Consolidating a Company's AI Tools onto One Platform

The problem

SunCo was paying for scattered ChatGPT seats with no policy, no training requirement, and no record of who used what. Nobody owned it. Some people relied on it daily; some hadn't touched it in months; a few used it for things it shouldn't have been used for, like customer data.

The business needed one approved tool, a way to know who was actually on it, rules for what data stays out, and a defensible exception process for the one thing the new platform couldn't do (photorealistic image editing).

Cutting people off from a tool they use is where these projects fail. The technical part of a migration is small. The people part is the whole job.

What I built

The plan (May to June). Researched the migration path first, including the finding that ChatGPT Business plans can't export history, which set expectations early. Wrote the company rollout plan: phased waves, standard seats by default, a training gate, weekly office hours, and success metrics.

The wave rollout (June to August).

• Wave 1 announced June 25, Wave 2 July 15.
• Every new seat came with a one-page usage policy (what the tool is and isn't, what data stays out, accountability, required training) and a required certification course.
• A manager sign-off round before the cutoff: five division managers each made keep or remove calls on their own people.
• A usage audit of the old platform to find who was still active, then individual follow-up with each of them.
• Documented exceptions for the five accounts that stayed, each with a stated reason and an owner.

The cutoff (August 24). 27 ChatGPT seats deleted in two passes. Access ended at 4:00 PM on the announced day, after written advance notice.

The support system around it.

• A single versioned context file holding the people directory, timeline, decisions log and open items, so any planning session starts fully briefed.
• A certification tracker with an intake checklist, because completions arrived in three different formats and got missed.
• An onboarding skill that encodes the full new-seat workflow (identity check, invite, intro email, records update, follow-up gate), built the day I noticed I was re-deriving the steps every time.
• A biweekly tips email series teaching one feature at a time.
• A connector approval process with a read-only default, because anyone could otherwise grant the AI write access to a customer-facing system.

Screenshots 1

Screenshots 2

What was mine

The wave structure, the cutoff dates and every slip of them, the manager sign-off round, every team-facing communication, the policy one-pager, the read-only connector standard, and the deletion passes. Ownership held budget and exception authority and set the shared-account structure. Division managers made the individual keep or remove calls. My self-hosted agent ran the usage audit; I caught what it got wrong.

Decisions that mattered:

• Slipping the cutoff a week rather than breaking a promise. I'd committed in writing to no cold cutoff. The advance notice didn't go out while I was away, so I moved the date instead of defending it.
• Holding a seat rather than defaulting it when a manager hadn't answered. Removing someone's tool on silence is how you lose trust.
• Answering "can we re-add people later?" before asking managers to cut anyone. Two managers independently asked. Once I confirmed yes, both released seats they'd been holding onto. Reversibility unlocked the decision.
• Verify in the admin console, never infer from usage. The audit read "zero messages on the old platform" as "switched to the new one." Two people were mislabeled and neither had a working seat. That became a standing rule.

Problems worth knowing about

• Invites fail silently. One invite was broken for two and a half weeks and nobody knew, because the failure only shows on the invitee's side. It also explained why that person's old-platform usage had looked like resistance. Close-out rule changed from "an invite exists" to "the person has logged in."
• A friendly reply is not a certification. One attachment turned out to be a screenshot of a question. "I think I'm good to go" was about access, not training. Both went into the intake checklist.
• The record asserted a seat that didn't exist. One person sat in the user directory for three weeks with no invite behind it. She surfaced it herself. Fix: a cross-check rule against the mailing list.
• The image-editing gap. One employee used the old tool for photo mockups. I verified the gap was real, told him so first, then priced the options: a documented single-seat exception or a one-time landscape design tool. Validate the concern, then fit the tool to the workflow.

Results

Savings: ChatGPT monthly spend before to Claude Team spend after per month.

Honest gaps: certification lags seat count (9 of about 29 as of September), and eight logins were still unconfirmed at the time of writing. Adoption is the lagging indicator here, and it's not finished.

What I'd do differently

• Never make a notice depend on me being present. The cutoff slipped because the announcement was tied to my schedule. Schedule the send, or hand it off, before going anywhere.
• "Logged in" as the close-out standard from day one, not "invited."
• Get the connector governance conversation done before the first request arrives. Requests stacked up for weeks against no policy, and three were approved on the wrong premise before the read-only standard existed.
• Protect the tips cadence. Every-other-Monday became whenever-I-got-to-it. The series works when it's predictable.