AI Solutions

What does a Copilot readiness assessment actually check?

Permissions come first, because Copilot inherits them exactly. So the assessment starts with SharePoint and OneDrive sharing settings, links set to 'anyone with the link', sites with no owner, guest accounts left over from projects that finished, and the company-wide groups that quietly grant everybody access to everything.

The rest follows from that. Data lifecycle: whether retention and disposal have ever been applied, because a tenant that has never deleted anything is one where a decade of superseded documents is exactly as available to be surfaced as this month's version.

Sensitivity labelling: whether the material that genuinely needs restricting is marked in a way the platform can act on. Identity and device posture: multi-factor authentication, Conditional Access and device compliance, because a Copilot licence raises the value of a compromised account. And the licensing position, which is the easiest part and the one most people start with.

Why does Copilot surface files people should not see?

Because it is doing exactly what it was designed to do. Copilot has no permissions model of its own: it queries content as the signed-in user, so anything they could already have found through search, they can now find by asking a question in plain English. The change is not access, it is discoverability. Files that were technically reachable but practically invisible, buried in a site nobody browses, twelve folders deep, named something meaningless, become one sentence away.

None of those are Copilot faults. All of them become Copilot's problem the moment you switch it on.

What has to be true before you assign the first licence?

A short list, and none of it is optional. Multi-factor authentication enforced for every user without exception. Sharing defaults set to named people rather than anyone with the link, with expiry on any anonymous link you do keep.

Company-wide access removed from sites that should never have had it. An owner assigned to every site and Team, because an unowned workspace has nobody who can make a decision about its content. Guest access reviewed by last activity rather than by whether the name still looks familiar.

Doing the permissions work first is not a delay to the project. It is the project. What follows it is comparatively straightforward.

What does a safe Copilot rollout look like, step by step?

Assess, remediate, pilot, measure, expand. The assessment produces a written picture of permissions, sharing, data and readiness. Remediation fixes what it found, and this is the phase that takes real time, because re-scoping site access and clearing legacy sharing is careful work that changes how people currently get at things. Only then does a pilot start, deliberately small and deliberately chosen: a group with real, repetitive, document-heavy work, rather than whoever asked first.

The pilot has to be measured against something. Pick two or three tasks the group genuinely does, record roughly how long they take today, and compare afterwards. Then expand by role rather than by enthusiasm, because the value is uneven: people who spend the day in Outlook, Word, Teams meetings and SharePoint get far more from it than people who spend the day in a line-of-business application Copilot cannot see. Reviewing usage after each expansion, and reclaiming licences from users who stopped, is what stops this becoming an annual renewal for something nobody opens.

How does Copilot licensing interact with what you already own?

Microsoft 365 Copilot is an add-on rather than a plan in its own right, so it sits on top of a qualifying Microsoft 365 subscription and is assigned per user. Two consequences are worth planning for.

First, the base licence you hold is part of the decision, and where a tier change is needed it is far easier to handle at renewal than mid-term. Second, it is a per-user cost that runs whether or not anyone opens it, so who receives a licence is a commercial decision rather than an inclusive gesture.

There is also more than one thing called Copilot, and conflating them causes most of the confusion in a licensing conversation. A general web chat experience is not the same as the paid Microsoft 365 Copilot that reaches into your own tenant content, and neither is the same as a purpose-built agent published to a named audience.

Establishing which one a request actually refers to settles half the argument before price comes up. The wider question of what you own, what you use and what you are paying twice for sits on our licensing and cost management page, and right-sizing there frequently funds a good part of a Copilot pilot.

What does the published evidence say about Copilot saving time?

There is now real published evidence rather than only vendor case studies, and it is worth reading carefully because it is more modest and more useful than the marketing.

HMRC published a Phase 3 evaluation covering 3,000 randomly allocated licences plus 500 for volunteers, run from September to December 2024 with 1,364 survey responses. Staff self-reported saving 2 to 3% of their working week, which HMRC put at around 60 minutes, and HMRC then reduced its own headline figures by about 20% to account for non-users and response bias. A separate cross-government trial covering more than 20,000 civil servants reported around 26 minutes a day.

Those two numbers are not comparable, and the gap between them is the point. Different populations, measured differently, both self-reported. Anyone quoting one of them at you as the figure your business will achieve is selling rather than advising, and we are not going to do it either.

That last figure is the one worth sitting with, because it is the readiness argument made by the users themselves. Nearly half the people who never touched a licence their employer had already paid for stayed away because they were not confident about what it could reach. That is not a training problem and it is not solved after rollout: it is the permissions and governance work, done first, and it is the difference between licences that get used and licences that get quietly abandoned.

