Forward Deployed AI from QuantSpark

How to roll out Claude, ChatGPT or Copilot across a company

Last updated 21 November 2026

To roll out an AI assistant across a company, put licences in a configured company account, connect the assistant to company data read-only first with a named approver, measure who uses it, and build the first tools with an engineer inside the team. The steps are the same for Claude, ChatGPT and Microsoft Copilot, which are assistants companies roll out. QuantSpark's method maps each step to one of five levels.

What to set up first

Set up the account before the people. Work out which assistant licences the business needs and who should hold them, so people work in a configured company account rather than personal ones. Load your own terms, templates and house rules into it, and write the prompts and reusable skills that carry them. Write a short use policy at the same time: what data may go into the assistant, and who to ask when unsure. This is level 1, Use, and its first six-week sprint ends with a configured assistant and a usage baseline.

Who needs which licence

Licences are easy to buy and hard to turn into daily use. Start with the teams that have a named job for the assistant, and record who holds a licence, who uses it and how often. That record is the baseline for every later measure. Official figures show why depth matters more than access: in September 2026, 38% of UK businesses with 10 or more employees reported using AI technologies (ONS, 8 October 2026), and limited AI skills, expertise or knowledge affected 67% of businesses that had adopted AI (DSIT, October 2026). Those figures measure whether a business uses AI at all, not how deeply.

Connect the assistant to company data, read-only first

An assistant is most useful when it can read your files, mail and systems. It is also where rollouts go wrong, because it makes any file shared too widely easier to find. A level 2 engineer connects the assistant to files and mail, fixes files shared too widely, and gives it read-only links to finance or sales systems, with a named approver for any change. Our engineers connect your tools through the Model Context Protocol, the open standard for linking AI to company systems.

Measure use

Agree the measure before the work starts. Each sprint goal has a measure, a baseline and a target, agreed at kick-off and reviewed in week six of its sprint. If a team's use is not rising against the baseline, adoption has stalled there, which is the sign to bring in level 2 to read usage figures and plan how more teams start.

Build the first tools

When the next opportunity on your list is a single-task agent, the work moves to level 3. An agent is software that carries out a task with AI, with checks. The engineer builds a tool or agent that routes exceptions to a person, and checks it against at least 20 known answers before anyone relies on it.

Move a pilot to production

When a tool or agent is going into production, a level 4 senior engineer joins to own its quality, because tests, monitoring and security review matter there. Gartner reported that at least 50% of generative AI projects were abandoned after proof of concept by the end of 2025 (Gartner, 26 January 2026). The steps that prevent it are in AI pilot to production.

Why put an engineer inside the team

Andreessen Horowitz writes that "even model providers like OpenAI and Anthropic have been hiring forward deployed engineers and solutions engineers to win at the enterprise" (a16z, 4 June 2025). Forward Deployed AI brings the same embedded model to mid-sized and private-equity-backed businesses. For the role itself, read what a forward-deployed engineer is.

Claude, ChatGPT or Copilot: choosing, and running more than one

Choose by where your files and mail live, which tools your teams already use and what your security team will approve, not by a ranking. Running more than one assistant is possible: the steps above apply to each, and one use policy can cover all of them. Our engineers are model-agnostic. They build on the models and agent platforms you already use, from Anthropic, OpenAI, Google or open models, and connect your tools through the Model Context Protocol, the open standard for linking AI to company systems.

Tool-specific plans: Claude Enterprise rollout and ChatGPT Enterprise rollout.

Where each step sits in the five levels

StepLevelWho does it
Accounts, licences, prompts1 UseLicences, prompts, simple automations
Connect to company data2 ConnectConnects tools, runs adoption
Build the first tools3 BuildBuilds simple tools and agents
Pilot to production4 ShipSenior engineer, owns quality
Roadmap for the board5 OperateAI operating partner

Questions, answered

How do we roll out an AI assistant across a company?

Put licences in a configured company account, write a short use policy, connect the assistant to company data read-only first with a named approver, and record who uses it and how often. Then build the first tools with an engineer inside the team, and move any pilot that matters into production with tests and review.

Claude, ChatGPT or Copilot: how do we choose?

Choose by where your files and mail live, which tools your teams already use and what your security team will approve. QuantSpark is model-agnostic, so the engineer works with whichever you choose. The rollout steps are the same for each, and the choice does not change the order of work.

Can we run more than one assistant?

Yes. The rollout steps apply to each assistant, and one use policy can cover all of them. Keep one baseline of who holds a licence, who uses it and how often, per assistant, so you can see which one people actually use before you widen any of them.

What should an AI use policy say?

A short use policy says what data may go into the assistant, who approves a new use, and who to ask when unsure. Write it before launch, alongside the account set-up, and have your administrators apply it as settings in the tool where they can, not only as a document.

How do we measure adoption?

Start with a usage baseline: who has a licence, who uses it and how often. Each sprint goal then has a measure, a baseline and a target, agreed at kick-off and reviewed in week six of its sprint. If use is not rising against the baseline in a team, adoption has stalled there.