How we work

This page sets out how NUVAI puts AI to work inside an organisation. It is the public version of the playbook we follow on every engagement. It covers what we build and in which order. It also covers what we need from you, and what stays with you when we leave.

What we build

A skill is a short work instruction that your AI assistant follows each time a particular task comes up, such as drafting a proposal or checking a contract against your standard position. When an organisation has a set of these, with a register, a fixed running order and an owner for each, we call it a skill estate.

We build skill estates for organisations that get too little from a standard AI roll-out and have no need for a large integrator. We start small, on work that has to be done anyway.

Live within 72 hours

Your first skills and sidekick instructions are live in your own AI environment within 72 hours once four conditions are in place.

1. —Admin access to your AI environment, delegated to us for the duration of the work.
2. —Single sign-on for your users.
3. —A read-only role in the systems the skills need, such as your CRM or document archive.
4. —One person on your side who decides.

In practice, delays usually sit with these dependencies on the client side. A read-only role may be waiting on an IT ticket, or a decision on the next meeting. If access arrives late, the date moves and the scope stays as agreed.

If we miss it, we tell you why and what is still missing.

Principles

We work in your own AI layer. Everything we build lives in tools you already use and control. That means your AI workspace, its connectors, your CRM and your design system. The workspace might be Claude, ChatGPT or Grok Bot, for example. There is no NUVAI platform in between and nothing to migrate away from later.

Copilot? That one is a special case. Talk to us.

Read before write. Every connection to your systems starts read-only. Writing comes later, after a quiet period and your written approval. Even then each change is proposed first, written only once approved, read back and labelled as made with AI. Deleting, bulk changes, sending, and finalising quotes or invoices stay blocked.

A person decides and sends. Every output is a draft. Someone in your team reads it before it goes outside the organisation. No skill sends mail, signs a document or commits you to anything. Method, scope and advice stay with your people.

The assistant as sidekick. For thinking work, such as an approach or an analysis, the assistant asks for your own view first and puts one to three sharp questions. Then it stops. Simple tasks it just does, stating its assumption in one line.

No invented facts. A field without a source says "not found" or carries a placeholder. An estimate is marked as an estimate and says how it was made. Figures come from your own sources.

Your data stays yours. Client data stays in your environment, and nothing from one client is reused for another. A data-classification check runs before any other skill. Personal data is handled under a signed processing agreement.

We work on copies. Nothing is deleted. What has to go is archived. Files are shared with named people only, never through an open link.

The order we work in

Every engagement follows the same sequence. After each phase you decide, and nothing new starts without your written approval.

Kickstart (two weeks). We look at how your company works today and choose three to five processes with you. Those processes then run on AI in your own accounts, with a baseline alongside. The first one goes live within 72 hours of your users having access. The foundations go in at the same time, even if they are small: a skill register, the data-classification check, a one-page usage policy and a short session on AI literacy. You finish with a written view of what is worth building next, and what is not.

Baseline (five working days). For the process from the Kickstart that matters most and crosses several systems or teams. We measure how it runs today, and you get a decision on one page: buy it, build it or leave it.

Sprint (four weeks). The process from the Baseline, built and live in your own systems. It is accepted on real cases before we call it done.

Ongoing build. We work through the list that came out of the Kickstart, one team at a time, moving on when the first team runs well. We maintain the register, measure against the baseline each quarter and add skills when you ask for them.

After the Kickstart we take one process at a time. A skill people open every day is worth more than five they forget about.

How a skill goes live

Each new skill starts as a pilot with two decision-makers, on at least three real cases. It moves on to two or three champions once the output keeps to the agreed length, contains no invented facts and passes the privacy check. After that it opens to everyone. You decide at each step and we advise.

Every skill has a version number, an owner, a review date and test cases. At least one test case uses fictional data, so we can test before any of your data is involved. To switch a skill off we set it to unavailable. Nothing is deleted, so there is always a way back.

Reworking a skill from your best work

Sometimes a skill works technically and its output still falls short of your best work. We don't rewrite it on instinct. We ask your own AI to describe the gap.

1. —You give us two to four of your strongest examples. In your AI environment, the assistant reads them next to the current skills and your design system.
2. —It sets out the differences under fixed headings (structure, length, tone, visuals and the building blocks your examples use) and proposes concrete changes per skill file.
3. —We list the choices only you can make and number them. We suggest a default for each, and you decide.
4. —We build a test version with those defaults and run it on a fictional brief, with no real client data and no amounts.
5. —You go through the choices with the test output beside them. The new version goes live after your written approval, with a changelog. The previous version is kept.

