AI does not create new risk. It exposes the risk you already had.
Who pays for the damage when a decision made under AI goes wrong?
Where does work created with AI stop being yours and start being plagiarism?
What is the difference between being accountable for AI and being responsible for it? And can you tell which one you are?
If those are sitting somewhere on your list, you are in reasonable company. They are not my questions. They came off a flipchart on Wednesday morning, put there by eight executives who had come to work out what to do about AI in their own businesses.
We ran "AI in Business: Who Owns the Risk?" at Microsoft, and what was good to see was how quickly the room opened up. Not polite questions. The kind that need answering.
Running a roundtable with executives can be challenging, and a little intimidating. What is worth remembering is that people are in the room because they have questions they need answered, and understanding those questions matters more than standing at the front pushing a product.
Most suppliers walk into a room like that with a silver bullet. One of the participants said it plainly: most of the AI sessions they had attended were really just a product push. There are a great many products on the internet now with an AI label on them. Some are the same old product with a new suit on. Others are hastily assembled, and missing most of the governance and risk management that should sit underneath them.
So this was as much an exercise in understanding and exploring for us as it was for the people who came to get their own questions answered.
The questions, and what they group into
We set the scene with a three-part approach. Map it. Measure it. Manage it. Simple as a mindset, considerably more involved as a process.
Then we asked the room what they wanted solved. What came back was practical and specific.
How do you measure AI, risk and ROI in a business?
How do you deal with policies and automated decisions? Who owns the accountability for a decision made through AI, or improved by a human using AI?
Who pays for the risk or the damage?
Where does self-created work end and plagiarism begin?
How do you quantify the risk, and measure the exposure that already exists?
What is the difference between a claim and evidence?
Which governance frameworks apply? Is this business process and good accounting practice, or do King V, ISO 42001 and ISO 27001 have a place in AI management?
How do you match accountability with input? And what actually separates accountability from responsibility?
Then a run about visibility. How do we surface the risk. How do we know what is risky and what is not. How do we see the risk in data management, mitigate it, and monitor it. How do we make sure governance processes are followed.
Tough questions. They group into three:
How do we create visibility. How do we understand the impact and the risk. And what do we do with the answers once we have them.
Which is Map it, Measure it, Manage it, arrived at by the room rather than handed to it.
Why a fingerprint
The analogy that kept fitting was a fingerprint.
A fingerprint tells you specific things. Who has been there. What they touched, and what they interacted with. It starts to paint a picture of what happened in an environment.
It is not somebody's view of what happened. It is not a claim, and it is not a description of what should have happened. It is an evidential trace of something that did happen.
That distinction ran through the whole morning, and it is the difference between a policy and a position you can defend.
Everyone went home with a UV torch and a card marked in invisible ink. Under normal light, the card is blank. Under the torch, a fingerprint.
We suggested they keep the torch as a reminder of the morning. Or, if they were feeling less charitable, to establish who has been raiding the cookie jar at home.
The point is not the trick. The trace was on the card the whole time, sitting in plain view on the table in front of them. What was missing was the light.
That is a fair description of where most businesses are with AI. The evidence of what is running, who is using it and what data it touches already exists, in logs, in billing, in the tools you are already paying for. It stays invisible until you point the right thing at it.
Which is the part worth holding onto: this reads like a governance conversation because we were in a room with executives, but every step of it runs on technology you either already own or can turn on.
Map it
AI enters a business through three doors.
Self-adoption. Personal accounts, personal payment, no procurement, no policy review. Somebody trying to make their Tuesday a little better and a little faster.
Software you already run. Your ERP, your finance system, your productivity tools. All of them now have AI embedded somewhere, in some flavour.
Vendor release schedules. The steady stream of updates, new features, new connectors and new ways of joining data sets together.
Which surfaced something the room felt immediately: their AI risk is largely being managed on somebody else's schedule rather than their own.
The discussion moved quickly to the realisation that nobody is working on green fields here. The fingerprints are already all over the business. The concern was how to catch up, and how to know what is actually there.
