What AI Is Actually Compressing in Cross-Functional Work — and What It Isn’t.
Most cross-functional “sprints” are not sprints. Product writes a brief. Engineering estimates. Legal reviews. Marketing waits. Finance pushes back. Six weeks later you have a deck, a revised timeline, and a team that has quietly forgotten why the initiative mattered.
If you have run a programme office, you already know the diagnosis: the delay is seldom in the work. It is in the gaps between the work.
Rob Cross, Reb Rebele and Adam Grant put numbers to this in Harvard Business Review back in 2016. Their finding, drawn from two decades of data across more than 300 organisations, was that time spent on collaborative activity had risen by 50% or more, and that at many companies people spend around 80% of their time in meetings or answering colleagues’ requests. Cross’s later work put the figure for managers at 85% or higher. That is the real shape of the problem. The analysis is not slow. The waiting is slow.
So the interesting question about AI in cross-functional work is not “does it draft faster.” It is: does it actually remove handoffs, or does it just make each handoff produce a longer document?
The honest answer, based on what is currently on the public record, is: sometimes the former, often the latter, and the evidence is thinner than the headlines suggest.
Here is what I can verify, and what I cannot.
Four Documented Cases — and How Much Weight Each Can Bear.
A note on sourcing before we start, because it matters.
Three of the four cases below are customer stories published by Microsoft. These are marketing assets. They are not audited; the metrics are self-reported by the customer, and no vendor publishes the deployments that went nowhere.
I am including them because they are the most detailed public accounts available, not because they are evidence in a strong sense. Read them as existence proofs — this happened somewhere — rather than as base rates.
1. BOQ Group (Bank of Queensland): review cycles collapsed.
Australia’s Bank of Queensland deployed Microsoft 365 Copilot across its operations. On Microsoft’s account: business risk reviews moved from three weeks to one day, training programme creation from three weeks to one day, and report sign-off from four weeks to one week. More than 70% of users reported saving 30 to 60 minutes a day.
The number I find more credible than the headline ones is the granular one: BOQ reported cutting analyst time on root cause analyses by roughly half, with a measured improvement in the quality of findings. That is a specific, testable claim about a specific process.
What it does not tell you: whether the one-day risk review received the same scrutiny as the three-week one, or whether the compression came from AI or from the process redesign that accompanied it.
Source: Microsoft Cloud Blog and Microsoft Customer Stories, 2024–25.
2. Motor Oil Group: the sequential relay, shortened.
Motor Oil Group, a Greek energy company, deployed Copilot across communications, HR, finance and operations. Tasks that consumed weeks — executive speeches, training programmes, job descriptions, financial documentation — now reportedly take minutes.
Worth noting: this is an 1,800-employee company with a small central IT team, and the story dates from May 2024. It is a good illustration of a lean function extending its reach. It is not evidence about how a 40,000-person matrixed enterprise behaves.
Source: Microsoft Customer Stories, May 2024.
3. A global media company (via TiER1): faster ideation, unquantified.
TiER1 Performance worked with a 10,000+ employee media organisation on AI-facilitated ideation. Synthetic focus groups mirroring different demographics and buyer personas enabled rapid concept testing; iterative sessions blended publishing-team expertise with AI-generated insight. TiER1 reports products launched more quickly and new voices entering product decisions.
I want to be precise here, because this is where I have seen this case may have misreported: TiER1’s published case study contains no timeline figures at all. No “days instead of quarters.” The outcomes are described qualitatively. If you see a specific compression ratio attached to this case, it was possibly an afterthought somewhere downstream.
Source: TiER1 Performance case study, “AI-Enabled Product Innovation with Market Relevance.”
4. Mondra: automation, not coordination.
Mondra Global built an AI copilot called Sherpa for Life Cycle Assessments, cutting LCA timelines from weeks or months to about four hours across a dataset of 60,000+ products and over a million ingredients.