Some things never carry over, such as typos, outdated dates, colours outside your palette or anything taken from another client's work. A rework may never weaken the privacy and review rules of a skill.

Onboarding every user

A new user should be able to get started in the first week without someone sitting next to them. Each client gets an onboarding checklist, written directly to the user and short enough to tick off.

Every step has a link and a screenshot. Links go to the official download page or help article, or straight to the right screen in the app, and we check them on the day of delivery. Screenshots come from your own environment, taken with an ordinary user account and cropped so that no names, mail addresses or chat titles show.

The checklist takes the user through signing in, the desktop and mobile apps, a short tour and the first connections, which are read-only. It explains meeting notes and consent, what each skill does with an example prompt, and a plan for the first week. It closes with good prompts drawn from your own work, the ground rules and where to find help. Each skill carries a status label, such as working, pilot, being reworked or not yet available, so nobody relies on something that isn't ready. Some settings we cannot check ourselves, for instance whether training on your data is switched off. Those stay flagged for your admin to confirm before the checklist goes to the whole team.

Routines with a human owner

A routine is a fixed question the assistant answers at a fixed moment, using data from your systems. Think of the pipeline every Monday morning, or the relationships nobody has spoken to in three months. Each routine has one owner who reads the output and acts on it. Routines start read-only, with automatic approval switched off. They prepare lists and draft texts. They don't send anything, and they don't place drafts in your mailbox or CRM. If a routine goes unread three times in a row, we pause it.

Illustrative example: customer onboarding

This example describes the mechanism. It is not a client result.

Take a B2B brand that sells starter kits to professionals. The period that matters most runs from the first purchase to the first repeat order. Our working assumption, to be measured per client, is that customers who get help with the product in their first weeks are more likely to come back.

Every morning the assistant prepares one list with the next best action for each customer in that period. For each customer it shows:

• —who they are and how many days have passed since the order;
• —which step is due, for example a welcome, a check-in after the first week, a tip or an event invitation a few weeks in, or a gentle reorder prompt after a month;
• —which channel to use, based on where the customer actually talks to the brand, whether that is mail, a messaging app or social media;
• —why this action, and why now;
• —a ready draft in the brand's tone and the customer's language.

Someone in the team goes through the list, adjusts or skips a line where needed and sends each message personally. For a customer with a complaint, a return or an open question, the flow pauses and the case is flagged. A log per customer makes sure nobody gets the same step twice. All connections are read-only, covering shop orders, the CRM and the conversation history.

Renout van Hove spent years building next-best-action programmes on customer data platforms as founder of GrowthAgent, long before generative AI. Back then this took a customer data platform and a marketing automation suite, plus an integration project to connect them. Today the same logic runs in your own AI layer, on top of the CRM and shop data you already have, and each draft is written for one person.

Measuring value: paid, owned and earned

We look at value in three layers.

• —Paid is what you rent, such as the AI workspace and the CRM. We track the cost per seat and how many seats are active each week.
• —Owned is what we build and you keep. That includes instructions, skills, templates and the register. We track how often skills are used and how long a core process takes before and after.
• —Earned is what the system builds up by running, such as win and loss outcomes linked to proposals, a decision log and a measurement history. It only has value if the data is recorded consistently. That is why the CRM foundations come before any analysis that depends on history.

We measure before we build, together with you, on recent and current cases. Afterwards we measure again with the same method. Every result states the number of cases behind it.

What you own when we leave

You should be able to carry on without us. At the end of a phase, or of the engagement, you receive:

• —the skill register, with owner, version, data used and risk class for each skill;
• —every skill as files in your own environment, with your content kept separate;
• —the approved workspace instructions and the one-page usage policy;
• —instructions for changing, testing and reloading a skill, with the test cases;
• —a list of every connection, with its owner, its rights and what happens if it fails;
• —the baseline and follow-up measurements, with method and number of cases;
• —open points, each with an owner and a proposed next step;
• —a list of our access rights and confirmation that they have been withdrawn;
• —an overview of the licences you rent, in whose name and when they renew.

Skills are kept portable, so you can load them into another model or workspace if you change supplier. The generic logic behind our templates remains ours and is licensed to you. Everything specific to your organisation belongs to you.

What we don't promise

We don't promise agents that run your work without a person. Every skill on this page prepares work, and a person decides.

We don't promise a percentage up front. We measure before and after, show you the difference and the number of cases behind it, and leave the judgement to you.

Getting started

Most engagements begin with a Kickstart. If you'd rather talk it through first, get in touch.