Nikhil | Building Systems

turn your records into revenue signals.

[Nikhil | Building Systems @unik_47]

@ryangrowthh Ps. don't rebuild the engine during traffic jams

[Nikhil | Building Systems @unik_47]

Most firms could use AI on client work without the leak risk with just two habits:

Label data public, internal or client-private
Let the AI see only what the person may see

The system is pretty simple.

Use the three labels to stop the all-or-nothing choice:

Data the person is allowed to see
Sources shown on every answer
Logs of who asked what
Client-private facts that stay out of training
The one-page policy a client can read
The human check before anything leaves

Then use the misses to decide what to label next.

Instead of banning AI and watching staff use phones, pull:

Requests that still go to outside tools
Answers that lack a source
Data that has no label yet
Tests where the AI should have refused

Index them by level, role, client and owner, filter for what still lacks a rule, then enforce the four rules on every request.

A company brain is basically the perfect place for this because the labels and the retrieval sit in one place.

Then reuse the same three-level rule across every tool and treat each refused request as a sign of what to tighten next.

[Nikhil | Building Systems @unik_47]

if you're hiring a content team, write your method down before you post another thing.

take a go-to-market agency with an onboarding method that works on almost every client.

they have a real method clients trust.

steps in delivery docs, examples in old decks, and mistakes only the senior team remembers.

they were about to pay for a content team and start every post from a blank page.

what i'd tell them:

1. the blank page is going to be a shock. work out what you already know, not what you need to say

2. the same words on every channel will not land with a buyer in new york. write one master file with steps, examples, mistakes and proof instead

3. try lots of completely different forms per master. posts, emails, slides, webinars, checklists. you have no idea which one your buyer reads, and neither do i

4. let the leads tell you which forms work before you commit to a content budget

one master file is a lot cheaper than a year of posts written from scratch.

[Nikhil | Building Systems @unik_47]

@ImMediii Different clocks, same game.

[Nikhil | Building Systems @unik_47]

@i_mika_el That's the missing control. Measure the old workflow before celebrating the new one.

[Nikhil | Building Systems @unik_47]

Every founder with an AI plan on the table thinks the firm needs a full strategy before it can start

That has never happened once in a busy firm

Your firm decides whether AI works from the very first task it tries, based on one result (fast or slow, sharp or confused)... and then spends the next few quarters in leadership meetings building plans, tool lists and hiring ideas, so nobody has to admit the first real test never happened

Planning is never the driver

Planning is just the comfort blanket the team picks up AFTER the first result already decided how it feels about AI, so you have something serious to show your partner, your board or your clients

Once you understand that every AI rollout is split into a first-proof half (one task that works) and a plan half (everything after), you immediately see why most firms are still "looking into AI":

They only build half the bridge

Type 1 is the "Strategist" (most firms with a leadership offsite):
They open with the whole system: the vision, the vendor list, the new hire. Because nothing ships in the first month, client work wins every week, and the plan sits in a folder. Meanwhile a smaller rival sends its first client draft in hours, not days

Type 2 is the "Quick Tester" (most firms that "tried AI once"):
They try a free AI tool on the hardest problem they have, with no examples and no rules. The answer comes back confused, because the AI knows nothing about the firm. The team decides "AI doesn't work for us" and goes back to the old way, with no number to prove it either way

To get AI into real client work, you have to feed both halves in the exact order they fire:

Step 1: Win one task first (to prove it)
List ten tasks your team does every week. Score each 1 to 5 for how often it happens and 1 to 5 for how much it hurts, then multiply. Take the top score with a clear "good" answer, give AI ten past examples and your rules, and name one owner. Within a month, the team says: "do the next one"

Step 2: Hand the plan its proof (to scale it)
Now the plan has a number to stand on. Put the plain math on the screen: minutes per task before and after, tasks per month, hours back, and the next task on the list. Pick boring, frequent and clear, never the founder's hardest call

The firm already decided to trust AI the day one task worked, because of the proof. You just handed the plan the numbers it needs to grow without feeling like a gamble

[Nikhil | Building Systems @unik_47]

@i_mika_el cleann steps there.