This one is genuinely impressive — and it is a different phenomenon from the other three. Mondra did not compress a cross-functional meeting cycle. It automated a data-intensive computation that previously required specialist labour. The bottleneck removed was calculation, not coordination. Filing it under “cross-functional collaboration” flatters the argument.
Source: Microsoft Customer Stories; Mondra press announcement, October 2024.
The Number You Should Hold Alongside These.
MIT’s Project NANDA, in The GenAI Divide: State of AI in Business 2025, examined 300 public AI deployments alongside leader interviews and employee surveys. Roughly 5% of pilots produced rapid measurable impact on the P&L. The rest stalled.
The methodology has been fairly criticised — small samples, a six-month measurement window, directional rather than definitive figures. I would not treat 95% as a law of nature. But the direction is consistent with what most operators report privately, and the study’s explanation of why is the useful part: the failures were not model failures. They were integration failures. Tools that could not retain context, could not learn from feedback, and were bolted onto workflows nobody redesigned.
Put the two bodies of evidence together, and the pattern is clear enough:
AI reliably compresses the drafting layer. It compresses the coordination layer only where someone deliberately redesigned the coordination.
BOQ did not get a one-day risk review by installing Copilot. It got one by rethinking who reviews what, in what order, with AI producing the first pass. The tool was necessary and nowhere near sufficient.
A Practical Way to Test This in Your Own Organisation.
If you want to find out whether this applies to you, do not run a transformation programme. Run one contained experiment.
Pick a decision, not a process. Choose a single cross-functional decision that currently takes two to four weeks — quarterly resource allocation, campaign approval, a feature prioritisation call. The deliverable at the end is a decision, not a document.
Instrument the baseline first. Before you change anything, measure the current cycle honestly: elapsed time from kickoff to decision, number of handoffs, and the number of days where the item sat waiting on someone. Most teams skip this and then cannot prove anything afterwards. This step is the difference between a result and an anecdote.
Pre-load the context. Use AI to synthesise the prior decisions, the relevant data and the constraints before anyone joins. No one should spend hour one asking what we are deciding.
Convert sequential review into parallel review. This is the actual mechanism. Legal, finance, and risk look at the same AI-produced first draft simultaneously rather than in a relay. The AI is doing the drafting; the humans are doing the judging.
Name what must not be compressed. Decide in advance which reviews are load-bearing — anything with regulatory, safety, or irreversible commercial consequence — and protect them explicitly. A compressed cycle that quietly skips a control is not a win; it is a deferred incident.
Measure coordination velocity, not output volume. Time from kickoff to decision. Number of handoff delays. Rate of decisions later reversed. That last one is the honest metric, and almost nobody tracks it.
The Part Nobody Puts in the Case Study.
Speed has a cost profile, and it is worth naming.
Fewer chances to object. A six-week review cycle is wasteful, but it has one accidental benefit: the person who has doubts gets three or four separate moments to speak up. Someone who misses the first meeting can raise the concern at the second. Compress that cycle into two days, and they get one chance — and only if they happen to be free and paying attention that day.
So before you speed anything up, ask an honest question about your own organisation: when someone junior disagrees with a decision, do they actually say so? If the answer is already “not often,” a faster calendar will make it rarer still.
A polished draft is not a checked draft. This is the risk that catches experienced people out. When AI writes a risk review, it produces something that reads exactly like a well-researched risk review — clear structure, confident language, the right professional vocabulary. It looks finished. And because it looks finished, the reviewer relaxes, skims it, and signs. But the polish came from the writing, not from the thinking. A document that has been checked properly and one that has not now look identical on the page. That means your reviewers need to read more carefully than they used to, not less — and in my experience, nobody starts doing that on their own. You have to ask them to.
The most useful knowledge is usually the unwritten kind. Sometimes a slow review catches a genuine problem, and it is worth understanding why. Usually it is because somebody in the room has watched this exact thing go wrong before at a previous employer. They cannot always explain how they know; they just recognise the shape of it. That kind of judgement is not in the project file, the policy manual, or last quarter’s report. AI can only work with what somebody wrote down — and a great deal of what your most experienced people know has never been written down anywhere.
