An AI enablement rollout across several locations is not a larger version of one at a single site. The number of sessions you pay for is set by how many people fit in a hands-on cohort, not by how many offices you have — and you only pay per office once you decide every cohort must be single-site.
One disambiguation, because the phrase is overloaded: this is not about training a model on compute split between data centers. It is the operational problem. You authorized AI licenses for a company with offices in three or six places, somebody has to turn them into capability, and the calendar has stopped cooperating.
Operating from more than one address is ordinary — the Census Bureau’s Statistics of U.S. Businesses covers “more than 2.0 million multi-unit establishments” — and nothing about the enablement market reflects that. Search for a provider who delivers across locations and you get booking pages, platform roundups, and playbooks borrowed from sales enablement. None does the arithmetic, so start there.
The session count is set by cohort size, not office count
Hands-on LLM fluency training has a hard ceiling on room size. What makes it transfer is a facilitator sitting with people while they rewrite a real deliverable of their own — a live proposal, a board memo, a client email — and that stops working somewhere around 20 to 25 participants per facilitator. Above the cap you are running a demo, and a demo changes nothing on Monday. So the first number is not your office count; it is headcount in scope divided by the cap.
Take a 240-person company putting 118 people in scope across five locations: 54 at headquarters, then 28, 19, 11, and 6. At a cap of 24, the two ways to slice that are far apart in cost.
| Cohort rule | How the math runs | Sessions |
|---|---|---|
| Every cohort is single-site | 54 → 3, 28 → 2, 19 → 1, 11 → 1, 6 → 1 | 8 |
| Cohorts may mix sites | 118 ÷ 24, rounded up | 5 |
Three sessions separate those rows. A single facilitated session for one cohort generally lands in the lower half of the $2,000 to $15,000 range a workshop engagement occupies, so insisting on single-site rooms costs roughly $10,000 to $18,000 more for the same 118 people and the same curriculum. The waste concentrates at the eleven- and six-person sites: each buys a nearly empty room at close to the price of a full one.
Mixed cohorts are usually the better product as well as the cheaper one: someone from the smallest site who practices alongside three peers from headquarters leaves with a cross-site contact list, the cheapest follow-through mechanism available. The constraint on mixing is not politics. It is whether those people share an hour.
What multiplies when you add a location
Four cost drivers behave differently, and vendors quote them as one number. That is what makes these quotes hard to compare.
| Cost driver | Scales with | Does a new office change it? |
|---|---|---|
| Platform seats | Headcount | No — seats are location-blind |
| Facilitator hours | Cohorts, and cohorts come from headcount ÷ cap | Only if you refuse to mix sites |
| Travel and facilitator days lost in transit | Sites you insist on visiting in person | Yes, directly |
| Manager follow-through | Managers, which track sites | Yes, and it is paid in payroll, not fees |
Seat-based platforms are indifferent to where you operate. Coursera’s Teams page is sized for “2 to 499 employees” and says nothing about where those employees sit. That indifference is the attraction and also the problem: a location-blind product cannot notice that one of your sites never logged in.
The driver that quietly dominates is the last row. Nobody invoices you for it, so it never reaches the business case, and it decides whether anything survives week three. One named person per site who asks in the next team meeting what people used the tool for is worth more than a second session — and a rollout with no local owner reverts at every location alike, as the case against treating licenses as capability argues in full.
Subtract the offset from eight
The scheduling question has a closed-form answer. For two sites each working a standard eight-hour day, the hours they share equal eight minus the offset between their clocks.
| Pair | Offset | Shared business hours |
|---|---|---|
| Eastern and Central | 1 hour | 7 |
| Eastern and Pacific | 3 hours | 5 |
| Eastern and Central European | 6 hours | 2 |
| Pacific and Central European | 9 hours | 0 |
Within the continental US, mixed cohorts are always schedulable: five shared hours between the coasts covers a three-hour working session nobody takes at dawn, and you can rotate which coast hosts the early slot.
Once an offset reaches eight hours the shared window is zero and the question stops being a preference. A company with sites on the US west coast and in continental Europe has no hour that is business time in both, so no three-way window with its east coast staff exists either — the European workday ends before the Pacific one starts. You are then not buying one program for five offices but one per region. Say so in the brief and the quotes come back comparable.
