The AI rollout plan for a 20 to 200-person company: phases, sequence and what to do in week one

Most AI rollout plans I'm shown are a purchase order with a training day attached.

Pick a vendor, buy the seats, run a lunch-and-learn, put an adoption dashboard on the wall. Three months later the dashboard reads "last opened: 47 days ago" and the plan is declared a people problem.

It was never a people problem. It was a sequencing problem, and this page is the sequence. An AI rollout plan for a founder-led company of 20 to 200 people, in five phases, with the tool decision after the people and process work, an exit after that, and a checklist you can start on Monday without calling anyone.

The diagnosis is elsewhere, in why AI rollouts fail. Short version: you bought a tool and kept the structure. This page is what comes after the diagnosis... the plan.

Most AI rollout plans are a purchase order with a training day attached

Look at where a typical plan spends its lines. Vendor selection, security review, provisioning, a training session, a champions channel, a dashboard. Every item is about the tool. None is about the work the tool is supposed to change.

MIT's Project NANDA saw the same shape from the outside. Its 2025 report, The GenAI Divide: State of AI in Business, found that roughly 95% of the organisations it studied saw no measurable return, and that the blocker was workflow integration, not model quality.

A plan whose every line is about the tool has priced in that outcome before anyone logs in. So the first move is to throw away the shape of the plan, not the tool.

The plan in one line: People before Process before Tool

Whether you call it an AI rollout strategy, a roadmap or an implementation plan, the sequence is the same, and it's the sequence below:

The rollout: People → Process → Tool → Handover.

Five phases sit inside that line. Two are people work, one is process work, one is tool work, and one is leaving. The proportions matter more than the calendar. The people and process phases are the majority of the effort, and if your plan has them as a two-week preamble before a launch, it has already reproduced the failure it was meant to avoid.

BCG's 10-20-70 rule says the same thing from the value side: about 10% of AI value comes from the algorithms, 20% from the technology and data around them, and 70% from people and processes.

Phase 1: map the organisation you actually have

Goal. Know how work really moves before you change anything, and put every person in front of a frontier tool on their own work.

Owner. The founder or COO. Not IT, whose mandate is systems and who will produce a tool-first plan by default. Outside help can run the mapping, but the founder has to be in the room, because the output is going to disagree with the org chart they drew.

Inputs. The formal org chart, calendars, a month of status updates and approvals, and every employee for one uninterrupted hour.

What you're mapping. Three things. Where information stalls: which decisions sat open, and who they waited on. Who routes versus who decides: for each role, how much of the week is passing information along versus making something or making a call. And where each role's attention goes.

Then the painful first hour. Every employee, their own backlog, the best tool you can put in front of them, no demo and no sandbox. The first hour is the only hour anyone resists. Past it, the team's floor for what AI can do today is current, and every later conversation is held against it.

Outputs. The "real org" diagram, which never matches the formal one, and a per-role attention snapshot.

Failure to watch for. Delegating the mapping to IT, or to the people being mapped. Both will hand you back the org chart with a tool attached.

Phase 2: sort the work before you automate any of it

Goal. The ten workflows that consume the most attention, each in exactly one of three zones, signed, then the rest. And a named owner for each of the five ORBIT functions.

Owner. The founder or COO signs the map. Each pod's Outcome Orchestrator drafts it for their own work.

Inputs. The Phase 1 outputs, plus the list of the ten workflows that consume the most attention.

The sort. This is the 3-Zone Co-Pilot (Replace / Augment / Refuse). Replace: the human contribution is coordination, synthesis of known information or routine generation, so the AI does it end to end. Augment: judgment under ambiguity, novel synthesis or stakeholder management, so the AI prepares and the human decides. Refuse: legal, ethical or relationship risk outweighs the gain, so it's named in writing and left alone.

The function map. Now name owners. The ORBIT Framework names five functions every AI-augmented team has to cover, and a rollout stalls at whichever one is unowned. Outcome Orchestrator: owns the outcome and routes work between humans and AI. Operator: runs the loops day to day and catches silent failures. Innovator/Builder: designs the workflows and grades the output. AI Champion: the peer who carries the change and works the refusals. Stakeholder Liaison/Translator: carries stakeholder signal in and direction back out. Names, not titles.

Outputs. A signed Replace / Augment / Refuse map for the ten workflows, and an ORBIT function map with a name in every column.

