AI implementation roadmap: a 90-day plan from first workflow to live

A practical AI implementation roadmap covers one workflow in about 90 days, in six phases: baseline, prototype on real data, build, shadow mode, limited live use, then a decision to scale. Each phase ends at a gate with a named deliverable, so a sponsor can stop, adjust or continue on evidence. This plan assumes you have chosen the first workflow. If you haven't, start with the AI readiness assessment.

#Why 90 days, and why one workflow

Programs that try to adopt AI everywhere at once tend to produce a strategy document and little else. One workflow forces real decisions about systems, data, approvals and ownership. It also produces a number your finance team will believe.

Ninety days is a planning frame, and nobody should treat it as a promise. A workflow with one system and clean inputs may go live sooner. One with a legacy ERP and scanned documents may take longer. The gates tell you which situation you are in.

#The plan at a glance

Phases overlap slightly, because integration work can start while the prototype is still being measured.

#Weeks 1 to 2: baseline and scope

Without a baseline, nothing later in the plan can be judged. Before any build, find out how many items arrive a month, how long each takes by hand, how often people fix them, and who does the work.

Concrete tasks:

1. Walk the workflow with the people who do it, not only their manager.
2. Collect 50 to 100 real items, including ugly ones.
3. Time a sample of items by hand.
4. Confirm access: API credentials, test environments, a sandbox if the system has one.
5. Name an owner on your side who can answer questions within a day.
6. Write down what will stay manual.

A scoped audit does this work and gives you a document to take to funding. Our sample audit shows what comes out: a baseline, ranked opportunities and an estimate of the automatable share.

On the money side, the AI ROI calculator lets you test assumptions before you commit.

#Weeks 3 to 5: prototype on real data

Build the smallest version that runs end to end. For invoices, that means reading a set of them, extracting fields and matching to purchase orders, with no polished interface. For support, it means classifying a batch of past tickets and drafting replies.

The prototype exists to produce a number. Label your sample with the right answers and score the prototype against them. You'll find which cases fail and why. Maybe one supplier's layout breaks extraction. Maybe a ticket category is ambiguous even to your own agents. You want those findings in week five, not week twelve.

Gate rule: if accuracy on the test set is far from your target and the failures look structural, pause. Fix inputs, narrow the scope or stop. Stopping here costs weeks. Stopping after the build costs months.

#Weeks 5 to 8: build and integrate

This is where the engineering gets heavy. The system gains the parts that make it safe to run:

• Integrations with your ERP, CRM, help desk or inbox, read access first and write access once tested.
• Validation rules, such as totals that must add up or fields that must match a master record.
• A review queue where people approve, edit or reject, with the reason for each flag.
• Permissions that limit what the system can do on its own.
• Logging of input, decision and action for every item.
• Failure alerts, so nobody learns about a break from a customer.

Bring your security and IT teams in during week five, not week eight. Access requests and data reviews eat calendar time. If your data can't leave your environment, settle hosting now and write the data-handling rules into the scope document.

#Weeks 8 to 10: shadow mode

In shadow mode the system processes live items while people do the work as usual. You compare what the system would have done with what your team did.

A test set can't answer some questions, and this phase can. How does the system cope with a Monday morning spike? What happens when a supplier sends a new format? Where do your own staff disagree with each other, and who is right? Say a team processes 2,000 invoices a month (illustrative). You might find that most disagreements trace to approval rules that live in one person's head. That's a useful finding, and a human has to resolve it.

Gate rule: agree the pass bar before the run starts. A typical bar sets accuracy on the fields that matter, the share of items sent to review, and zero unauthorized actions.

#Weeks 10 to 12: limited live

Start with a slice: one supplier group, one ticket category, one region, or 10 to 20 percent of volume. The system acts and humans approve where you agreed. Keep the manual path available and write down the switch-off procedure.

Track daily at first:

• Items processed and share sent to review
• Corrections made by reviewers, and why
• Failures and how quickly they were fixed
• Time per item, compared with baseline
• Anything the system did that nobody expected

Raise the slice as the numbers hold. Don't raise it because the calendar says week twelve.

Key point: Every phase should end with something a sponsor can look at: a baseline, a measured accuracy figure, a shadow-run report, a results table. If a phase ends with only a status update, you have no evidence for the next decision.

#Weeks 12 to 13: review and scale

The final gate is a decision memo covering four things: results against the success metrics you set in week two, actual run cost against the estimate, what broke and what you changed, and what comes next.

Outcomes vary. Sometimes you expand volume. Sometimes you fix a weakness first. Sometimes you stop, which is a legitimate result for a cheap experiment. Build the next-workflow shortlist from what you learned.

Before the project closes, confirm ownership. Who monitors the system? Who handles supplier or model changes? Who approves rule updates? A live automation with no owner decays. AI automation cost explains why monitoring is a recurring line.

#Sponsor checklist

Use this list to see whether the plan is on track.

• A named owner on your side with time to spare
• Baseline numbers that finance accepts
• A labeled test set you own
• A written pass bar for shadow mode
• Security and IT involved before the build starts
• Approval points defined for any action that moves money or messages customers
• A switch-off procedure
• A support and monitoring plan before go-live

To see how we organize a delivery against these gates, read how we work. To plan your first workflow with us, book a strategy call.

#Frequently asked questions

#What is an AI implementation roadmap?

It is a dated plan that takes one workflow from baseline to live use, with gates and deliverables at each stage. A strategy deck produces opinions. This produces evidence at every phase: measured accuracy, a shadow-run report, results against agreed metrics.

#Can AI be implemented in 90 days?

One well-scoped workflow often can reach limited live use in about that time, if system access and sample data arrive promptly. Broader programs take longer. Treat 90 days as a frame for the first workflow, with the scale decision at the end.

#What should happen before implementation starts?

You need a chosen workflow, a baseline, a named owner and confirmed system access. The readiness assessment tool can show you the gaps, and a short audit can fill them.

#What is shadow mode?

The system processes real items while people keep doing the work. You compare its output against theirs without acting on it. It exposes real-world problems a test set misses, and it gives you evidence for the go-live decision.

#What are the common reasons AI implementations stall?

No named owner, no baseline, a late security review, unclear approval rules and expanding scope. Most show up by week four, which is why the early gates exist.