# Build Your AI Operating Model, Starting From Nothing

> You own AI delivery and nothing is mapped. Gather what your organisation has already written down, then build your operating model from it in Claude, in an afternoon. Free, no engagement.

Source: https://kowalah.com/resources/templates/map-your-ai-operating-model

## Overview

- **Who it's for:** The person who owns AI delivery inside an organisation with nothing mapped yet: AI Operations Leads, internal Forward Deployed Engineers, AI enablement leads, and anyone running discovery on their own
- **When to use it:** Your first week or two in the role, before you have built anything, when you need somewhere for what discovery turns up to land
- **What you get:** One business function mapped by the end of an afternoon, built from documents your organisation already has rather than from memory, in the tool you already work in

## How to Use This Guide

1. **Gather before you start.** An hour finding the org chart, process maps and checklists your organisation already has saves a day of describing them from memory
2. **Install the plugin.** Ten minutes in Claude. On Team or Enterprise, your Claude owner makes it available first. The install guide is linked in step one
3. **Work through the conversation.** The prompts below are the ones to use. Adapt the specifics to your business, keep the sequence
4. **Map one function, not the company.** Finish one function end to end before starting the next. It teaches you the shape
5. **Come back for the next one.** Each function takes less time than the one before. Coverage matters more than depth early on

***NOTE**: This is a walkthrough, not a document to fill in. Work through it with
the plugin open in Claude.*

# Build your AI operating model, starting from nothing

You own AI delivery inside a business. Nothing is mapped. You have people asking
what you are going to do, a vague sense of where the value might be, and a
growing pile of notes from discovery conversations with nowhere to put them.

This is the shortest route out of that. By the end of an afternoon you will have
one business function mapped properly: its processes, the steps inside them,
where the time goes, and the first use cases worth doing.

The part people get wrong is starting from memory. Most of what you need has
already been written down by someone in your organisation, for a completely
different reason, and Claude can read it. That is the real reason to do this in
Claude rather than in a platform: it is already sitting next to your documents,
and it will take material you would not upload anywhere else.

## What you will have at the end

- An org tree, built from your own org chart rather than from memory
- One business unit filled in, with the processes it owns
- One of those processes broken into steps, with the current state honest
- A list of what the documentation does not tell you, which is your agenda for
  the next conversation with that team
- Two or three opportunities logged against real processes rather than a
  wishlist
- A clear sense of how long the next function will take

## What you will not have

A complete map of your company, and you should not try. A first pass across five
or six functions is a couple of weeks of part-time work alongside your discovery
conversations. Trying to do it in one sitting produces a map you do not believe
and cannot defend.

---

# Before you start: gather what already exists

The instinct is to sit down and describe your business from memory. Resist it.
Most of what you need has already been written down by someone else, usually for
a completely different reason, and reading it is faster and more accurate than
recalling it.

An hour of gathering saves a day of guessing, and it produces a map other people
recognise rather than one they have to be argued into.

## Where the material already is

| Source | What it gives you | Who has it |
|--------|-------------------|------------|
| Org chart export from the HR system | The unit tree, headcount, who leads what | HR or People Ops |
| The last big systems implementation | Process maps, as-is and to-be | IT, or whoever ran the ERP or CRM programme |
| ISO, SOX or internal audit documentation | Processes with controls and named owners, written to a standard | Risk, compliance, finance |
| IT service catalogue | What each function actually runs on | IT |
| Job descriptions | Tasks and responsibilities, function by function | HR |
| Onboarding and training material | How the work is supposed to be done, step by step | L&D and team leads |
| The month-end close checklist | A finance process already broken into steps | Finance |
| RPA or automation documentation | Someone already identified the repetitive work | IT or the automation team |
| A consultant's operating model review | Often most of a map that nobody actioned | The executive who commissioned it |
| Recorded discovery calls | Your own conversations, already transcribed | Your notes or meetings tool |

Nobody in your organisation calls these an operating model. Together they are
most of one.

## Asking for it

The specific ask works where the general one does not. "Do you have a process map
or an SOP for the month-end close?" gets a yes more often than you would expect.
"Tell me how Finance works" gets you a meeting in three weeks.

Ask for whatever exists rather than whatever is current. A three-year-old process
map beats a blank page, because correcting a document is faster than authoring
one, and people engage with something wrong far more readily than with something
empty.

---

# Step 1: Install the plugin

Ten minutes, once it is available to you.

