The First 90 Days of a Custom AI Project With Ereos

When most people picture a custom AI project, they picture a long one. A discovery deck that takes a quarter to produce, an integration phase measured in seasons, a go-live date that keeps sliding, and a budget that grew teeth somewhere in month nine. That picture is real, and it is the reason a lot of mid-market operators decide custom AI is for someone larger than them. The eighteen-month enterprise project is a thing that exists. It is just not the only shape an engagement can take.

Watchtower is the AI system Ereos builds to run inside your operation, and we structure the first engagement so you see useful output in ninety days, not eighteen months. The work runs in three phases of roughly thirty days each, and they happen in a specific order for a specific reason. This is what each phase actually looks like, and why the sequence matters more than the speed.

Phase one: discovery (days 1 to 30)

The first month is not spent in a conference room looking at slides. It is spent watching the work. We sit with your leadership to understand what they are trying to fix, and we sit with the front-line operators who actually run the systems, because those two groups almost never describe the same business. We watch where the day catches: the report someone rebuilds by hand every Monday, the handoff that drops things, the question that takes four people to answer.

What we are looking for is the friction your team has stopped complaining about. Every operation has a layer of workarounds that became invisible because people learned to live with them, and that layer is usually where the value is. Alongside it we map the systems you run, the data trails they leave, and where the regulated data lives. Discovery ends with three deliverables you can hold: a written architecture, a phased scope, and a fixed-price quote for phase two. You decide whether to proceed knowing what it costs and what it builds.

• Time with both leadership and front-line operators, because they rarely describe the same business.
• A map of your systems, your data trails, and where regulated data lives.
• A written architecture, a phased scope, and a fixed-price quote for phase two.

Phase two: foundation (days 31 to 60)

The second month builds the thing nobody sees in a demo and everybody depends on. Watchtower deploys into your environment, inside your own Microsoft 365 tenant or Azure subscription, on your existing identity and security controls. The integrations get wired through supported APIs to the systems discovery identified. And then, before a single AI call reaches production, the governance stands up.

That means the scrubber that strips sensitive content before any model call, the audit log that records every interaction, the per-pipeline cost tracking that keeps spend predictable, and the user permissions that decide who sees what. All of it is built and running before the first pipeline produces anything. By the end of month two there is a foundation in your environment that can carry real work safely, even though no AI output has shipped yet. That is by design, and it is the part of the engagement people are most tempted to skip.

Phase three: first value (days 61 to 90)

The third month is where output starts. The first pipeline runs against your historical data first, where the stakes are low and the answers are known, so you can see how it reasons before you trust it on anything live. Then it tunes against your team's calibration: where your people agree with it, where they override it, and what those overrides teach it about your operation specifically. Only after that does it turn on for current data.

Inside the third month, the first weekly digest goes out the door. Not a chatbot and not a science project, but a short list of patterns worth your attention, each with a proposed next step, landing in the inbox of the person who can act on it. That digest is the proof that the foundation was worth building, because it ships safely, with every interaction logged and every cost tracked, the way the system was built to from day one.

Why governance comes before pipelines

The temptation in every AI project is to build the exciting part first. Get a pipeline producing output, show it off, and add the controls later once the value is proven. We sequence it the other way on purpose, because the other way produces a system you cannot defend. A pipeline running against regulated data without a scrubber, an audit log, and spend caps is producing risk alongside its output, and you cannot retrofit a record of what happened before the controls existed.

Governance first is the only sequence that lets you put AI into production and still answer the questions that matter: where did the data go, who decided, and what did it cost. We learned that order by building Watchtower for our own operation first, where skipping it would have been our own problem to clean up. You can read more about the controls themselves in our pieces on keeping regulated data safe and on predictable AI spend. The short version is that the boring phase is what makes the useful phase safe.

What you have at day 90

At the end of ninety days you are not holding a pilot or a proof of concept. You have Watchtower running inside your own environment, integrated with the systems you already use. You have at least one pipeline in production against your real data, tuned to your team's judgment, with the scrubber, audit log, and cost tracking all live. You have a weekly digest landing with the person who can act on it. And you have something your compliance officer and your CFO can both stand behind.

You also have a backlog. Discovery surfaced more than one phase of work could fix, and what did not make the first ninety days becomes a ranked list of pipelines to add over the following quarters, each scoped on a foundation that is already built and already paid for. The first engagement is the start of a system that grows, not a one-time install. If the eighteen-month version is what has kept you out of custom AI, this is the shape worth a conversation.