This is where the torch stops being a metaphor. The tooling that makes this visible is not exotic and in most cases you are already licensed for it. Microsoft and the other major providers have discovery and reporting that will tell you which AI services are in use, by whom, and what data is moving through them.
You also do not need a paid tool to start.
One of the more technically minded people in the room made the suggestion, and it was a good one. Pull your DNS, proxy and firewall logs and look at where the traffic is going. It will not tell you what people are doing, but it will tell you where they are going, and it takes an afternoon.
Then run a survey. Draft a simple mail asking your employees to help you understand where data is being used and shared on AI platforms, so the business can deal with it properly.
Which is where the conversation turned.
The part about people
Many organisational cultures are hesitant, and in places fearful, about telling management where AI is being used.
The concern raised most directly was employees losing their jobs to AI. If you demonstrate how much faster the work can be done, you may be demonstrating that fewer people are needed to do it.
There was a second thread underneath it. AI can give an employee the impression, and give their colleagues the impression, that they are suddenly the expert in an area. Often it is shallow. You see it in email: long, complex, confident messages from someone who does not have the grounding behind what they have sent. It is usually three paragraphs longer than it needs to be, and somewhere in the middle there is a word nobody in the business has ever said out loud.
Both of those make people less willing to tell you what they are using. And you cannot map what people will not tell you about.
Which is why the survey is not really a survey. Bringing your people into the mapping is the start of bringing them into the whole thing. If the first time your teams hear the business's position on AI is when a policy arrives, you have lost them before you have started.
This is a change management exercise as much as a technology one. The fear does not get resolved by tooling. It gets resolved by involving people early enough that they can see where they fit in what comes next, and by being straight with them about what you are doing and why.
Measure it
Once you have found the traces, measurement runs on two prongs.
The first is the obvious one. How much is being used, what is being used, what it costs, how much data is going through it, what kind of data, and what the security exposure looks like. Again, largely a tooling question, and largely answerable with what you have.
The second is the one the room spent longest on, and it was the most uncomfortable conversation of the morning. Is any of it actually adding value?
Everybody assumes AI improves productivity. The global numbers are mixed at best, and only the strongest performers are showing meaningful returns.
Here is the example that landed hardest. A business spends money on AI tooling. The employees genuinely do execute faster. But the work produces the same return it did before. The customer is not billed more. Satisfaction may improve, quality may improve, and revenue has not moved.
Which means the ROI is now lower than it was before the AI was introduced, not higher.
That opened the more useful question. If you are aiming at productivity improvement, what are you going to do with the capacity you get back?
That is the part worth planning. Recovered time only becomes return if it is pointed at something: more inbound work, more value-adding conversations with customers, capacity to take on projects you were turning away. Saving time is not the outcome. What you do with the saved time is the outcome.
And it has to be measured. The other measures that came up were customer satisfaction, quality of work, errors reduced, risk reduced or exposed, revenue, conversion rate and time to market.
The point is not to go in blindly and hope that unleashing AI produces a return on its own. Over the last eighteen months plenty of proofs of concept have succeeded, and the two concerns that have not gone away are how to govern and manage the risk, and how to measure the return.
Those are the two we had to unpack.
Two scenarios worth sitting with
Somebody resigns on Friday. Could you establish what they put into a public AI tool over their last month?
A supplier switches on AI summarisation across your customer records. When would you find out? Who would you tell? And what does it mean for your customer, whose data is now being read and summarised by something nobody told them about?
The measurement problem gets complicated quickly, and it always comes back to the same things. Where the data sits. Who has access to it. What the AI is doing with it. And where it goes next.
Which took us back to a distinction that mattered to this room more than any other:
You cannot outsource accountability. You can only outsource responsibility.
Somebody else can be responsible for operating the tool, holding the data or running the process. The accountability stays with you.
Manage it
So you have found the fingerprints and you can read what they tell you. What do you do with it?
The first answer is what not to do. Do not come in and try to suppress the use of AI. Bring your employees into it instead. They already know where the friction is in your business. That is why they went looking for AI in the first place.
Then build a cycle.