Failure to watch for. Putting everything in Augment. It's the only zone that requires nothing to change, which is why it's the default and why it produces a faster version of the old company.

Phase 3: redesign roles, handoffs and "done"

Goal. Change the structure so that using the tool is the job, not a thing done in addition to the job.

Owner. The founder or COO, with each Outcome Orchestrator. This is nobody's job on the current org chart: not IT's, not L&D's. It sits with whoever can change how work is measured.

The redesign. Three rewrites, per workflow. For each Replace workflow, write the definition of "done" as the loop will deliver it, name the Operator, and remove the human reviewer who was there "to be safe". For each Augment workflow, write the handoff: what the AI prepares, what the human decides, what the human may not delegate. For each Refuse workflow, the document is the deliverable.

Then the approval chain. Every approval that existed to coordinate work that is now Replace comes out. Every status update that told someone what the loop now tells them is deleted, not automated. This is where roles change shape, and it's the Hourglass Collapse: the narrow middle of the chart existed to route information, and once the loops route it, the coordination-heavy roles from Phase 1 need redrawing, not a copilot.

Outputs. A definition of "done" per workflow, a handoff per Augment workflow, a shorter approval chain, and roles that match the work that's left.

Failure to watch for. Automating the status update instead of deleting it. If the loop does the work, the report about the work has no reader.

Phase 4: tools last, training inside the new workflow

Goal. One tool decision per Replace workflow, a working loop for each, and a team trained on the changed workflow rather than on the tool.

Owner. The Innovator/Builder builds; the Operator runs; the AI Champion carries the training; the Stakeholder Liaison/Translator feeds signal back in.

The build. Only the Replace column gets a tool decision; Augment gets a handoff and Refuse gets a document. Per Replace workflow, one loop: what triggers it, what policy it follows, which tool it calls, what the quality gate checks, how it learns from what the gate rejects. Pick the tool per workflow, not per company. A single-vendor decision made before the sort is how the plan became a purchase order in the first place.

Then train, inside the workflow. Now there is something specific to teach: how the redesigned workflow runs and where the judgment calls are. Hours with a human in the room, on live work. Anything earlier is training for a workflow that's about to be obsolete, which is why AI training doesn't stick.

Outputs. Working loops on the Replace workflows, each with a named Operator, and a team using them as the starting point for real work from the day they ship.

Failure to watch for. Forty agents and four working workflows. A Builder handed an unsorted org builds whatever is easiest to build. The signed map from Phase 2 is what stops that.

Phase 5: hand over and leave

Goal. The team runs the system, improves it, and doesn't need the outside help to do either.

Owner. The Outcome Orchestrators, collectively. If outside help ran the earlier phases, this is where it works itself out of the room.

The handover. Ownership transfers by function, not by document. Each named owner runs their function for a full cycle with the outside help watching and not intervening. An Outcome Orchestrator chairs a monthly review (which workflows moved zone, which loops the Operators flagged, what the Liaisons brought back) and the team picks the next Replace candidates. The test I use: when that review has run twice without anyone external in the room, the rollout is done.

Outputs. A self-improving system with internal owners, and an exit.

Failure to watch for. The consultant who stays. If the plan has no exit written into it, the outside help has become the coordination layer you were removing. That's the structural reason the Fractional CAIO role is temporary by design: senior enough to change the chart, outside enough to say it, and gone when the team runs it.

A week-one checklist

Everything here is Phase 1 and the front of Phase 2. None of it needs a vendor, a budget or a tool decision.

1. Draw the real org on one page. For each person: who they get information from, who they hand it to, who they wait on. Ignore the reporting lines.
2. For each role, estimate how much of the week goes to routing, summarising and chasing versus creating or deciding. Mark every role where routing wins.
3. Ask each person for the last thing they waited on too long, and who they were waiting for. That's your stall map.
4. List the ten workflows that consume the most attention. One line each: what comes in, what goes out, who signs it.
5. Book one uninterrupted hour for every employee on their own backlog with the best frontier tool you can give them. Enforce it.
6. Find the shadow users. Someone is already doing real work on a personal account. Ask them to show you.
7. Run the 30-minute self-audit in the ORBIT Framework essay. Which of the five functions has nobody's name next to it?
8. Write the Refuse column first. It's the fastest to agree on, and it makes the other two safer to argue about.
9. Pick one obvious Replace candidate and write its definition of "done" in one sentence. If you can't, it isn't Replace yet.
10. Freeze tool purchases until Phase 3 is finished. Cancel nothing, buy nothing.

