The 90-Day AI Implementation Roadmap: From Scope to Production

The pattern I keep seeing is this. Someone reads a McKinsey deck, books a kickoff, then spends three weeks running "AI opportunity workshops" while the actual workflow stays unwritten on a whiteboard. By the time anyone touches an API, half the runway is gone and the team still cannot name a single ticket type they want the system to handle. Most of the public 90-day templates make this worse by treating discovery as its own deliverable. It is not. Discovery is the cost you pay to write one paragraph: which user, which workflow, which metric and what number that metric is at today. If you can't write that paragraph in five working days, you do not have a project, you have a wish.

MIT's Project NANDA reported in July 2025 that 95 percent of enterprise GenAI pilots showed no measurable P&L return. The common thread in the failures was scope: pilots that explored instead of shipping one workflow to production.

I write the scope artifact in a Linear doc or a Notion page, never in Slack. The medium matters because Slack lets scope drift in silence. A page has a URL, a last-edited timestamp and a comment thread. That audit trail is what stops the founder from quietly adding "and also voice" in week 6. If you want a deeper map of the broader category landscape before scoping, the piece on agents vs chatbots vs copilots is the lens I use to talk operators out of building the wrong thing.

“

Discovery is the cost of writing one paragraph. Not a deliverable.

The 2026 stack for a first production AI workflow is small. One frontier model API (I default to Claude Sonnet 4.5 via the Anthropic Agent SDK because the tool-use loop is the cleanest in the market right now, with GPT-5 as fallback for tasks where the cost of an extra round trip is fine). One eval tool, usually LangSmith, because it writes traces you can actually read back at 11pm three months in. One vector store if the use case needs retrieval – pgvector when you already run Postgres, Pinecone serverless when you want to skip ops, Weaviate when you genuinely need hybrid search and are willing to pay for it in operator time. One workflow runner, often n8n self-hosted on a Hetzner CCX33 box at roughly €60 a month, for the glue around the model. That is the whole list.

What I would not do in the first 90 days. I would not stand up Modal or Replicate for inference hosting if a hosted API solves the problem, because you are paying with your team's attention. I would not write a custom RAG framework when the Vercel AI SDK plus pgvector covers 80% of cases. I would not pick a vector store on the strength of a benchmark blog post written by the vendor of that vector store. Boring beats clever for the first ship. The a16z emerging architectures piece is still the most honest single map of the space and worth a read before you sign any procurement contract, even where it has aged. If budget realism is what you want, the breakdown in what AI implementation actually costs matches the prices I see in invoices.

The schedule below is the one I run when the use case is already chosen. If the use case is not chosen, the schedule does not start. I have watched too many teams launch a 90-day clock against an undefined target and then act surprised at day 70.

I see operators staring at a $499/month vertical SaaS and a $40K custom build and concluding the SaaS wins. On a 90-day horizon, it does. On an 18-month horizon, the SaaS lock-in tax and the inability to change the prompt or the retrieval logic compounds into a six-figure regret. The honest comparison is below. Read it as a starting point, not a verdict. Your specific workflow may genuinely live inside a Zendesk macro forever, and that is fine.

If you reach day 90 and three or more of the right column apply, the most useful thing you can do is stop, freeze the project for a week and re-scope. I know that is the advice nobody wants to hear in month three. It is still the cheaper move. The full audit framework I use on stalled programs is closer to what is in measuring AI ROI – the gap between "we are doing AI" and "the AI is doing the work" is usually a measurement gap, not a model gap.

Deloitte's June 2022 automation survey put the average payback period for scaled programs at about 22 months, up from 16 a year earlier. A 90-day scope exists precisely to beat that curve.

The published 90-day playbooks tend to come from three places. Consultancies selling six-month engagements. Vendors selling the platform that conveniently sits at the centre of the diagram. And LinkedIn personalities who have never owned a production on-call rotation for an LLM system. None of these sources have to live with the consequences of a bad week-2 decision, which is why their advice rounds to "do everything carefully" rather than "skip discovery if you already know the workflow and start the eval set on day one."

The other thing the playbooks get wrong is the assumption that the model is the hard part. It is not. The model is a paid API. The hard part is the integration into a CRM that nobody has touched since 2019, the consent flow your legal team will rewrite three times and the question of who carries the pager when the model decides to refuse a legitimate refund request at 2am. Talking to operators running this in production, the time split looks roughly 20% model, 60% integration and operations, 20% policy and rollback. The model is the easy part because Anthropic and OpenAI have done that work for you.

When operators come to us with "we want to do AI in 90 days," the first thing we do is push back on the timer. If the use case is named, the metric is baselined and the data is reachable, the timer makes sense and we run the schedule above with weekly checkpoints. If any of those three are missing, the first conversation is about getting them in writing, and we will say so before quoting anything. Our services page has the shape of the engagements we take, and the contact form is the right place to start if you want a thirty-minute call where we mostly listen and then tell you which of those three things is missing.