The first 100 days: a value-creation map for AI in a portfolio company

Why a 100-day plan

The first hundred days after close decide whether AI in the company becomes a value lever or just another line of IT spend. Without a plan, teams chase whatever looks best in a demo, and a year later they have three working prototypes with no effect on results. With a plan, those same resources go where they actually move EBITDA at an acceptable level of risk.

The premise is calm and concrete: this isn't about autonomous systems replacing people, but about AI agents automating repetitive slices of work under human oversight. The stakes are throughput and margin, not a slogan about a revolution.

Weeks 1–2: inventory and sequencing

Start with a map, not a tool. List the company's processes that are repetitive, text- or data-based, and expensive in people's time. Score each on two dimensions: EBITDA impact and risk of error. The sequence follows a simple rule — high impact and low risk first.

EBITDA impact	Risk of error	Decision
High	Low	Do this first — it's your first value
High	High	Do it, but with guardrails and a human in the loop
Low	Low	Leave for later, or use as a quick test
Low	High	Skip it — the cost of oversight outweighs the gain

The demo is not a criterion. A process that looks good on stage but touches little money goes to the bottom of the list.

Weeks 3–6: the foundation, before you scale anything

The temptation is to deploy right away. That's a mistake — without a foundation, you scale your mistakes along with everything else. Three things must be in place before the first process goes into production.

• Governance. AI governance answers the questions of who owns what, how decisions are logged, and where a human approves high-stakes actions. It's a prerequisite, not an add-on.
• Measurement. Without evaluation on a fixed set of cases, you can't tell improvement from regression. You put measurement in place alongside the first process, not after it.
• Vendors and data. Decide which models you run on, on what terms, and what data is allowed into them.

This foundation is also the first candidate for centralization — more on that shortly.

Weeks 6–12: the first process in production

You take process number one off the list and build it as an agentic workflow: clearly described steps where the agent does the repetitive work and a human approves the output wherever the stakes are high. The goal of the first deployment isn't a flashy demo but a countable result — shorter handling time, higher throughput, lower unit cost.

Once the first workflow is running and has numbers behind it, you add more. With several agents working together, agent orchestration comes into play: coordinating who does what and in what order. That's a step for a mature deployment, not for the first week — orchestration on an empty foundation only multiplies points of failure.

Centralized or per company

This is the most common question for a fund with several portfolio companies, and there's no single answer. There is, however, a clear dividing line.

1. Centralize what's common. Governance, evaluation, vendor contracts, shared components and team knowledge repeat across companies. Done separately in each one, they multiply cost and let the standard drift apart.
2. Leave the company what's local. Processes, data and integrations depend on the specific company. Steered centrally from a distance, they're slow and detached from reality.

Pure centralization is slow and deaf to each company's specifics. Pure decentralization repeats the same costs in every company and makes comparison harder. The practical arrangement is a shared core plus local deployment: the fund supplies the standard, the tools and the measurement; the company supplies the processes and ownership of the result.

Operator's rule: don't ask "where could we use AI." Ask "which repetitive process, if it sped up, would change the result" — and start with that one, with measurement from day one.

What should be standing after 100 days

By the end of the hundredth day you don't need ten deployments. You need one process in production with a countable result, a standing foundation of governance and measurement, and an ordered list of the next projects ranked by EBITDA impact. That's enough to show the board not a promise but proof that AI creates value in this company in a way that can be repeated.