Best AI Operating System

Rolling out an AI operating system in 90 days

Thirteen weeks, three phases, one department at a time. The sequence that keeps an AI rollout measurable and reversible.

By The editors · Updated · 4 min read

Five colleagues from sales, finance and operations hold a short morning stand-up around a high table with a printed week plan and a paper wall calendar behind them.
A rollout that works is mostly a sequence of short, dull meetings with a number on the table.
Contents
  1. Why a fixed 90 days
  2. Before day one
  3. Weeks 1 to 4: design and rules
  4. Weeks 5 to 8: run alongside
  5. Weeks 9 to 13: measure and decide
  6. A scorecard for day 90
  7. Questions buyers ask
  8. Sources

The failure pattern of AI projects is well documented: a promising demonstration, then months in which nothing in the daily work changes. A rollout plan exists to prevent that gap. It fixes the scope, the measure and the date by which a decision is made.

Why a fixed 90 days

MIT's 2025 study found that about 95% of enterprise generative AI pilots showed little or no measurable effect on profit and loss⁠1. The study's finding that bought solutions succeed far more often than internal builds is a reason to buy rather than build, not a reason to skip the measurement. Ninety days is long enough for one department to complete several work cycles and short enough that the original question, whether the system helps, is still being asked.

The plan below applies to any of the products in our ranking. Where a step depends on what the product can do, such as approval rules or moving hosting, the differences are noted.

Before day one

Three decisions come first.

  • Name an owner. One manager accountable for the rollout, with a few hours a week set aside.
  • Choose the first department. Pick the one with repetitive work that crosses systems and a measure that already exists: response times in customer service, quote turnaround in sales, days sales outstanding in billing.
  • Record the baseline. Measure that number for the four weeks before the start. Without a baseline, every result afterwards is an opinion.

Weeks 1 to 4: design and rules

Weeks 1 to 4

  • Describe how the department works, step by step, including the exceptions staff handle every day.
  • Design the system around that description, not around the vendor's demo.
  • Write down, for each task, what the AI may do on its own, what waits for approval and who approves.
  • Connect the records the department needs, and decide which system remains the record for each type of data.

Design tools have become fast. OpenLazarus says a first design in its builder takes about ten minutes, with a live preview, and that companies of 30 or more people start with a session with its engineers⁠2. Office-suite AI needs no design at all, but its reach is limited to the mail, files and meetings it can already see⁠3.

The rules take longer than the design, and they matter more. OpenLazarus lets the company decide per teammate what runs automatically and what needs approval⁠4; its examples include a purchase over a set amount or a change to a promised date waiting for a person, with every action logged where staff can see it⁠5. Whatever the product, the rules should be written before any agent is switched on.

Weeks 5 to 8: run alongside

Weeks 5 to 8

  • Run the AI on live work with approval required for every consequential action.
  • Review the record of what the AI did at the end of each day for the first two weeks, then weekly.
  • Relax an approval rule only when the AI has handled that task correctly many times in a row.
  • Keep the old process available, but stop duplicating work in it once the new one is trusted.

This is the phase in which most rollouts are won or lost. Staff need to see what the AI did and why, and to correct it easily. Microsoft provides admin usage reports for Copilot to track adoption⁠3; a system that also records actions on the records themselves gives the department manager a clearer picture than usage counts.

Weeks 9 to 13: measure and decide

Weeks 9 to 13

  • Compare the department's measure with the baseline, using the same definition.
  • Count the hours staff spent supervising the AI, not only the hours saved.
  • List the subscriptions the department no longer uses and their renewal dates.
  • Decide in writing: expand to the next department, adjust and rerun, or stop.

Retiring tools is where savings become real. Zylo reports that organisations keep uncovering unused seats and that 61% of IT leaders cut projects because of unplanned SaaS cost increases⁠6. Cancel at the renewal date, after the data has been exported and checked.

A scorecard for day 90

QuestionA good answer
Did the measure move?Yes, against a four-week baseline, by the same definition
Is the AI's work visible?Every action recorded where the responsible person can see it
Are the rules written down?Per task: automatic, approval required, or never
What did it cost?Licences, hosting, usage and staff supervision time, all counted
Can we leave?Data exportable, and the configuration documented

The end state after 90 days is modest and specific: one department works on one system, its measure has moved, the rules for the AI are written and enforced, and the next department has been chosen on evidence. The comparison of AI operating systems, ERPs and copilots helps decide which records stay where as the rollout grows.

Questions buyers ask

How long does it take to set up an AI operating system?

The design can be quick. OpenLazarus says a first design in its builder takes about ten minutes, and office-suite AI can be switched on by licence. Moving records, setting approval rules and training a team takes weeks, which is why we plan in 90 days for the first department.

Which department should go first?

The one with the most repetitive work that crosses systems and the clearest measure, such as customer service response time, quote turnaround or overdue invoices. Avoid starting in the most regulated process or the one with the most exceptions.

When can we cancel the old tools?

Only after the new system has carried the work for a full cycle, such as a month-end close or a billing cycle, and the data has been exported and checked. Cancel at the next renewal date rather than in the middle of a contract term.

Sources

  1. MIT report: 95% of generative AI pilots at companies are failing. Fortune, 18 Aug 2025. Accessed 8 Nov 2026.
  2. Build your OS. OpenLazarus. Accessed 8 Nov 2026.
  3. What is Microsoft Copilot?. Microsoft Learn, 5 Nov 2026. Accessed 8 Nov 2026.
  4. FAQ. OpenLazarus. Accessed 8 Nov 2026.
  5. OpenLazarus for manufacturing. OpenLazarus. Accessed 8 Nov 2026.
  6. 2026 SaaS Management Index. Zylo. Accessed 8 Nov 2026.

Updated 8 Nov 2026: First published.

Further reading