[Nikhil | Building Systems @unik_47]

how to feed ai your firm's data with no data team

knowing ai matters is the easy part. knowing how to feed it your own records is what turns the plan into real work

the play: list your sources > pick one job > clean what it needs > connect it > test it with real questions

> first steps

- fix your plan first: one job, one team, one question the ai must answer well. vague plans stay ideas
- pick one source to start. connecting everything = nothing works by friday

> how to find the right sources

- write down every tool your team opens in a day, sort by how often
- open the last job that went wrong and note where the missing fact lived. that's your first source
- build 3 private lists:
- tier 1: the sources the job can't run without, like the client record and the fee sheet
- tier 2: the sources that make the answer better, like past work and call notes
- tier 3: the sources you can add later, like old email folders and archived chats
- give tier 1 a named owner so somebody notices when it breaks

> how much to learn a week

- one working test beats a stack of courses
- split: tier 1 first, then tier 2, then tier 3
- do it in two passes: one to connect the source, one to test the answers

> how to structure a test

- use one of these: a real client question, a past job, a known answer, or a known mistake
- the first question has to have a right answer you already know
- score each answer: right, half right, wrong
- fix the source behind each wrong answer, not the prompt
- never test with made-up questions nobody at the firm would ask

> what you get out of it

- a plan that leaves the slide deck and reaches the team
- staff who trust the answers because they saw them tested
- a clear list of which data is missing, by team
- you learn what your firm needs from ai before you pay for more tools

> what not to do

- don't hire a data team you can't keep busy
- don't buy another ai tool to fix a data problem
- don't run a test, see one bad answer, and drop the whole plan. the bad answer shows which fact is missing

> the catch

- it's slow. weeks of setup before the first job runs without help
- if you're a founder, you don't have the hours to learn data work on the side

> that's why i build ai-ready data for small firms

- the finding, cleaning, linking and testing happen around the tools you already use
- a monthly test with your team's own questions shows where the answers are weak, so they get fixed

advice: be the firm that runs one tested ai job, not the one with a folder of ai ideas

which ai plan at your firm stalled first?

start with one real client question today

[Nikhil | Building Systems @unik_47]

@kamhossieny lock it.

[Nikhil | Building Systems @unik_47]

AGENCY COMPANY BRAIN: DATA MAP
│
├── 1. Client Files (what you already own)
│ ├── Signed Briefs
│ │ ├── Goals & KPIs
│ │ ├── Target Buyer
│ │ ├── Budget & Scope
│ │ └── Approval Contacts
│ ├── Approved Work
│ │ ├── Ads That Ran
│ │ ├── Posts That Went Out First Try
│ │ ├── Landing Pages
│ │ └── Email Sequences
│ ├── Rejected Work
│ │ ├── Drafts Sent Back
│ │ ├── Client Notes on Each
│ │ └── The Reason Behind Each No
│ └── Brand Rules
│ ├── Tone of Voice
│ ├── Banned Words & Claims
│ ├── Logo & Colour Rules
│ └── Legal Lines
│
├── 2. Results Data
│ ├── Campaign Reports
│ ├── Ad Account History
│ ├── Best & Worst Performers
│ ├── Test Results (A/B)
│ └── Results Clients Let You Share
│
├── 3. Conversation Data
│ ├── Kick-Off Call Notes
│ ├── Weekly Status Calls
│ ├── Slack & Email Threads
│ ├── Complaints & Escalations
│ └── Praise in the Client's Own Words
│
├── 4. Money Data
│ ├── Retainer Terms
│ ├── Rate Cards
│ ├── Scope Changes & Who Approved Them
│ ├── Unbilled Hours
│ └── Margin per Client
│
├── 5. New Business Data
│ ├── Winning Proposals
│ ├── Losing Proposals
│ ├── Lost-Pitch Feedback
│ ├── Case Studies
│ └── Price Objections Heard
│
├── 6. How-We-Work Data
│ ├── SOPs
│ ├── Onboarding Checklists
│ ├── QA Checklists
│ └── Templates That Actually Get Used
│
├── 7. Where It All Sits Today
│ ├── CRM
│ ├── Project Tool
│ ├── Shared Drive
│ ├── Ad Platforms
│ ├── Inboxes
│ └── People's Heads
│
├── 8. Access Layer
│ ├── Who Sees Which Client
│ ├── Who Sees Money Data
│ ├── Freelancer Limits
│ └── What the AI May Read
│
├── 9. Connect Order
│ ├── 1st: Brand Rules + Approved Work
│ ├── 2nd: Rejected Work + Client Notes
│ ├── 3rd: Results Data
│ ├── 4th: Conversations
│ └── 5th: Money + New Business
│
└── 10. What You Ask It
├── "Write this the way this client approves"
├── "What did this client reject last time?"
├── "Which ad angle won for this type of buyer?"
├── "Which proposal won a pitch like this one?"
└── "Where is this client unprofitable?"

