# 90-day AI implementation roadmap

> A realistic first quarter for an AI project runs in five phases: pick one process and measure it, connect the systems read-only, let the agent draft while people review, grant narrow write access for the cases it gets right, then widen the scope and hand it to the team that owns the process. Each phase ends with a decision that gates the next, so the project either earns the right to expand or stops early and cheaply.

Source: https://risendigital.io/guides/90-day-ai-implementation-roadmap  
Publisher: Risen Digital  
Updated: 2026-09-21

## At a glance

- **Shape of the quarter:** Five phases across roughly thirteen weeks, each ending with a decision to continue, adjust or stop.
- **First deliverable:** A baseline: how the process performs today, measured the same way you will measure the agent.
- **Longest phase:** Draft mode, where people review the agent's output and the team fixes tools and instructions rather than prompts.
- **Gate to write access:** Case types the agent handles correctly almost every time, with a limit on value, count or scope and a full audit trail.
- **Definition of done:** The owning team runs the agent, reads its metrics and can pause it without calling the builders.

Most first AI projects fail on sequencing rather than technology: they start with write access, skip the baseline, and never earn the trust to expand. This playbook lays out a quarter that runs the other way round, from a measured starting point to a hand-over, with the decision and the traps at each stage.

![A project team in a planning session in a meeting room](https://risendigital.io/images/experience/meeting-seated-1600.jpg)

## Why plan a first AI project around ninety days?

Ninety days is long enough to prove something on real work and short enough that a wrong bet is cheap. It also forces the sequencing that separates projects that get adopted from demos that get shelved.

A quarter matches how most companies already plan and budget, which matters more than it sounds. It means the project has a review date on the calendar, a sponsor who expects a result, and a natural point at which to decide whether to expand, adjust or stop. Open-ended AI initiatives drift because none of those exist.

The other reason is trust. An agent that touches your systems has to earn permissions the way a new employee does: first watching, then drafting, then acting within limits. Compressing that into a fortnight skips the evidence; stretching it over a year loses the sponsor. Thirteen weeks is enough time for the agent to meet the awkward cases, for people to get used to reviewing its work, and for the figures to show whether the drafts are getting better.

> **This is a shape, not a promise** Your phases will stretch or shrink with how documented the process is and how accessible the systems are. Treat the week counts as a starting plan and revisit them at every gate.

## What happens in each phase of a 90-day AI roadmap?

Pick and baseline, connect read-only, draft with review, grant narrow write access, then widen and hand over. Each phase ends with a decision that gates the next one.

_Interactive on the web page: Walk the quarter week by week. The timeline shows five phases across about thirteen weeks. Weeks one and two pick a single bounded process and measure how it performs today. Weeks three and four connect the systems read-only and confirm the agent can find what it needs. Weeks five to eight run in draft mode, with every output reviewed by a person and the tools and instructions fixed as the misses show up. Weeks nine to eleven grant narrow write access for the case types the agent gets right almost every time. Weeks twelve and thirteen widen the scope on evidence and hand the running agent to the team that owns the process._

## What do you need to decide in each phase?

Each phase has one decision that gates the next. Making it explicit, with a named person and the evidence they will use, is what keeps the project honest.

| Phase | The decision | Who makes it | What they need to see |
| --- | --- | --- | --- |
| Pick and baseline | Is this the right process, and is the baseline believable? | Sponsor and process owner | A definition of done people agree with and figures nobody argues about |
| Connect read-only | Can the agent find what it needs, and is the access acceptable? | Process owner with IT or security | Every missing data item has an owner; logging is in place |
| Draft mode | Which case types has it earned the right to complete alone? | Process owner with the reviewers | Drafts accepted unchanged for that type over several weeks, and misses that are understood |
| Narrow write access | Are the limits right, and is the audit trail usable? | Process owner | Nothing in the log that surprised anyone, and reversals that took minutes |
| Widen and hand over | Expand, hold or stop, and who runs it from here? | Sponsor and the owning team | The team can pause it, read its metrics and explain it to a new colleague |

Notice that the sponsor appears at the start and the end, and the process owner appears everywhere. That is deliberate. A first AI project belongs to the people who run the process, not to the people who built the agent, and the hand-over at the end is only possible if they have been deciding all along.