What to measure

The adoption dashboard is the wrong instrument: it measures whether people did the extra thing you asked for on top of an unchanged job. Measure the structure instead.

Workflows that moved zone. Named, dated, signed. The count going up is the plan working. The count stuck at zero after Phase 2 is the plan being a purchase order again.

Handoffs removed. Approval steps deleted, status updates that no longer exist, reviewers taken out of Replace loops.

Decisions that stopped stalling. Re-run the stall map from week one. Items that were waiting on a routing layer should now be waiting on nobody.

Loops with a named Operator and a measured first-time-right rate. Not adoption of the loop. Whether its output was right without a human fixing it.

Every one of those is an org-design measure. The wider argument for why a rollout is an org-design decision and not an IT one lives under AI strategy.

The binary

Two versions of your next rollout.

In the first, the plan is a purchase order with a training day attached. You pick the vendor, buy the seats, run the session, watch the dashboard. Three months later you're reading the same numbers with a different logo on the invoice.

In the second, you spend the first week drawing the company you actually have. You sort the work. You rewrite "done". The tool decision arrives late, small, per workflow, and mostly obvious by the time it's made. And the plan has an exit in it.

Both are AI rollout plans. Only one of them changes how the company works.

Not choosing is the first one.

FAQ

What is an AI rollout plan?

An AI rollout plan is the sequence a company follows to change how its work gets done with AI. It is not the schedule for installing a tool. A good one covers people first (mapping the real organisation, frontier-tool exposure), process second (sorting workflows into Replace, Augment and Refuse, redesigning handoffs and "done"), tools third, and a handover at the end. If the plan is mostly vendor selection and training, it's a purchase order.

What are the phases of an AI rollout?

Five, in this order. Map the organisation you actually have and put everyone in front of a frontier tool on their own work. Sort every workflow into Replace, Augment or Refuse and name owners for the five ORBIT functions. Redesign roles, handoffs and the definition of "done". Pick tools per Replace workflow and train inside the changed workflow. Hand over, so the team runs it and the outside help leaves.

How long does an AI rollout take?

Longer than buying a tool, shorter than a data-readiness project, and the calendar is the wrong test. The useful test is proportion: the people and process phases should be the majority of the effort, the tool phase the smallest, and the handover ends when the team has run its own monthly review without anyone external in the room. A plan with a two-week people preamble before a launch has already failed.

Who should own an AI rollout?

The founder or COO, because a rollout is a decision about who does what, not a technology decision. IT will produce a tool-first plan by default, correctly given its remit, and the coordination layer can't be asked to redesign itself. Inside the plan, each of the five ORBIT functions gets a named owner. If nobody internal can credibly redraw the chart, that's the gap a Fractional CAIO fills, temporarily.

How do you roll out AI in a small business?

The same sequence, with less ceremony. In a 20 to 200-person company the coordination layer is often informal, a few people who hold context in their heads, so the mapping in Phase 1 matters more, not less. Start with the week-one checklist: draw the real org, list the ten heaviest workflows, give everyone the painful first hour, and freeze tool purchases until the work is sorted and "done" is rewritten.

Read this next: The Co-Pilot Zone... the Phase 2 sort, run in full, workflow by workflow.

Sources

1. MIT Project NANDA: The GenAI Divide: State of AI in Business 2025 (July 2025). 300+ initiative reviews, 52 interviews, 153 survey responses; finding that ~95% of the organisations it studied saw no measurable P&L return, with the blocker identified as workflow integration rather than model quality. Widely mirrored PDF: https://cloudelligent.com/wp-content/uploads/2026/02/v0.1_State_of_AI_in_Business_2025_Report.pdf
2. BCG: the 10-20-70 rule (10% algorithms / 20% technology and data / 70% people and processes). See From Potential to Profit: Closing the AI Impact Gap: https://www.bcg.com/publications/2025/closing-the-ai-impact-gap and The Leader's Guide to Transforming with AI: https://www.bcg.com/featured-insights/the-leaders-guide-to-transforming-with-ai

Frameworks referenced are my own: the ORBIT Framework and the 3-Zone Co-Pilot (Replace / Augment / Refuse).

---

The rollout sequence, one phase per issue, in your inbox

The frameworks behind this page go out to the list, one per issue.

Get the frameworks →

Or if you already know this is your situation: book a 30-minute scoping call

---