Two companies can run the same AI enablement program, with the same trainer and the same agenda, and get opposite results, because one of them did the steps in the wrong order. Sequencing AI enablement is the part nobody writes down. The published guidance is a row of phase diagrams, assess then govern then pilot then scale, which tells you what the stages are called and nothing about what has to be true before each one starts. The rules below are the order itself, written from the delivery side, for a company of 50 to 500 people that already pays for ChatGPT, Claude, Gemini, or Microsoft Copilot seats and has started to wonder what they bought. Why that gap exists in the first place is the argument in the case for building capability rather than buying seats.
Why Sequencing AI Enablement Beats Curriculum
BCG splits AI investment ten, twenty, seventy: 10% to the algorithms, 20% to technology and data, 70% to people, process, and new ways of working. Most buyers accept that split and then spend their 70% in an order that wastes it. Stanford’s Digital Economy Lab, after 51 enterprise deployments, put it more bluntly: “The difference was never the AI model. It was always the organization.” Same tools, same use cases, results measured in weeks at some companies and years at others.
What goes into each step is settled ground, and our guide to turning a license rollout into measurable team capability covers it. The order is not. There are two ways to get it wrong, and the published advice only warns about one.
Tool-first. Licenses go out, an all-hands announces them, training is scheduled for whenever the vendor has a slot. Seats look active for a month and then flatten, because no workflow changed and the dashboard was never measuring capability.
Policy-first. A committee drafts an acceptable-use standard before anyone has watched a real person paste a real client document into a real tool. It regulates imagined behavior, ends up either too strict to roll out or too vague to answer anything, and gets rewritten the week after the first session.
Both come from treating the program as a stack of layers to complete rather than a sequence of preconditions. Seven rules, in order, each naming the failure it prevents.
Rule 1. Pick One Workflow Before You Pick Any People
Start with a named piece of recurring work that has an owner and a number attached. Not a department, not a headcount. One workflow: the monthly close, first-draft proposals, inbound lead qualification, candidate screening, the weekly client report.
This inverts the usual first move, which is to ask who should be trained. A room assembled by job title shares no task, so headcount-first scoping can only produce a generic agenda. Workflow-first scoping tells you who is in the room as a consequence, and tells the facilitator what the exercises are built from.
The owner matters as much as the workflow. Somebody has to be able to say the current version is slow, hand over real examples of it, and accept a changed version afterward. A process with four part-owners will absorb training and revert, because reverting is nobody’s fault.
The failure this prevents: a workshop that teaches prompting in the abstract. Fluency is a transfer problem, and few people carry a skill from a sample dataset into their own inbox. The hard part was always the messy input they already have.
Rule 2. Take the Baseline in Week One, Not Week Twelve
Measurement gets sequenced first and reported last. It is the most skipped step in the order and the only one that cannot be recovered.
A baseline here is small. Two questions per attendee about the workflow from Rule 1, asked before anyone opens a tool: how long does it take you, and how many did you do last month. Add a quality marker if one exists, like rework rate or escalation count. Fifteen minutes per person.
Skip it and week twelve produces an opinion. You will have good stories, visible artifacts, enthusiastic attendees, and no answer to the only question your CFO will ask. Nobody reconstructs a cycle time after the fact, because the people who knew it have changed how they work and will now estimate the old number generously.
McKinsey’s State of AI found only 21% of organizations using generative AI had “fundamentally redesigned” any workflow, and that redesign moved earnings more than any other practice. You cannot show a redesign without a before. What each milestone can honestly claim is laid out in our realistic program timeline.
The failure this prevents: an unprovable program. Not a failed one. An unprovable one, which is worse, because it cannot be defended at renewal.
Rule 3. Write One Page of Policy Before Session One
One page, before the first session, and no earlier than the week you pick the workflow.
The page answers four questions. What may go into a tool and what may not. Which tools are approved. What a human checks before anything leaves the building. Who to ask when the page does not cover it. That last line does most of the work, because the page will not cover plenty.
