AI adoption ยท Rollout
Put AI into everyday work
A rollout that works is mostly ordinary groundwork. Set up governed access, get the sources and their permissions right, run a bounded pilot against a baseline you measured first, train people on the work they already do, write down what worked, and hand over to somebody inside the company who owns it from then on.
Whichever platform you chose, the sequence is the same. The product-specific detail sits in the playbook for that platform.
The sequence
How does a rollout run?
Six steps, each with one condition that has to hold before the next one starts. The gates are the point. A rollout that skips the permissions gate usually discovers the problem in front of an auditor.
Sponsor and access
A named sponsor, the group that goes first, and centralised sign-in so accounts are created and removed in one place.
Gate. Someone owns the outcome, and there is a written answer to who gets access and who takes it away.
Source preparation
The documents and systems the assistant will read, with their permissions checked before it reads anything.
Gate. The sources are current, and nobody can reach through the assistant to something they could not open directly.
Bounded pilot
One group, a defined set of tasks, a fixed period, and the baseline already measured.
Gate. Success and stop criteria are written down and agreed before the first session.
Training by role
Sessions built on the work that group actually does, not a product tour. People leave having finished a real task.
Gate. Each role has at least one task it can now do differently, and knows which tasks to leave alone.
Reusable instructions
The prompts, templates and checks that worked get written down and shared, so the second person does not start from nothing.
Gate. The company method exists as something a new joiner can run, rather than in four people's heads.
Champions and handover
People inside each team who answer the everyday questions, and a named owner who runs the monthly review from then on.
Gate. The practice survives without us in the room.
The split
Who does what?
Some of this cannot be outsourced, and saying so early avoids the week where everybody waits for somebody else.
| Area | Your side | Our side |
|---|---|---|
| Sponsor and objective | Names the sponsor and the constraint. Decides who goes first. | Turns it into a pilot scope with criteria for stopping as well as continuing. |
| Accounts and sign-in | Owns the identity system and the joiner and leaver process, usually with the IT provider. | Specifies what the platform needs, configures it and tests it with you. |
| Permissions on content | Decides who should see what. This is a business decision and it stays with you. | Finds where the current state does not match that, and says so before anything is switched on. |
| Sources | Confirms which documents and systems are current and authoritative. | Prepares connections, tests what the assistant can and cannot reach, and reports the gaps. |
| Training | Gives people the time. An hour promised and not protected is the commonest reason adoption stalls. | Builds and runs the sessions on your own work, by role. |
| Evidence and approval | Names who signs off output that leaves the building. | Puts the review step into the workflow so the approval is recorded rather than assumed. |
What does the change look like on one ordinary task?
An illustration of the shape of the change. There are no measured figures in it, because the honest answer varies by company and by task.
Before
- Someone pastes a meeting transcript into whatever tool they have and asks for a summary.
- The output is decent, and it is different every time depending on who asked and how.
- Nothing records which version of the policy was applied, so the file note reads differently across the team.
- The person who leaves takes the way they did it with them.
After
- A shared instruction produces the file note in the company's own format, with the checks built into it.
- It pulls the client context from the system of record rather than from whatever was pasted in.
- The reviewer sees a consistent draft and the points it flagged as uncertain, and signs it.
- A new joiner runs the same instruction on day one and gets the same shape of output.
The review step does not disappear in the second column, and pricing it out is how a case that looked like a week of savings turns into two days. The business case calculator keeps it in.
Evidence
Has this been done?
Key Capital, an Irish corporate finance and investment management firm, rolled out in April 2026. The configuration work ran with the firm's own IT provider and data-protection team: single sign-on, the Microsoft 365 connector, access permissions and the governance boundaries. The firm set its acceptable-use policy and explained to its people why it was doing this. Softblues then led nine weekly training sessions built around the work that firm actually does.
That is the work the case describes. It does not report measured time savings, cost savings or financial return, which is why no such figure appears on this page.
Training and shared instructions take a team a long way. Where a workflow has to start on an event, run at volume or write back into a core system, that is a build rather than an assistant, and it belongs under AI Automation.
Common questions about rolling AI out
Why does a rollout need more than switching the licences on?
Because a licence buys access and nothing else. Setup, preparing the sources it reads, training and ongoing support all sit outside it, and they are what decide whether people use the thing. So does whether the permissions on those sources are right, whether anyone showed each group a task worth doing with it, and whether a named person looks at the result each month. Switching it on and waiting can produce a burst of curiosity that fades once the first busy week arrives.
Does single sign-on mean access is managed for us?
It means people sign in through your identity system rather than setting up separate accounts, which is worth having. It does not by itself remove somebody's access when they leave, and it does not mean the assistant sees the right things. Removing access on departure depends on how the account lifecycle is configured, and what the assistant can read depends on the permissions sitting on the content. Both are separate pieces of work, and treating sign-in as though it covered them is how a rollout ends up with an assistant that can read the folder nobody meant it to.
How do we train people without taking them off client work for a week?
Short sessions by role, built on the work that role already has in front of them. An hour with a group that shares the same task beats a whole-company session where most of the room is watching something they will never do. The time still has to be protected. An hour promised and quietly reclaimed by a deadline is a common way for a rollout to stall, and it is the one thing we cannot fix from outside.
What does a pilot need before it starts?
A measured baseline, a defined group, a defined set of tasks, a fixed period, and written criteria for both continuing and stopping. The baseline is the one that gets skipped. Take it before anyone starts using the tool, because a before state reconstructed afterwards is hard to defend and the conversation turns into whether it felt faster.
How do we stop it becoming ten people doing ten different things?
By writing down what worked. When somebody finds a way of producing the client file note that stands up to review, that becomes a shared instruction with the checks in it rather than a habit belonging to one person. That is the difference between a company where AI helps some individuals and a company whose method a new joiner can run on day one.
What happens when you hand over?
A named owner inside your company runs the monthly review: who is using it, whether the quality holds, what it is costing, and what the next decision is. Champions in each team answer the everyday questions. We stay involved where there is ongoing work to do, and the point of the handover is that the practice does not depend on us being in the room.
Already bought it, and nobody is using it?
That is a fixable problem and it rarely needs more licences. Tell us what was rolled out and to whom, and we will say where it stalled.
The first call covers what you want to improve, the tools you already run, and whether a scoped paid discovery is the right next step.