On a Claude Team or Enterprise plan, your Claude workspace owner has to go first:
they make Kowalah available under **Organization settings > Plugins & skills** and
add its connector under **Organization settings > Connectors**, because members
cannot add it themselves. Until they do, Kowalah will not show in your **Discover**
tab. The [install guide](https://docs.kowalah.com/integrations/plugins/kowalah-plugin)
has the owner's steps, so send them the link. On Pro or Max, you do it all yourself.

Kowalah is in Claude's plugin directory. [Open the listing](https://claude.ai/customize/plugins/id/a4be1a55-fa04-473d-90f0-a5e020071f3a%40anthropic-plugin-directory) and click
**Add**, then open the plugin's **Connectors** tab and connect Kowalah, because
adding a plugin does not sign you in. In Claude Code it is two commands. Full
instructions are in the
[install guide](https://docs.kowalah.com/integrations/plugins/kowalah-plugin),
and the [plugin page](/platform/plugin) explains what you are getting.

When you connect, **sign in with your work
email.** An email address Kowalah does not already know is fine, and you do not
need to be invited first.

# Step 2: Check where you landed

This matters more than it looks, and it takes one prompt.

```text
I've just connected Kowalah. What can you see?
```

Your email domain decides where you land, and there are two outcomes:

**A new organisation, created for you.** Nobody from your company is in Kowalah
yet, so you get your own organisation with you as its admin and an empty model.
This is the normal case if you are reading this guide, and it is what the rest of
the walkthrough assumes. You can author everything.

**Your company's existing organisation.** Your company already works with
Kowalah, so you have joined theirs, usually as a member. There will be a map
already. Stop and talk to whoever runs the programme internally before adding
anything, and ask them to raise your access if you need to edit.

Claude can tell the two apart. The reliable tell is the shape rather than the
name: exactly one organisation where you are `admin`, a model with a single unit
and no processes, and no vision. A new organisation is *named* from your email
domain, so it will carry your employer's name and look convincing.

# Step 3: Point Claude at what you gathered

This is the reason to do this in Claude rather than in a platform: Claude is
already sitting next to your material, and it will take documents you would not
put anywhere else.

**Use your connectors.** If Claude already reaches your Google Drive, OneDrive,
SharePoint, your CRM or your call transcripts, it can go and find things without
you hunting for them.

```text
Search our SharePoint for anything that looks like a process map, an SOP
or an org chart for Finance, and tell me what you found before you read
any of it in detail.
```

**Upload the rest.** Anything not connected, or anything sensitive enough that
you would rather it were not in a shared drive, goes straight into the
conversation.

## What lands where

This is the question people ask before they upload anything, so here it is
precisely.

The document stays in your Claude conversation. What gets written into Kowalah is
the **structure** you and Claude derive from it: unit names, process names, the
steps, what each step takes in and puts out, and how it runs today. There is no
upload path into the model at all.

So a sensitive SOP, a draft reorganisation chart or a set of call transcripts can
inform the map without ever becoming part of it. That is a different trust
decision from loading the same file into a vendor's platform, and it is usually
the difference between mapping the real process and mapping the sanitised one.

# Step 4: Set the context once

Tell Claude who you are and what you are doing, at the start of the thread. It
changes every answer that follows.

```text
I'm the AI Operations Lead at [company]. We're [size] people in [industry].
Nothing is mapped in Kowalah yet. I've got our org chart, the finance
close checklist and the process maps from our ERP rollout. I want to map
one function properly today rather than sketch the whole company.
Where should we start?
```

If you already know which function to start with, say so. If you do not, a good
default is the one where you have already had the most discovery conversations,
because you will be able to answer the follow-up questions without guessing.

# Step 5: Build the org tree from the chart

Business units are the top level of the map: the divisions or functions the
company is actually organised into. If you have an org chart, this takes one
prompt rather than twenty minutes of typing.

```text
Here's our org chart. Build the unit tree from it, top two levels only,
with the leader and approximate headcount for each. Ask me where it's
ambiguous rather than guessing.
```

**Top two levels is usually the right depth.** Going deeper produces a tree you
will spend the afternoon maintaining instead of a map you can use. You can always
push into a unit later when you map its processes.

**The org chart is the reporting line, not how work flows.** Order to cash runs
through Sales, Operations and Finance, and the chart will not tell you that. Say
so when it comes up, and let the processes cross units rather than forcing them
into one.