FINAL GOAL:
An agency that answers from its own files, not from a blank page

[Nikhil | Building Systems @unik_47]

@matt_kub copy that.

[Nikhil | Building Systems @unik_47]

@noacontent fact.

[Nikhil | Building Systems @unik_47]

how to stop sending pitches that read like everyone else's (agency owner edition)

in most agencies the team rents the same AI as every rival, so every pitch comes out in the same voice and price is the only thing left to compare

the fix is to show it your own taste once, then let the AI draft and you edit

how it works ↓

1. build the taste folder

> approved/: your last 15-20 pitches the client said yes to
> rejected/: the drafts you killed, with one line on why
> voice .md: how you sound, the words you use, the words you never use
> never-say .md: the claims and phrases you refuse to put in a pitch
> proof .md: the real work you're allowed to point to
> pushback .md: what clients question in a pitch and how you answer it

build them from material you already have. old decks, sent pitches, the emails where you told someone "this isn't us"

2. give it the brief, not a summary

after the client meeting, drop the brief or your raw notes into one file

briefs/client-2026-11-10 .md

3. the drafting prompt

"read voice .md, never-say .md and proof .md first. then read the 3 approved pitches closest to this brief. read rejected/ and avoid every reason listed there. draft a pitch for briefs/client-2026-11-10 .md in our voice. flag any line that could come from any agency. mark every claim you can't trace to proof .md with [check]"

4. you edit, it learns

every change you make goes into voice .md as one line
> "cut any sentence a rival could sign"
> "lead with the client's problem, not our process"

the next pitch reads voice .md first

5. nothing goes out without you

it drafts, you decide the idea and the price, you send

start with a pitch you already sent last month. give it the same brief and compare its draft with what you wrote

[Nikhil | Building Systems @unik_47]

BUSY managing partners will hear that you can:

- search every file you ever stored
- find any past answer at once
- stop paying people to redo work
- turn your archive into an asset
- answer clients faster than rivals
- grow without adding headcount

and keep scrolling

your AI can answer from your own archive, not from the open web

here's exactly how you run it:

a lot of firms pay every month to store years of client notes and past advice, then write the same answer again because nobody can find the old one

the archive isn't empty. it was built to store, not to use

the same AI that gives a vague answer from a blank chat gives your firm's answer from your files, because it finally sees what you told the last client who asked

you can try it in one afternoon: take a question you answered two years ago and ask it both ways

WHY THIS DATA

your team already answered most client questions at least once. the answers sit in old letters and email threads

and your chat history is blunt
"didn't we do this for another client?"
"does anyone have the old memo?"

senior people rebuild answers from zero while the bill for storing the first one renews every month

WHERE THE ANSWERS COME FROM

for two weeks, have each person write down every question they ask a colleague or search for

then note which tool held the answer each time

the questions show the value and the tools show where it hides

don't build an AI that searches everything

pick the three tools with the most questions, connect those first, and build the answers around them:

- file: the memo a partner wrote for a client two years ago -draft: "same question, new client, check what changed since"

- file: a long email thread where the team settled a scope question -draft: "the answer we agreed, the date, who signed off"

every answer is two assistants, same question. one sees only the question, one sees the three tools. both answers side by side