And in the Indian context specifically: most of the organisations, in my opinion, are not slow because of handoff friction. They are slow because approval authority sits three levels above where the information sits.
AI will not fix a delegation problem. It will simply produce beautifully drafted material that waits just as long for a signature. If your cycle time is a governance artefact rather than a workflow artefact, start there — the tooling conversation is premature.
The Harder Question.
Assume it works. Assume you take a six-week cycle down to a week. The competitive advantage does not come from the six weeks you saved. It comes from what you do with them.
- If we can plan a quarter in a fraction of the time, what does that free us to actually examine?
- If compliance review takes hours instead of weeks, does that change our risk appetite — and have we consciously decided that it should?
- If cross-functional alignment stops being the bottleneck, what becomes the bottleneck? In most organisations, it may just be the decision rights, and that is a much harder problem than tooling.
Compression is not the destination. It is a diagnostic. It tells you what your organisation was actually spending its time on — and whether it was ever worth the wait.
What’s one cross-functional process in your organisation that takes six weeks and shouldn’t? More usefully — do you know why it takes six weeks? Please tell me in the comments.
What AI Is Actually Compressing in Cross-Functional Work — and What It Isn’t.
Most cross-functional “sprints” are not sprints. Product writes a brief. Engineering estimates. Legal reviews. Marketing waits. Finance pushes back. Six weeks later you have a deck, a revised timeline, and a team that has quietly forgotten why the initiative mattered.
If you have run a programme office, you already know the diagnosis: the delay is seldom in the work. It is in the gaps between the work.
Rob Cross, Reb Rebele and Adam Grant put numbers to this in Harvard Business Review back in 2016. Their finding, drawn from two decades of data across more than 300 organisations, was that time spent on collaborative activity had risen by 50% or more, and that at many companies people spend around 80% of their time in meetings or answering colleagues’ requests. Cross’s later work put the figure for managers at 85% or higher. That is the real shape of the problem. The analysis is not slow. The waiting is slow.
So the interesting question about AI in cross-functional work is not “does it draft faster.” It is: does it actually remove handoffs, or does it just make each handoff produce a longer document?
The honest answer, based on what is currently on the public record, is: sometimes the former, often the latter, and the evidence is thinner than the headlines suggest.
Here is what I can verify, and what I cannot.
Four Documented Cases — and How Much Weight Each Can Bear.
A note on sourcing before we start, because it matters.
Three of the four cases below are customer stories published by Microsoft. These are marketing assets. They are not audited; the metrics are self-reported by the customer, and no vendor publishes the deployments that went nowhere.
I am including them because they are the most detailed public accounts available, not because they are evidence in a strong sense. Read them as existence proofs — this happened somewhere — rather than as base rates.
1. BOQ Group (Bank of Queensland): review cycles collapsed.
Australia’s Bank of Queensland deployed Microsoft 365 Copilot across its operations. On Microsoft’s account: business risk reviews moved from three weeks to one day, training programme creation from three weeks to one day, and report sign-off from four weeks to one week. More than 70% of users reported saving 30 to 60 minutes a day.
The number I find more credible than the headline ones is the granular one: BOQ reported cutting analyst time on root cause analyses by roughly half, with a measured improvement in the quality of findings. That is a specific, testable claim about a specific process.
What it does not tell you: whether the one-day risk review received the same scrutiny as the three-week one, or whether the compression came from AI or from the process redesign that accompanied it.
Source: Microsoft Cloud Blog and Microsoft Customer Stories, 2024–25.
2. Motor Oil Group: the sequential relay, shortened.
Motor Oil Group, a Greek energy company, deployed Copilot across communications, HR, finance and operations. Tasks that consumed weeks — executive speeches, training programmes, job descriptions, financial documentation — now reportedly take minutes.