## What goes wrong in a first AI project, and how do you avoid it?

Most failures come from skipping a phase: no baseline, write access on day one, or a review queue nobody staffs. The fix is almost always to go back one phase rather than to change the model.

- **Starting with write access** The agent acts before anyone has seen how it behaves on the awkward cases. Go back to draft mode; the weeks feel slow but they are where trust is built.
- **No baseline** Without a measured starting point there is no way to show the agent is better, so the project ends in opinions. Measure the process before the agent touches it, even if the measurement is rough.
- **Review that becomes rubber-stamping** If reviewers approve everything, the acceptance figures mean nothing. Sample reviewed cases, rotate reviewers, and make rejecting a draft as easy as approving one.
- **Fixing the prompt instead of the tools** When a draft misses because the agent could not find the customer's history, rewording the instructions will not help. Record the reason for each miss and fix the tool, the data or the access.
- **Scope that widens quietly** Someone adds a second inbox or raises a limit without a decision. Every widening should be a dated entry in the log with the name of whoever approved it.
- **Builders who never leave** If the owning team cannot pause the agent or read its metrics, the project has produced a dependency, not an outcome. Hand-over starts in phase one by involving them in every gate.

> **The stop decision is a good outcome** A project that stops at week four because the data is not there has saved a quarter of effort and produced a list of what to fix. Plan for it as a legitimate result, not a failure.

## What should be true at the end of the quarter?

The agent runs one bounded process within written limits, the owning team operates it without the builders, and the next decision is on the calendar.

- One process runs with the agent completing the agreed case types alone and routing the rest to people.
- The baseline and the current figures sit side by side, measured the same way.
- Limits, forbidden actions and the pause switch are written down and have been tested.
- The owning team has run the review queue and read the metrics for at least a fortnight without help.
- The next process is chosen, or the reasons for stopping are written down.

From here the pattern repeats at lower cost. The integrations exist, the review habit exists, and the team knows what a gate looks like. The second process usually moves faster not because the technology improved but because the organisation did.

We run first projects this way because the alternative, a big launch followed by a quiet retreat, is what most people have already seen. Starting read-only and widening on evidence is slower for the first month and faster for every month after.

## Frequently asked questions

**Q: Can a first AI project really be done in ninety days?**  
A: A bounded process with accessible systems can be running in draft mode within a few weeks and completing agreed case types alone by the end of a quarter. What stretches the plan is undocumented rules and systems with no way in. If either is true, the first phase will tell you, and the honest response is to fix that before the agent rather than push the timeline.

**Q: What if we do not have a baseline?**  
A: Create a rough one in the first fortnight. Count the cases for a week, time a sample of them, and note how many needed rework. It does not have to be precise, only measured the same way you will measure the agent, so that the comparison at the end of the quarter is like for like.

**Q: Who should own a first AI project?**  
A: The person who owns the process the agent works in, with a sponsor above them who controls the budget and can settle disputes about scope. Builders, whether internal or a partner, support those two rather than lead them. A project owned by the technology team tends to stall at hand-over.

**Q: How much of the quarter is spent on the model?**  
A: Very little. Most of the effort goes into access, tools, review and limits. Choosing a model is a decision you can revisit in an afternoon; earning write access to your order system is not.

**Q: What happens after the ninety days?**  
A: Either the scope widens, one change at a time, or a second process starts using the same integrations and review habits, or the project stops with a written list of what would need to change. All three are sound outcomes if the decision was made on evidence from the log.

## Related pages

- [The RISE method](https://risendigital.io/guides/the-rise-method)
- [AI readiness checklist](https://risendigital.io/guides/ai-readiness-checklist)
- [How to measure AI automation ROI](https://risendigital.io/guides/how-to-measure-ai-automation-roi)
- [What is an AI agent?](https://risendigital.io/guides/what-is-an-ai-agent)