4 LANES, 1 BRAIN

- client questions
- new-hire questions
- proposal research
- renewal prep

one connected archive serves 4 kinds of question, instead of 4 people digging through the same drive

most days it saves a search

and the day it finds the answer a partner was about to write again, the storage bill finally earns something

and every month you wait, you pay to store answers your team still rewrites

i've said this for months yet here we are

so TONIGHT you need to start a question log for your team or you're cooked

[Nikhil | Building Systems @unik_47]

if i took over an agency tomorrow, here's the AI setup i'd build tonight so our work stops sounding rented

1. a taste file - holds every pitch a client said yes to. the AI reads it before it writes a word

2. a reject log - 3 lines on every draft we killed: what it said, why it failed, what we sent instead

3. a client lens - drop in a client, get the tone they approved, the words they hate and their last brief

4. a voice checker - flags lines in a new draft that any rival could have written

5. a pitch drafter - brief in, first draft out in our own style. a person still picks the angle

6. a proof finder - finds the past project closest to this pitch, so we show work instead of adjectives

7. a style keeper - one agent's whole job is reading what the team changes in each draft. any edit we make more than 3x becomes a rule the drafter follows from then on.

[Nikhil | Building Systems @unik_47]

@surlymochan easy w.

[Nikhil | Building Systems @unik_47]

Getting value from AI is much more than buying the right AI tool.

If you want AI that actually knows your firm, start understanding these areas:

1. The Data You Already Own

- Matter files and closed cases
- Ticket history and past fixes
- Client books and fee records
- Candidate notes and past briefs

2. Why It Sits Unused

- Files split across many tools
- Old versions next to new ones
- Duplicates nobody cleared
- No owner for each source
- Know-how stuck in inboxes
- Search that finds the wrong copy

3. Turning Files Into an Asset

- Mapping every source and its owner
- Clearing out duplicates and old drafts
- Picking one trusted version of each fact
- Sorting files by client, matter and project
- Linking related items across tools
- Keeping it live as new work comes in

4. Letting AI Use It

- Look up first, then answer
- Answers built on your own records
- A source next to each fact
- A plain "not found" when a fact is missing
- Read only before any write access

5. Checking the Value

- Real questions from each team
- Known answers to score against
- Right, half right and wrong
- The source behind each miss fixed
- The same test run again monthly
- Weak spots shown by team

6. Keeping It Safe

- Access rules for every client
- Fenced clients kept fenced
- A record of who opened what
- No client files in public AI tools
- Approval before a record changes
- No AI edits to money or contracts

7. What Unused Data Costs

- Questions asked again
- Work redone from scratch
- Fees set without past numbers
- New hires waiting on senior people
- Rivals who decide faster
- Storage that returns nothing

The real skill is understanding how all these parts work together.

Buying an AI tool is one thing. Running it on the files you already own is a completely different game.

[Nikhil | Building Systems @unik_47]

@thebaldogarza Treat every client correction as a system update, not another isolated instruction.

[Nikhil | Building Systems @unik_47]

Quick tip for agency owners:

If you want your AI drafts to go out as they are, so your team's time actually matters.

Your agency needs a "Client Context" layer.

It helps you:

give the AI each client's voice, buyer and rules
stop writers rewriting every draft
get drafts that sound like the client, not a stranger
It's the first thing I build into a company brain, because the rewrite hours show up there first

Here's one, written the way an agency owner would tell it:

"We used to rewrite AI drafts all afternoon

After the morning was gone on briefs

And even after the client had sent the same notes three times

Every draft had to be fixed before a client saw it

Every night my writers went home tired from fixing a tool we paid for

At some point, that had to change

I refused to keep paying people to correct a machine

A few weeks later, the AI was writing from our approved work

Now most drafts go out with light edits (the writers write again.. we've come a long way!)

Honestly, we didn't buy a better AI

What got us out of that rewrite loop can be in your client folders right now

The only difference is, you don't have to wait weeks to make that change"

Try it on your next client draft and let me know how it goes.

[Nikhil | Building Systems @unik_47]

@_niihhhall 🚀