Thin is the point. A long standard written before anyone has used the tools regulates guesses. Write it after Rule 1, so it names real document types, and expect to revise it once you have seen what people try to paste.
For any company with an EU footprint there is a floor underneath this. Article 4 of the EU AI Act, applicable since 2 February 2025, puts a duty on providers and deployers to ensure a sufficient level of AI literacy among staff operating these systems. In that market, training is a compliance question as well as a return-on-investment one.
The failure this prevents: both halves of the governance trap. No page, and your most careful people stay frozen while your least careful ones improvise. Forty pages written in advance, and the rollout waits on a committee.
Rule 4. Open Access Days Before Training, Not Months
Licenses and training have to arrive close together, in that order, with a gap measured in days.
Most companies get this backwards by accident. Seats were bought in March, training lands in October. By then the curious people have formed habits nobody taught them, the cautious ones have decided the tool is not for them, and the session has to undo both before it can teach anything. The reverse gap is as bad: a workshop on a tool people cannot open afterward converts to nothing by Friday.
A week of access before session one is about right. Long enough that nobody is fighting a login during hour one, short enough that the room is still curious.
BCG’s AI at Work 2025 survey, covering more than 10,600 workers across 11 countries, found 72% using AI regularly overall but only 51% of frontline workers, with that frontline number flat year over year. Access was never the bottleneck.
The failure this prevents: paying twice. Once for seats that went unused for seven months, once for a session that spends hour one on password resets and hour two unpicking bad habits.
Rule 5. Go Deep With One Group Before Going Wide
Train 15 to 20 people until the workflow from Rule 1 has changed. Then widen. Not the reverse.
Company-wide first is the more tempting order, because it looks fair and books one date. It also guarantees a lecture. Above roughly 20 people, hands-on work stops being possible and the event becomes a webinar with a chat window. Spread across four functions, the curriculum collapses to the intersection of their jobs, which is close to nothing.
Depth first also buys the content for everything after it. The first cohort generates the examples, the objections, the two drills that worked, and the one that fell flat. Wave two is built from that, not a vendor template.
BCG’s survey found 36% of employees feel adequately trained, and that regular use rises among people with five or more hours of training, highest where it is in person with coaching. Five hours with twenty people beats one hour with two hundred, and costs less.
The failure this prevents: a company where everybody has heard of the tools and nothing runs differently, plus a budget conversation with no finished example in it.
Rule 6. Put the Skeptic in Wave One
Standard advice says to start with volunteers and a leader who wants AI. Half right: a room of volunteers produces a flattering pilot and a useless signal.
Build wave one from the people who do the work, which means including the senior person who thinks this is a fad and the careful one who worries about client data. Their objections are the ones that will stop wave three, and you want them raised in week two, in a room, with a practitioner answering them, rather than in month five in a leadership meeting where nobody can test the claim.
In the sessions we run, the question that decides a room is almost always an unscripted one from a doubtful senior person. Answer it with a demonstration on their own document and the rest of the room shifts. Deflect it and the hour is spent.
One caveat on who to leave out. A team mid-reorganization, or one whose system of record is in motion, is a bad first cohort for reasons unrelated to attitude. Pick the workflow for stability, then take the skeptics inside it.
The failure this prevents: a pilot that reports success and a rollout that stalls, because the real objection was never raised.
Rule 7. Book the Follow-Up Before You Book the Workshop
Put the 30-day checkpoint on calendars before the training date is confirmed. If it cannot be scheduled, the program is not ready.
This looks like a scheduling detail and is the whole retention question. Capability built on a Tuesday decays within weeks unless something holds it in place. The follow-up is the cheapest hour in the program and the first one cut when calendars tighten, which is why it gets committed before there is anything to cut it for.
Three things sit in that slot. A short working session on the same workflow, so people bring what they tried. A recording and document pack for whoever missed the live session and whoever is hired next quarter. And the second half of Rule 2’s baseline: same two questions, same people, same task.
What that cadence looks like once it is years old rather than months is covered in our account of the second year of an enablement retainer.
The failure this prevents: a strong workshop with no trace of it by December, and a renewal conversation holding nothing but satisfaction scores.