Review the traces. Make sure every one of them has a named owner. Be clear on who owns each data point, how it gets measured and how the exposure gets reduced. Then report it to somebody outside your own team, in language they can read and act on.
Make it a rhythm you can hold. Quarterly is a sensible place to start. Map the environment using the tooling and the surveys. Measure it using the tooling and the insight from your teams. Then decide what you allow, what you do not, and how you manage the exposure. Train your people around those decisions.
The reason this works as a rhythm rather than a project is that the reporting can be automated. If the dashboard builds itself every quarter, the review is a conversation about what changed. If somebody has to assemble it by hand, it will happen twice and then stop.
And keep going back to your teams for where the friction is, because that is where AI will do the most good.
There is one more thing, and it is the difference between keeping up and getting in front.
The cycle above keeps you current. It tells you what is in the estate, who owns it and whether it is safe. What it does not tell you is where you intend to win.
That part has to come from you. Every quarter, each item on the list should resolve to one of three things: scale it, put conditions on it, or retire it. And that call should be measured against where the business is trying to get to, not against how nervous the item makes you feel.
The loop keeps you current. The thesis is what puts you in front. Run the loop without one and you have a very well documented business that is still following.
AI, or automation?
One more thing came out of this, and it is becoming a grey area.
The default assumption is that AI is the answer to everything. Looking honestly at the problems being discussed, most of them did not need AI. They needed automation. AI is what closes the gap on the parts automation cannot reach.
Knowing which one you are actually buying is worth a great deal of money.
The demonstration
After we had worked through mapping, measuring and managing, Inayeth Govender and Kamil Hirjee took the room through the tooling that sits behind each step.
This is the part worth being direct about, because it is what we do. Each step has technology behind it, and for most businesses it is technology they already hold licences for and have never switched on. Discovery and reporting for the mapping. Data classification, sensitivity labelling and audit for the measuring. Dashboards and scheduled reporting for the managing, so the cycle runs without somebody rebuilding it by hand every quarter.
The half that mattered most was not the risk half. Seeing adoption and usage as a business signal, who is using what, how much of it, and where it is actually landing, is what turns this from a control exercise into a performance one. A demonstration that only proves control proves you are safe. It does not prove you are ahead.
The session drew plenty of questions, particularly around data security, sensitivity labels, and handling those across multiple platforms, services and data sources. That was the area the room wanted most, and it is usually where the gap between what a business owns and what it actually uses is widest.
What people took away
The session closed on what they had learned and what they would do next.
That adoption needs to be deliberate. That there is real preparation to do before deploying AI, and that the controls and governance which should already have been in place are exactly what surfaces when you start looking.
That quicker does not automatically mean better.
That not all AI is equal, and that data privacy, data sharing and how models are trained deserve proper attention.
And the one that summed up the morning: AI does not necessarily create new risk in your business. It exposes the risk that was already there.
What they said they would do immediately:
Work out their usage and follow the breadcrumbs. Understand their data, and who is using what. Measure the exposure and the risk. Document what the return could be for one focused project rather than for AI in general. Bring employees and stakeholders along rather than doing it to them. And make it a regular cadence instead of a one-off exercise.
Where we come in
This was not a product pitch, and the follow-up is not either. It is worth being precise about which parts of this you should hand over and which you should keep.
Seeing it, evidencing it and watching it is a job for technology. It has to run continuously and it cannot depend on anyone remembering to do it. An estate that changes every week cannot be watched by hand.
Reading the signal, sorting it and deciding is judgement, and most businesses want a hand with that. Turning usage data into a funding decision is a business call rather than a technical one. That is what our diagnostic is actually for: an executive conversation much like the one we had in that room, followed by proper due diligence to map and measure what your business is exposed to, using the tooling you already own wherever we can.
The thesis and the ownership stay with you. Where your business intends to win, and who carries the risk, are the two things worth keeping in-house.
The third one is what earns the first two.
The feedback on the day was that people were surprised by how little they knew of what is running in their own environment, and equally surprised by how much is already available to map and manage it.
If you want the worksheet the room used, reply and I will send it over. If you want the conversation with your own team around the table, that is what we do next.
Comments