monday report for two people, about 300h on price and stock monitoring) — on the order of $35,000. Payback lands between five and ten months, and the saving repeats every year after that, because the work was removed rather than sped up.

Payback 5–10 months, then the saving repeats every year

Social media services company · approx. 50 employees

Recurring reports, handed over to a machine

BEFORE Sources Manual build Report The same hours vanish here every month AFTER Sources Automatic Report
The problem
A meaningful share of the team’s month went into assembling the same reports from the same sources.
What we built
A delivered report-automation solution, designed around their existing review and approval flow rather than replacing it.
How it ran
2 months, build-and-deliver project.
Scale of the saving
The assembly step is gone rather than faster: the monthly hours that went into rebuilding the same report from the same sources are no longer spent at all, and are not spent again next month, or the month after.
Note
Originally scoped as a workshop. We changed it to a build engagement because that was the honest answer to their problem — and we still recommend that trade when it is the right one.

Payback The assembly work stopped recurring — every month, from then on

Professional association · accounting & tax

Sixty minutes of work, done in five, in front of the room

BEFORE 60 min AFTER 5 min
Format
An entry session for professionals with no prior AI experience: a live demonstration on their actual routine work, followed by hands-on.
Where it leads
Our entry point, not a product line. Where a room wants to go further, it becomes a project.

BtoB manufacturer · Communications & marketing

The bottleneck was never the writing

BEFORE Draft Find sources (human) Fix wording (human) Publish Every rejection goes back to step one Idea to publish: 11.5 days / 3 a month AFTER Draft Source linking & wording check (automatic, stops before publish) Approve (human)
The problem
They wanted to publish, and could not. Every post carried figures and comparisons that needed checking, so drafts went back and forth between the writer, the source check and the wording check. The team was producing about three posts a month, and could not say which of those steps was actually costing the time.
What we built
A publishing queue where every figure in a post is tied to a source, plus six machine checks that run before anything goes out — anonymity, banned wording, source verification, overclaiming, quotation practice, and live disputes. A post that trips any of them stops, with the reason written down. The weekly view then shows what is queued, what is waiting on whom, and what each format actually reached.
How it ran
Three weeks, three sessions, fixed fee. Delivered and handed over.
What changed
Idea to published went from 11.5 days to 3.1, and monthly volume from about 3 posts to 14. More usefully, the dashboard showed that approval had become the constraint — it now accounts for 57% of the time spent on a single post. That was not visible before, and it is the next thing they will fix.
Note
We did not automate approval, and declined to. The check that a figure actually appears in the cited document is done by a person opening the document. Flipping that flag automatically would have removed the only thing that makes the rest trustworthy.

Fee ¥1,000,000 — three weeks, three sessions, nothing after

New business development · nine concurrent projects

Not the project that is late — the stage that gets skipped

BEFORE — by project Nine reports, one by one. Each looks "almost there" AFTER — cut by stage Idea 9 Build 8 Launch 5 Reach 0 ↑ No owner, no estimate — only here
The problem
Nine projects, reviewed one at a time in a 90-minute weekly meeting. Every project had a reasonable-sounding explanation for where it stood, and none of them pointed at anything the team could act on. Hours were being spent close to plan while finished work was running at a third of target, and nobody could say where the gap was going.
What we built
A board that reads the existing project tool, time records and approval log, and cuts them by stage instead of by project: how many of the nine cleared each stage, how long each one sat untouched, who each decision is waiting on, and how the hours split between work that survives in a deliverable and work that does not.
How it ran
Three weeks, three sessions, fixed fee. Delivered and handed over.
What changed
The first view found something a year of project-by-project reviews had not: one stage — getting the finished thing in front of anyone — had no owner, no estimate and no definition of done, so zero of nine had ever cleared it. A stage with no row on the chart cannot show up as late. The weekly meeting went from 90 minutes to 15, and the meeting now ends with decisions rather than status.
Note
No new input was added. Everything on the board already existed as a by-product of the work — ticket history, time records, approval requests. A dashboard that needs people to type things in for it is a dashboard that stops being true within a month.

Fee ¥1,000,000 — three weeks, three sessions, nothing after

Investment

Start by building one.
Then there are four ways to keep going.

One-shot · fixed price

One dashboard. Three weeks, three sessions. Then we are done.

  1. 1wk Hearing Which number has to move, and who decides on it. We leave with the spec, and you send over the data to use.
  2. 2wk Base lecture & prototype How the thing works, and a first version on your actual data. You tell us what is wrong with it.
  3. 3wk Build & hand over We fix what you flagged, hand it over, and walk you through how to change it yourselves.

¥1M Flat. No monthly commitment, no retainer attached.

Ask about the one-shot build

If it turns out that a build is not the honest answer for you, we say so in the first session and stop there.

