The tabs they keep open
Which six systems a single piece of work crosses, and which one of them is the reason the whole thing takes three days instead of twenty minutes.
How we work
Five steps. It starts with watching.
Most AI projects start with a tool and go looking for a use. We start with a week of watching your team do the actual job. We don’t build anything until we can name the work we’re taking off their plate.
Why it’s in this order
Every step is useless without the one before it.
Skills built on guesses automate the wrong work. Workflows that can’t reach your billing system stop halfway. Rules written after rollout are suggestions nobody follows. All three, installed for one volunteer, produce the pilot you already ran once.
So the order isn’t a preference. Watch, connect, teach, fence, deploy. Taken out of sequence, each step quietly undoes the value of the last.
Step 1 of 5 · Week 1
We shadow your team.
No questionnaire, no discovery workshop, no slide where you tell us what your process is supposed to be. We sit with your people and watch the real work happen, at the speed it actually happens.
What we watch for
What we watch for
The waiting
Where work sits still because it is queued behind one person’s approval, one person’s expertise, or one person’s inbox.
What we watch for
The rebuilding
The report, the proposal, the reconciliation that gets assembled by hand every single week from the same four sources.
- What you get at the end of the week: a ranked list of what is worth automating, with the hours behind each line.
- And the other list: what we think you should leave alone, and why. There is always one, and it is usually the thing somebody was most excited to automate.
- No build starts until you have read both and told us which ones we are doing.
This is the step most AI projects skip, and it is why most of them automate something impressive that nobody needed. A week of watching is cheaper than a quarter of building the wrong thing.
Step 2 of 5 · Weeks 2–4
We connect it to everything you use.
Real work crosses six systems. Miss one and the chain stops and waits for a human. The work didn’t get automated, it got interrupted. So we don’t stop at the easy integrations.
01
The ones with a connector
Slack, Gmail, Drive, Notion, HubSpot. Installed and working the same week. This is the part everyone can do.
02
The ones with an API and no connector
Your billing platform, your scheduling tool, the vertical software your industry runs on. We build the connector against their API.
03
The ones you built yourself
The internal dashboard that will never have an off-the-shelf integration. We wrap it so your AI can call it like anything else.
04
The ones with no API at all
Some vendors simply don’t offer one. We drive the browser directly, signed in, clicking the same screens your team does, and turn that into a tool.
- Connected once, for the team. Not a setup guide each person follows on their own laptop and half of them abandon.
- Hosted where your security team needs it. Your infrastructure if that is the requirement, ours if you would rather not think about it.
- Portable. If you switch AI tools next year, the connections come with you.
Step 3 of 5 · Weeks 2–4
We teach it how your team does the work.
The workflows we watched in step one, the way you write a proposal, close a month, escalate a ticket, get encoded as skills anybody can run. Not generic best practice. Yours.
Request lands
Email or Slack
Pull the account history
CRM · billing
Draft it your way
Your encoded skill
Owner approves
Human, every time
Send and log it
CRM updated
One encoded workflow. The dark step is a person. You decide which steps stay that way, and step 4 is where that gets enforced.
- Sourced from your best person, not from us. The hundred hours they spent getting good at this becomes something the new hire runs on day one.
- Written down, not trapped in a prompt. Skills live in your repository, get reviewed like anything else, and change when your process changes.
- Tested on your real data before anyone relies on it — never a demo account with clean sample records.
Step 4 of 5 · Weeks 2–4
We set the rules it can never break.
Nobody hands real work to something they can’t predict. Managers decide what can be delegated and what can’t, in plain language, and those decisions are enforced rather than written in a document and hoped for.
Who it may contact
Which customers, which vendors, which internal lists. And which messages never leave the draft folder without a person on them.
What it may spend
Hard limits on money and on cost per run. A budget that stops the work is better than a surprise invoice.
How it sounds
Your brand voice and your legal language, applied everywhere, including the fast internal reply nobody would have proofread.
What it must never guess
Pricing, dates, commitments, anything a customer will hold you to. When it doesn’t know, it stops and asks a named person.
What stays human
The approvals that don’t get automated at any price. Marked explicitly, so the boundary is a decision and not an accident.
What it did, and when
Every action logged and attributable. If a manager asks why something went out on Tuesday, there is an answer.
Rules are what make the rest of it safe to use. Teams that skip this step don’t get a disaster. They get something quieter and worse: nobody trusts it with the work that matters, and the pilot dies of politeness.
Step 5 of 5 · Week 5 and after
We roll it out to everyone.
Not three volunteers. The whole team gets the same connections, skills, and rules installed. The ten-year veteran, the person who has never opened an AI tool, and the hire who starts next month.
- Training that assumes nothing. Built for the person who has never used any of this, not for the one who already got it working alone.
- Day-one default for new hires. The setup is part of onboarding, so the gap never reopens after we leave.
- Tuned on what people actually use. We watch the first weeks of real usage, cut the skills nobody touches, and fix the ones people are working around.
- Then you own it. It is yours to run and change. We stay on if you want us to, and plenty of teams do — but nothing here depends on us being in the room.
The point of the whole process is this step. Everything before it is setup; this is where the difference between your best person and everybody else stops being a hundred hours of private effort.
On a calendar
What this looks like in real weeks.
Typical shape for a Foundation build. Bigger teams and messier systems stretch the middle; nothing skips the front.
Week 1
Shadowing
We are with your team watching the work. You get the ranked list at the end of it and pick what we build.
Weeks 2–4
Connections, skills and rules
Steps two, three and four run together. Connections go in, your workflows get encoded, and your managers set the rules alongside it. You see it working on your own data.
Week 5
Deployment and training
Installed for everyone, with training that assumes no prior experience.
Weeks 6–8
Tuning on real usage
We watch what people actually run, drop what they don’t, and fix what they are working around.
After
You own it
Yours to run. Add Nexus later if your team outgrows a shared assistant.
What we need from you
Three things, and they are small.
01
A week of access to real work
Time with the people doing the job, on live work. Not a summary of it from someone who manages them.
02
Someone who can grant access
One person who can approve connections to your systems. Most delays on these projects are this, not engineering.
03
A manager who will set the rules
An hour or two from someone with the authority to say what may be delegated. Step four doesn’t work without a real decision-maker.
Working together
Three ways to run this with us.
The five steps don’t change. Who does the building does.
01
Advise your AI team
You already have people on this. We run the process alongside them, tell you what is going to break, and set the direction, so your team stops burning months on the wrong approach.
02
Build with you
We embed and build it together. You own it at the end, and your people know how it works because they helped make it.
03
Built for you
You don’t have people to spare. We run all five steps, deploy to your team, train them, and hand you the keys.
Worth saying plainly
What we don’t do.
We don’t run prompting workshops. Your team doesn’t have a knowledge problem, it has an access problem. The tool can’t reach the CRM and nobody told it how your company works.
We don’t build a demo on sample data. If it hasn’t run on your real records, we haven’t shown you anything.
We don’t automate a job we haven’t watched, and we will tell you when something on your list isn’t worth it. Both lists come back from step one for a reason.
Step one starts with a conversation.
Thirty minutes. Tell us what your team does all day and we will tell you honestly whether there is a week of shadowing worth doing.