How do you tell whether Copilot is actually delivering value?

By choosing how you will measure it before the rollout rather than after. The mistake is to look for a single productivity percentage, which nobody can produce honestly. What works is a small number of concrete tasks with an observable before and after: how long a first draft of a routine report takes, how long it takes someone to catch up on a meeting they missed, how long the monthly summary takes to assemble.

Ask the pilot group to record the starting position candidly, including how much they dislike the task, because time recovered from work people resent is worth more than the minutes suggest.

Then watch usage rather than licence count. Active use per person, which applications it is being used in, and how many licensed users have stopped entirely are the numbers that predict whether renewal is worth it. A licence assigned to somebody who used it twice in March is a recurring cost with no return, and reclaiming it is not a failure of the project, it is the project working properly.

What should an AI use policy actually say?

Less than most drafts, and far more specifically. A policy people will follow states which tools are approved and which are not, what categories of information must never be pasted into an unapproved tool (client data, personal data, anything under NDA, credentials, unpublished financials), that AI output is a draft and the person who sends it owns it, and where to ask when something is unclear. A policy that opens with a definition of machine learning is one nobody reads to the end of.

The part most often missing is the route to yes. If there is no approved way to use a tool people find genuinely useful, they will use it anyway on a personal account, and you lose the control and the visibility together.

Naming approved tools and providing a simple request process for new ones converts shadow AI into something you can see. Where automated output feeds a decision about a person, whether that is recruitment, credit or anything else with a consequence, the policy also needs to say who reviews it, because that is a legal position rather than a preference.

What is shadow AI, and how do you find it?

Shadow AI is the use of AI tools nobody approved, on accounts nobody manages, with company information nobody agreed to share. It is almost always well intentioned: somebody had a tedious task and found something that helped.

The risk is not the intent, it is that the data has left your control, the output has no audit trail, and when a client or an auditor asks whether their information has been put into a third-party model, the honest answer is that you do not know.

Finding it starts with signals you already hold rather than with a survey, because a survey asks people to confess. Sign-in and application consent records in Entra ID show which third-party applications have been granted access to your tenant and by whom.

Expense claims and card statements show personal subscriptions. Browser and network telemetry, where you have it, covers the rest. What matters more than the inventory is what happens next: a genuine approved option offered quickly does more to reduce shadow AI than any prohibition.

Where does Copilot help, and where does it genuinely not?

It helps most where the work is language-shaped and the source material already lives in Microsoft 365. Summarising a long thread or a meeting you missed, producing a first draft from documents that already exist, finding the thing you know you read but cannot locate, turning rough notes into something presentable, and restructuring content you already have. In those tasks it changes the shape of the work: instead of starting from nothing you start from something imperfect, which for most people is considerably easier.

It helps least where the answer depends on data it cannot see, or on being exactly right. Anything sitting in a line-of-business system outside Microsoft 365 is invisible to it unless deliberately connected. Numerical work needs checking, because a confident wrong figure looks identical to a confident right one.

And any output going to a client, a regulator or a court needs a human who is accountable for it, which is a governance decision rather than a technical one. Saying this plainly at the start protects the project, because expectations set too high in month one are the most common reason adoption stalls by month three.

What are Copilot agents, and what changes when you publish one?

An agent is a Copilot pointed at a defined set of content, given instructions, and published to a specific audience. That focus is what makes agents useful, and it is also what makes three questions unavoidable: what content it can reach, who can use it, and who owns it.

An agent holds a standing pointer at company data, so one published broadly against a site with loose permissions is the oversharing problem again, in a form that is harder to notice because it presents as a small helpful tool rather than as a search index.

The workable position is that every published agent has a named owner, a stated purpose and a review date recorded somewhere central, and that it joins the same review cycle as unowned sites and stale guest accounts. Start narrow. An agent scoped to one well-governed document library and used by one team is a good first agent. An agent scoped to everything is a governance problem with a friendly interface.

When do you not need this, and where does our scope end?

You do not need a readiness engagement if your tenant is small and recently built, sharing is already locked down, multi-factor authentication is everywhere and every site has an owner. In that position, buy a handful of licences, run a pilot and measure it.

We do not sell AI as a headcount reduction, and we will say so when the honest answer is that a particular team will get little from it.

Where the requirement is AI answering from your own data, or built into your own application, that is a different discipline and it sits on our AI engineering page rather than this one.