Everything below is for teams who want the second, third and tenth one built by their own people.

Entry

Diagnosis / spot advisory

¥100k – ¥300k

Half a day to a full day. We look at the actual work and come back with a specific answer on what AI changes — and what it does not. Deliberately not free.

Core

Implementation project

¥3M – ¥6Mover 3–6 months

We work alongside your two-to-five person team, on your real data, until they run without us — then stay long enough for it to hold.

Weekly at the start, monthly once they are running. The density is meant to fall — that drop is the result we are aiming for.

Build

Agent development

¥3M – ¥8M

PoC through production. When building it for you is the honest answer, we build it — and hand over the parts your team should own.

Embedding

Keeping it running

¥200k – ¥500k / month

After hand-over, the phase most projects skip. We keep what was built in daily use as the work changes around it: monthly review, a channel for “is this worth doing?”, and what actually changed in the tooling this month.

Pricing works backwards from the hours that disappear. On a recent project the design removed roughly 700 hours of manual work over two years — on the order of $35,000.

We do not discount. If the budget doesn’t fit, we cut scope and say so plainly — a discounted engagement is one we cannot staff properly, and you would feel it by week three.

Fit

Who this works for — and who it doesn’t.

Saying this up front saves both of us a month.

We work well together when

  • You can describe your own work in detail — the exceptions included
  • Someone on your side will commit three months of real time
  • The decision maker is in the room, or one seat away
  • You want your own team to own it in the end
  • Can move a project budget of ¥3M and up

We are the wrong partner when

  • You want to hand it over entirely and see a result
  • The brief is “something with AI, for now”
  • Price negotiation is the main conversation
  • You want the deliverable, but nobody inside will own it afterwards
  • AI is expected to fix a problem that is actually organisational

Team

Not one person’s consultancy.
Everyone here has run a business, not advised one.

Yutaka Kesamaru

Founder & CEO

Yutaka Kesamaru

Hitotsubashi UniversitySonyDeNAHakuhodo

  • Now — Western tech company: leads the sales organisation, carries the P&L

For now, the founder is the producer on every engagement. That is a function of how many we can run at once, and it is not a promise we can keep indefinitely.

A career spent on the side of the table where the number has to move: carrying a revenue line, reading the data underneath it, and growing the business on what the data said. That is why these sessions are not training. We start from the business question — which number, whose decision, by when — and the agent only comes up after that.

LinkedIn →

The bench

The people we bring in have run the businesses you are trying to change

Not a fixed org chart. We pick the background the engagement actually needs. Company names are withheld under confidentiality, but the following sit within a phone call.

  • Sitting board member at a major venture company — marketing
  • Head of Marketing at Japan’s largest IT conglomerate
  • Ran the recruiting business in Japan and overseas, and a mobile payments service
  • Product manager at a major SaaS company that has raised ¥10B
  • Large-scale direct-to-consumer commerce
  • Office of the President at an online securities brokerage
  • Founder-CEO of a games venture

Implementation lead

Agent architecture & templates

Designs the templates and prompt libraries clients keep after hand-over. Runs the hands-on half of every session — the part where the thing has to actually run.

Infrastructure & review

Environment, security posture, program review

Sets up the working environment before session one, keeps the “public data only” boundary enforced, and reviews each program from outside the delivery team.

We deliberately keep the permanent headcount low and compose the team per engagement, so the experience in the room matches the problem in the room rather than whoever happened to be free.

Field notes

What we write down after each engagement

We publish what worked, what didn’t and the numbers behind it on note (in Japanese). Details are generalised to protect clients.

All field notes (Japanese) →

FAQ

Questions we get before the first call

Our data can’t leave the company. Is that a problem?

It is the normal case, not the exception. Our overseas program was delivered end-to-end on public data only, with no access to internal systems, official accounts or company data. We design the boundary with your security team before session one.

Nobody on our team is technical. Can they keep up?

Our participants are marketers, accountants and business leads — not engineers. The sessions are built for people who work in natural language. If your team can write a clear brief, they can build an agent.

Can you run this in English, or across time zones?

Yes — that is our home ground. We have run a full multi-week program across a time zone in Japanese and English, for a Japanese-affiliated team overseas with local non-Japanese members participating.

How soon do we see something?

The first session ends with something running. Whether it is worth keeping is a question for week three — and we will tell you honestly if the answer is no.

Which tools do you lock us into?

None by default. We pick per engagement and tell you the running cost before you commit. If the honest answer is a ¥3,000/month tool instead of a build, we say so.

Get in touch

Thirty minutes. Bring the problem, not the solution.

No deck, no pitch. We will tell you within the call whether this is something we should work on together — including when the answer is no.

押すと、下の相談内容に文章が入ります。書き換えても、消しても構いません。

We reply within one business day. If we are not the right partner, we will say so in that reply.