Two housekeeping notes. Offsets shift twice a year and not in step across countries, so lock session times to one site’s clock rather than converting. And never let the same site absorb the awkward hour twice running — the second time is when attendance collapses.
Three shapes you can buy
Everything on offer is a variation on three structures.
| Shape | What you pay for | Where it breaks |
|---|---|---|
| Replicate — run the same session for each site | One session per site | Partial cohorts at small sites; cost rises with addresses rather than with people |
| Converge — one synchronous session for everyone | Fewest sessions on paper | Breaks the cohort cap, so it degrades into a demo; one region always takes the bad hour |
| Cascade — train a cross-site group deeply, they run local practice | A small number of deep sessions plus internal hours | Needs a named owner per site and real time in their week; fails silently if those hours are not budgeted |
Most mid-market companies should buy a blend closer to cascade than it looks: mixed-site cohorts sized to the cap, a deeper second session for whoever will run local practice, and a written expectation of what those people do for the next month. Replicate is defensible when a site has a genuinely different workload. Converge is defensible only for a 30-minute policy briefing, which is not training.
A provider’s published terms tell you which shape their pricing assumes. NobleProg’s public delivery pages offer training “at a location of your choice,” distinguish private courses “for you and your colleagues only” from open ones, and list small group sizes — typically four for remote delivery, two for classroom — with a guaranteed course that “will be organized even for one participant.” Read that as procurement intelligence: a vendor who will run a private remote session for a handful of people sells by the session — the unit you need for a small satellite site.
Making the quotes comparable
Ask for every quote in four lines: facilitation per cohort session, materials and platform access per person, travel as a pass-through at cost, and anything the vendor expects your staff to do. A blended per-head number answers the wrong question, and you cannot compare two of them.
Travel, separated and capped. A facilitator flying to a site loses most of a day each way, and that day is in your number somewhere. Quote it as a pass-through with a ceiling, then decide site by site whether in-person earns it. For a nineteen-person office it usually does not.
Recording and reuse rights. If a cohort is recorded, put in writing who may watch it, for how long, and whether a later hire counts. This is where the “one office gets the live session, another gets the video” failure starts: a recording is a reference, never a substitute for a cohort, and the two sites diverge within a month of your treating it as one.
Per-session cancellation, not per-engagement. Multi-site schedules slip, and one site will postpone. Confirm whether the cancellation ladder applies per session or re-prices the whole engagement, and whether moving a date costs a flat administration fee rather than a percentage. How that meets the payment schedule is covered in our piece on milestone versus upfront payment for these engagements.
Clauses a single-site agreement does not need
Four provisions earn their place only when more than one location is involved.
- A per-site deliverable, not a per-site session. Name the artifact each site produces: five documented workflows with before-and-after times, signed off by the local owner. A session date is an event; an artifact is evidence.
- A named local owner per site, in the agreement. A person, not a job title, with the hours they are expected to spend in the four weeks after their cohort.
- Attendance and completion records per person, per site, delivered to you. You will want them for the next budget conversation, and if any location employs staff in the EU they are your only evidence of meeting the AI literacy duty in Article 4 of the EU AI Act — a duty that sits with you as the deployer rather than with the provider you hired, applied since February 2025, and written as an outcome rather than a required curriculum. That text has been under revision through the Commission’s digital omnibus work, so check the current wording on EUR-Lex before quoting it in a policy.
- A substitution and reschedule right that does not re-price. Swapping an attendee between sites, or moving one site’s session by three weeks, should not reopen the fee.
Sequence the locations, do not launch them together
A simultaneous five-site launch gives you five copies of the same mistake and no control group. Run one site first, far enough ahead to change the curriculum before the second cohort, and treat that run as a test of the vendor rather than of your people — the logic in the sequencing decision between one department and the whole company applies to locations the same way.
Pick that site for diagnostic value, not convenience. Headquarters is the tempting choice and usually the worst: the most self-taught users, the most direct access to whoever authorized the licenses, the least representative conditions. A mid-sized site with an engaged manager and no AI enthusiasts tells you far more about what the other four will do. Paying for the run on its own terms, rather than as the opening session of a committed program, is the point of running a paid pilot before a company-wide commitment.