Worth noting: this is an 1,800-employee company with a small central IT team, and the story dates from May 2024. It is a good illustration of a lean function extending its reach. It is not evidence about how a 40,000-person matrixed enterprise behaves.
Source: Microsoft Customer Stories, May 2024.
3. A global media company (via TiER1): faster ideation, unquantified.
TiER1 Performance worked with a 10,000+ employee media organisation on AI-facilitated ideation. Synthetic focus groups mirroring different demographics and buyer personas enabled rapid concept testing; iterative sessions blended publishing-team expertise with AI-generated insight. TiER1 reports products launched more quickly and new voices entering product decisions.
I want to be precise here, because this is where I have seen this case may have misreported: TiER1’s published case study contains no timeline figures at all. No “days instead of quarters.” The outcomes are described qualitatively. If you see a specific compression ratio attached to this case, it was possibly an afterthought somewhere downstream.
Source: TiER1 Performance case study, “AI-Enabled Product Innovation with Market Relevance.”
4. Mondra: automation, not coordination.
Mondra Global built an AI copilot called Sherpa for Life Cycle Assessments, cutting LCA timelines from weeks or months to about four hours across a dataset of 60,000+ products and over a million ingredients.
This one is genuinely impressive — and it is a different phenomenon from the other three. Mondra did not compress a cross-functional meeting cycle. It automated a data-intensive computation that previously required specialist labour. The bottleneck removed was calculation, not coordination. Filing it under “cross-functional collaboration” flatters the argument.
Source: Microsoft Customer Stories; Mondra press announcement, October 2024.
The Number You Should Hold Alongside These.
MIT’s Project NANDA, in The GenAI Divide: State of AI in Business 2025, examined 300 public AI deployments alongside leader interviews and employee surveys. Roughly 5% of pilots produced rapid measurable impact on the P&L. The rest stalled.
The methodology has been fairly criticised — small samples, a six-month measurement window, directional rather than definitive figures. I would not treat 95% as a law of nature. But the direction is consistent with what most operators report privately, and the study’s explanation of why is the useful part: the failures were not model failures. They were integration failures. Tools that could not retain context, could not learn from feedback, and were bolted onto workflows nobody redesigned.
Put the two bodies of evidence together, and the pattern is clear enough:
AI reliably compresses the drafting layer. It compresses the coordination layer only where someone deliberately redesigned the coordination.
BOQ did not get a one-day risk review by installing Copilot. It got one by rethinking who reviews what, in what order, with AI producing the first pass. The tool was necessary and nowhere near sufficient.
A Practical Way to Test This in Your Own Organisation.
If you want to find out whether this applies to you, do not run a transformation programme. Run one contained experiment.
Pick a decision, not a process. Choose a single cross-functional decision that currently takes two to four weeks — quarterly resource allocation, campaign approval, a feature prioritisation call. The deliverable at the end is a decision, not a document.
Instrument the baseline first. Before you change anything, measure the current cycle honestly: elapsed time from kickoff to decision, number of handoffs, and the number of days where the item sat waiting on someone. Most teams skip this and then cannot prove anything afterwards. This step is the difference between a result and an anecdote.
Pre-load the context. Use AI to synthesise the prior decisions, the relevant data and the constraints before anyone joins. No one should spend hour one asking what we are deciding.
Convert sequential review into parallel review. This is the actual mechanism. Legal, finance, and risk look at the same AI-produced first draft simultaneously rather than in a relay. The AI is doing the drafting; the humans are doing the judging.
Name what must not be compressed. Decide in advance which reviews are load-bearing — anything with regulatory, safety, or irreversible commercial consequence — and protect them explicitly. A compressed cycle that quietly skips a control is not a win; it is a deferred incident.
Measure coordination velocity, not output volume. Time from kickoff to decision. Number of handoff delays. Rate of decisions later reversed. That last one is the honest metric, and almost nobody tracks it.
The Part Nobody Puts in the Case Study.