Add the siblings you are not mapping today anyway. A tree with eight units and
one of them filled in tells a better story than a single unit on its own, because
it makes the gaps visible.

If you have no org chart to hand, describe it instead:

```text
Add Finance as a unit. It reports to the CFO, has around 40 people,
and covers financial control, accounts payable, accounts receivable
and financial planning.
```

# Step 6: Map a process from its documentation

The part that carries the value. Start from the document rather than from memory.

```text
Here's the month-end close checklist. Turn it into a process under
Finance with its steps. For each step capture what goes in, what comes
out and who does it. Flag anything the document doesn't tell you.
```

**The flags are the most useful thing you get back.** A checklist tells you the
order of operations and almost never tells you how long a step takes, who does it
when that person is on holiday, or what happens when it fails. That list of gaps
is your agenda for the conversation with the team, and it is a far better opening
than an empty page.

## Where documentation runs out

Some processes are genuinely not written down anywhere, and every documented one
has undocumented reality around it. That is where you talk it through.

```text
The checklist stops at "reconcile intercompany". Walk me through
capturing what actually happens in that step. I'll describe it and
you capture it. Ask me what you need.
```

Describe it the way you would explain it to a new joiner: what happens, who does
it, what it takes, and what goes wrong. Claude will ask about what you skip, and
the questions are where you find out what you do not know yet.

**Record the real version, not the documented one.** A map of how things are
supposed to work is worth nothing. If step four is done in a spreadsheet that one
person maintains and the SOP says it happens in the ERP, record the spreadsheet.
The gap between the document and the practice is the most valuable thing you will
find all afternoon, and those are exactly the steps where the value is.

# Step 7: Find where it hurts

With the steps captured, ask the question you came for.

```text
Looking at these steps, where is the cycle time actually going, and which
steps have never been measured?
```

```text
Which of these steps could realistically change this quarter, and what
would we need to prove first?
```

Then log what is worth doing, against the real process rather than into a
separate list:

```text
Log the reconciliation step as an opportunity. Impact is high, we think
it's three days of the nine. Confidence is medium, we haven't measured it
properly yet.
```

Scoring honestly matters more than scoring high. Confidence is the number people
inflate, and an opportunity you have not measured should say so.

# The rest of the afternoon

You have a shape now. Two useful ways to spend what is left:

**Go wider.** Add the processes each of your other units owns, without the steps.
Half an hour of this gives you the coverage view, which is what tells you where
to concentrate next. Value concentrates in a small number of domains and you
cannot see which until you have the breadth.

**Go back and check.** Take what you mapped to one person who does the work and
let them correct it. Being corrected is worth more than being approximately
right, and it is the fastest way to establish what you are actually for.

```text
Show me the coverage across all units, and where the gaps are.
```

---

# What to do next

**The next function.** It will take about half as long. Repeat until you have a
first pass across the functions that matter, then stop and build something.

**Pick your domains.** Once four or five functions are mapped, choose the two or three
where AI moves a number the business already cares about, and concentrate there.
The [first 90 days guide](/resources/templates/ai-operations-lead-first-90-days)
covers how to choose and how to defend the choice when everyone wants to be
first.

**Bring the process owners in.** This is where solo mapping stops being enough. A
map you built alone is your account of how the business runs. It becomes true
when the people who own those processes confirm their own and keep them current. That is a different kind of
artefact: it stops being a document you maintain and becomes an operating model
the organisation maintains.

---

# Implementation notes

**Map in the conversation, not afterwards.** The point of doing this in Claude is
that it happens while you are still talking to people about how their work runs.
A discovery conversation in the morning and a mapping session in the afternoon
loses half of what you heard. Map as you go.

**Gathering is the hour that pays for itself fastest.** It is also the part
people skip, because it feels like preparation rather than progress. A map built
from documents lands differently in the room: you are showing a team something
assembled from their own material and asking where it is wrong, rather than
showing them your interpretation of a conversation.

**Coarse beats complete.** The model is a decision-making instrument, not
documentation. If a level of detail would not change which use case you pick
next, you do not need it yet.

**Record who told you.** When you capture a process or log an opportunity, note
who described it. It costs nothing and it means you can go back to them when it
reaches the top of the list, which is how a backlog builds advocates.

**Your first map will be wrong.** Write it anyway. Correcting a draft in front of
a team is a conversation; asking them to describe their work into a blank page is
an interview. The first is faster and people prefer it.