The AI Enablement Sequence on One Page
| Order | Step | Precondition it depends on | Recoverable later? |
|---|---|---|---|
| 1 | Name the workflow and its owner | Nothing | Yes |
| 2 | Capture the baseline | Step 1 | No |
| 3 | Write the one-page policy | Step 1 | Yes |
| 4 | Open tool access | Steps 1 and 3 | Yes |
| 5 | Run the deep cohort of 15 to 20 | Steps 2, 3, 4 | Yes |
| 6 | Hold the 30-day checkpoint | Booked before step 5 | Partly |
| 7 | Redesign the workflow, then widen | Steps 5 and 6 | Yes |
The last column matters most. Steps 1, 3, 4, 5, and 7 can all be fixed late at some cost. Step 2 cannot, which is why a vendor who does not ask for your current cycle times in week one is not running a program that intends to be measured. Our twelve rules for choosing an enablement partner covers what else that first conversation should surface.
Step 7 earns its own line. Train people and leave the process alone and the savings stay individual: somebody finishes a draft in 35 minutes instead of 90, and the spare hour disperses. Redesign is where it reaches a P&L, through fewer handoffs, a review stage removed, a deadline pulled forward. It is also the step most programs never reach, which is why that 21% is as low as it is.
Recovering a Program That Is Already Out of Order
Most readers are not starting clean. Licenses went out last spring, something was run in June, and the question is what to do from here. Triage by what is still available. No baseline means picking a second workflow and baselining that one properly this month, rather than inventing a first number from memory. No policy means writing the one page this week. A company-wide session that already happened counts as awareness, not training, which makes the deep cohort cheaper now than it would have been.
One call is worth making on purpose: when to stop. A workflow that has been through a cohort, a checkpoint, and a redesign attempt without moving its numbers was a bad candidate. Change the workflow, not the curriculum.
Frequently asked questions
What does sequencing AI enablement actually mean?
It means the order of operations in a capability program: which steps have to finish before the next one starts. Workshop content is roughly the same at most providers. The order around it, when the baseline is captured, when the policy is written, when access opens, how wide the first cohort goes, separates a program that shows a result from one that shows attendance.
What is the first step, honestly?
Name one recurring workflow and the person who owns it. Everything else, including who attends, follows from that. Picking people first is the most common opening mistake and it costs you the whole design.
Do we need an AI policy before any training?
You need one page before the first session, not a full standard before the program. Write it after you have picked the workflow so it refers to real document types, and expect to revise it once you have seen what people try. A policy drafted before anyone has used the tools regulates guesses.
Can we skip the baseline and measure afterward?
No. It is the only step in the sequence that cannot be recovered. Once people have changed how they work, nobody can state the old cycle time accurately, and estimates drift in the flattering direction. Two questions per person in week one is the entire cost.
How long should the gap be between buying licenses and training?
About a week. A long gap lets habits form that training then has to undo, and no gap at all means hour one goes on logins. Companies that bought seats months ago should assume part of session one is unpicking workarounds.
Who should own the sequence internally?
One person with roughly two hours a week and the authority to get people into a room. At 50 to 500 people that is usually a COO, chief of staff, or head of operations rather than a dedicated AI lead, and it is almost always a second job. The hours are the constraint, not the knowledge.
What does this cost to run properly?
Workshop programs sit in a $2,000 to $15,000 market range depending on how many roles need their own material, and a custom automation build after it runs $25,000 to $150,000. Do not scope the build before the cohort finishes: a team eight weeks into these tools knows which task is worth automating, and the same team in week one will guess.
How do we know the sequence is working before the numbers land?
Watch for one thing in weeks two to four: people bringing their own messy work to the checkpoint without being asked. That precedes every measurable result we have seen, and its absence is an early warning that no workflow is going to change.
Bring three things to a first conversation and the order falls out of them: the one workflow you would most like to be faster, the person who owns it, and whether you can hand over real examples. A free AI-readiness call is enough to turn those into a sequence with dates on it.
Arthur Wandzel