Speed has a cost profile, and it is worth naming.
Fewer chances to object. A six-week review cycle is wasteful, but it has one accidental benefit: the person who has doubts gets three or four separate moments to speak up. Someone who misses the first meeting can raise the concern at the second. Compress that cycle into two days, and they get one chance — and only if they happen to be free and paying attention that day.
So before you speed anything up, ask an honest question about your own organisation: when someone junior disagrees with a decision, do they actually say so? If the answer is already “not often,” a faster calendar will make it rarer still.
A polished draft is not a checked draft. This is the risk that catches experienced people out. When AI writes a risk review, it produces something that reads exactly like a well-researched risk review — clear structure, confident language, the right professional vocabulary. It looks finished. And because it looks finished, the reviewer relaxes, skims it, and signs. But the polish came from the writing, not from the thinking. A document that has been checked properly and one that has not now look identical on the page. That means your reviewers need to read more carefully than they used to, not less — and in my experience, nobody starts doing that on their own. You have to ask them to.
The most useful knowledge is usually the unwritten kind. Sometimes a slow review catches a genuine problem, and it is worth understanding why. Usually it is because somebody in the room has watched this exact thing go wrong before at a previous employer. They cannot always explain how they know; they just recognise the shape of it. That kind of judgement is not in the project file, the policy manual, or last quarter’s report. AI can only work with what somebody wrote down — and a great deal of what your most experienced people know has never been written down anywhere.
And in the Indian context specifically: most of the organisations, in my opinion, are not slow because of handoff friction. They are slow because approval authority sits three levels above where the information sits.
AI will not fix a delegation problem. It will simply produce beautifully drafted material that waits just as long for a signature. If your cycle time is a governance artefact rather than a workflow artefact, start there — the tooling conversation is premature.
The Harder Question.
Assume it works. Assume you take a six-week cycle down to a week. The competitive advantage does not come from the six weeks you saved. It comes from what you do with them.
- If we can plan a quarter in a fraction of the time, what does that free us to actually examine?
- If compliance review takes hours instead of weeks, does that change our risk appetite — and have we consciously decided that it should?
- If cross-functional alignment stops being the bottleneck, what becomes the bottleneck? In most organisations, it may just be the decision rights, and that is a much harder problem than tooling.
Compression is not the destination. It is a diagnostic. It tells you what your organisation was actually spending its time on — and whether it was ever worth the wait.
What’s one cross-functional process in your organisation that takes six weeks and shouldn’t? More usefully — do you know why it takes six weeks? Please tell me in the comments.
Sources:
- Cross, R., Rebele, R., and Grant, A. “Collaborative Overload.” Harvard Business Review, January–February 2016. hbr.org/2016/01/collaborative-overload
- Cross, R. “Collaboration Overload Is Sinking Productivity.” Harvard Business Review, September 2021.
- Microsoft. “AI-powered innovation: How leading organisations are shaping the future.” Microsoft Cloud Blog, 30 June 2025. (BOQ Group figures)
- Microsoft. “Bank of Queensland evolves operations, delivers business value with Microsoft Copilot.” Microsoft Customer Stories, 2024. (Vendor-published; metrics self-reported by customer.)
- Microsoft. “From HR to marketing, Motor Oil Group eliminates mundane tasks with Microsoft Copilot.” Microsoft Customer Stories, May 2024. (Vendor-published.)
- TiER1 Performance. “AI-Enabled Product Innovation with Market Relevance.” tier1performance.com/case-studies/ (No timeline metrics published.)
- Microsoft. “Mondra Global Limited facilitates transition to net-zero food systems with AI and Microsoft Azure.” Microsoft Customer Stories. (Vendor-published.)
- Mondra. “Mondra launches AI assistant Sherpa.” Press release, 23 October 2024.
- MIT Project NANDA. The GenAI Divide: State of AI in Business 2025. Reported in Fortune, August 2025. (Methodology criticised for sample size and measurement window; treat as directional.)