## Frequently Asked Questions

### Do I need a Kowalah engagement to do this?

No. Installing the plugin and signing in creates your account, and if nobody from your company is there yet it creates an organisation for you with you as its admin and an empty model to fill. You can author the whole map on your own: units, processes, steps, vision and opportunities. There is no form, no sales call and nothing to procure. If your company already works with Kowalah you will join their organisation instead, as a member, and an admin there can raise your access.

### What is an AI operating model, in plain terms?

A map of how your organisation actually works and where AI could support it: the business units, the processes each one owns, the steps inside those processes, and what each step looks like today against what it could look like. It is deliberately coarse. You are not documenting the business to the last detail, you are building a shared picture good enough to decide where to point an AI programme and to defend that choice when someone asks.

### How long does this take?

An afternoon gets you one function mapped properly and a clear sense of the shape of the rest. The second function takes about half as long, because you stop deciding how to describe things and start just describing them. A useful first pass across five or six functions is a couple of weeks of part-time work alongside discovery conversations, not a project.

### Should I map everything before I build anything?

No, and mapping the whole company before shipping is a common way to spend a quarter with nothing to show. Map a function, find the use case inside it worth doing, build that, then map the next. The point of the model is to make the next decision better, not to be complete. Coverage matters more than depth early on: a shallow map of six functions tells you more about where value concentrates than a deep map of one.

### What if I get the model wrong?

You will, and that is the intended way to use it. A first pass written from your own understanding is a draft you take back to the team and get corrected. Being corrected in front of the people who do the work is the fastest way to get it right and the fastest way to establish that you are interested in how they actually operate. The model is easier to argue with than a blank page.

### What actually gets stored in Kowalah, and what stays in the chat?

The document stays in your Claude conversation. What gets written into Kowalah is the structure you and Claude derive from it: unit names, process names, the steps, what each step takes in and puts out, and how it runs today. There is no upload path into the model at all, so a sensitive SOP, a draft reorganisation chart or a set of call transcripts can inform the map without becoming part of it. That is a different trust decision from loading the same file into a vendor's platform, and it is usually the difference between mapping the real process and mapping the sanitised one.

### What if we have no process documentation at all?

You almost certainly have more than you think, just not filed under that name. Look for the last big systems implementation, ISO or audit documentation, the IT service catalogue, job descriptions, onboarding material, the month-end close checklist, and anything an automation team wrote down. Nobody calls those an operating model and together they are most of one. Where there is genuinely nothing, talk the process through with Claude instead: it is slower and it still works, and the guide covers how.

### Should I use my existing connectors, or upload files?

Both, and the choice is about where the material lives rather than about capability. If Claude already reaches your Google Drive, OneDrive, SharePoint, CRM or call transcripts, let it search rather than hunting for files yourself. Upload anything that is not connected, or anything sensitive enough that you would rather it were not sitting in a shared drive. Either route gets the same result in the model, because what lands there is structure rather than the source document.

### Can I do this if I am not technical?

Yes. This is a conversation, not a configuration exercise. You describe how a process runs and Claude captures it into the model. The technical half of the AI Operations Lead role matters later, when you are building agents and wiring up connectors. Mapping is the half that needs someone who can hold a credible conversation with a finance director, not someone who can write Python.

### What happens to what I build if my company later works with Kowalah?

Talk to us before you get far, because the two organisations are separate today. If your company's domain is not yet registered with Kowalah you are building in your own organisation, and a later engagement would set up the company's organisation alongside it rather than absorbing yours automatically. It is a solvable problem and much easier to solve early than after six functions are mapped.

### When do I stop doing this on my own?

When the map starts making claims you cannot personally stand behind. A model built by one person is that person's account of how the business runs. It becomes true when the people who own those processes confirm their own and keep them current, which is also the point at which it stops being a document and starts being an operating model. That is what adding the rest of the organisation is for.

### Does this replace the Vision Map Workshop?

It is the same first move, done alone and at your own pace rather than with a facilitator and your leadership team in a room. What you cannot do on your own is get eight executives to agree on the same picture in five days, and for an organisation where alignment is the actual problem, that is the whole value. If you are one person trying to get started, begin here.

---

Free and ungated, from Kowalah, a UK Anthropic implementation specialist. More templates and skills: https://kowalah.com/resources/templates. Talk to us: https://kowalah.com/contact