Measure the variance between sites
A multi-site program hands you a natural experiment a single-site one cannot. Same curriculum, same facilitator, different local management, so whatever differs afterwards is mostly down to local conditions. Compare the spread rather than the mean. Four weeks after each cohort, count the workflows still being done the new way, per site. A program averaging 60% retention across five locations can be two sites at 90% and three at 40%, and the remedy is not another vendor session — it is finding out what the two sites did and requiring it everywhere else. Whether to build that measurement capability in-house or buy it is the asset-by-asset question in the buy-versus-build playbook for enablement capability.
Questions only a multi-site buyer asks
Put these to every provider on the list. They sort the field faster than a capability deck.
- What is your maximum cohort size for hands-on work, and what changes above it?
- Will you run a cohort that mixes people from different sites, and have you?
- How is travel quoted, and what is the ceiling?
- Who facilitates each session — the person in this meeting, or a subcontractor?
- If one site postpones by a month, what does that cost and does it re-price the rest?
- What attendance and completion records do we receive, per person and per site?
Frequently asked questions
How much does AI training cost for a company with multiple offices?
Price it per cohort session, not per office. Divide in-scope headcount by a cohort cap of roughly 24 and that is the number of sessions you are buying; one session for one cohort generally sits in the lower half of the $2,000 to $15,000 range these engagements occupy. Offices add cost two ways only: travel, if you want in-person delivery at each, and partial cohorts, if no room may mix sites. A 118-person rollout across five locations needs five mixed sessions or eight single-site ones.
Should every office get the same curriculum?
Yes, in a US mid-market company. What differs by site is the follow-through, not the content — who asks about it in the next team meeting, and which local workflows get rebuilt first. Localization is billable, which is why it is sold, and it is rarely the constraint at this size. Spend that money on a named owner per site.
Is remote delivery as effective as in-person for a satellite office?
For hands-on LLM fluency work, a live remote session with a small cohort is close to equivalent and far cheaper than flying a facilitator to a nineteen-person site. What does not work is giving one location the live session and another the recording of it.
What do we do when no shared business hour exists?
Stop treating it as a scheduling problem. Subtract the clock offset from eight: at eight hours or more the shared window is zero and no amount of goodwill produces one. Buy one program per region rather than per site, same curriculum and same artifact, and let each region run on its own calendar.
Who should own a rollout across several locations?
One central owner for curriculum, vendor relationship, and measurement; one named person per site for follow-through. Split it the other way — local curriculum, central follow-up — and you get five incompatible versions and no comparable data. That central owner is usually doing this as a second job, which argues for fewer moving parts.
Can we buy a platform and skip the sessions?
You can, and seat-based platforms are indifferent to how many addresses you have, which is their real advantage. What they cannot do is notice that one site never logged in, or that usage at another is all first-draft email. They work as a reference layer under facilitated practice and poorly as a replacement for it.
How do we keep the smaller sites from being ignored?
Put them in mixed cohorts with headquarters instead of running thin single-site sessions, and give each a named owner with budgeted hours. The failure mode is predictable: small sites get scheduled last, their session moves twice, and by the time it happens the program has lost its sponsor.
How long before we can tell whether it worked at each site?
Four weeks after each cohort, per site. Count the workflows still being done the new way and compare locations rather than averaging them. One blended number across five sites hides the information you need.
Key takeaways
- Sessions come from in-scope headcount divided by the cohort cap, not from office count. Offices multiply cost only when you refuse to mix sites in a room.
- Shared business hours equal eight minus the clock offset. At eight or more, plan by region rather than by site.
- Split every quote four ways: facilitation per cohort, per-person access, travel as a capped pass-through, and your own staff’s hours.
- Buy the curriculum centrally, pay for follow-through locally. A named owner per site beats a second vendor session.
- Compare sites instead of averaging them. The spread between best and worst is the most actionable number the program produces.
Working out how many cohorts your headcount implies, and which locations can share a room, takes half an hour with the org chart in front of you. For a second set of eyes on that arithmetic before you request quotes, book a free AI-readiness call.
Arthur Wandzel
