Slide 01 · Title 0:00 – 0:10
Do not read the slide. Stand still. Let the wordmark sit for two seconds before you speak.
"We're CoWork. In seven minutes I want to convince you that the thing standing between enterprises and AI agents is not intelligence. It's an org chart."
Slide 02 · The market today 0:10 – 0:35
This slide sets the frame for everything after it. Say the headline, pause, then sweep a hand across the four boxes. Do not read them.
"There's a new AI tool for building software every week. Coding assistants. Autonomous agents. Cloud workspaces to run them on. Every one of them ships the worker: something that can do the job."
"Not one of them ships the workplace. Who is allowed to act. Who approved it. What is on the record. That gap is why we're here."
Slide 03 · What changed 0:35 – 1:00
"And those workers are genuinely good now. So every large company is about to point hundreds of them at its own systems, and not one of those companies has an org chart for them. No job titles. No manager. Nobody's signature on file."
Beat. Then the caveat, said out loud — it buys you credibility for every number that follows.
"That number on the right is a Gartner forecast, not measured spend. I'll flag every number I use that way."
Slide 04 · The problem 1:00 – 1:35
This is the spine of the whole pitch. Slow down. Land the three causes one at a time.
"Gartner forecasts that more than forty percent of agentic AI projects get cancelled by 2027. The interesting part isn't the number, it's the reasons they give. Cost. Unclear value. Inadequate risk controls. Read that list again — not one of them is a model problem. Every one of them is an organisational problem. Better models will not fix a single line of that list."
"And again: that's a forecast, not an observed rate. It tells us the gap is real. It does not tell us we close it."
Slide 05 · What that looks like 1:35 – 2:00
"Here's what an organisational problem looks like in practice. A security flaw is found in a login library that sixty services share. Someone tries an automated find-and-replace, and it fails on the first service that was built its own way. So it becomes sixty tickets in sixty backlogs, and six months later it's still open in fifteen of them."
"The work was never writing the fix. The work is everything around the fix."
Slide 06 · Why nobody has fixed it 2:00 – 2:30
This is the slide the whole pitch exists for. Ask the question first, then stop moving. Let the room sit in silence for a full second after “the organisation.”
"So why has nobody fixed that?"
"Everyone says AI is eating SaaS, and they mean the code. I think that's the wrong read. Enterprise SaaS was never valuable because it was software. It was valuable because it encoded organisational problem-solving — roles, approvals, compliance, coordination. The software was just how you delivered those."
"Agents eat the delivery vehicle. They don't eat the organisation. So that layer is now sitting there with nothing to run on. Agents need a workplace, not just a workbench. CoWork is the workplace."
Slide 07 · The loop 2:30 – 2:50
"Four steps. Send one change to every service that needs it, at once. Every agent works in isolation and has to pass that service's own checks. Where an agent hits something only a person may decide, it stops and asks. And everything is signed with that person's name."
"Let me show you what that looks like."
Slide 08 · Live demonstration 2:50 – 5:00
Show this slide for about ten seconds as the map, then switch to the live demo. Everything below is said over the demo, not over the slide.
"This is one security upgrade, sent to sixty services at once."
Switch to the demo. Leave the full grid on screen.
"Forty-five finished on their own. Fifteen stopped. And the fifteen that stopped are the whole product, because an agent that guesses its way through those is exactly what your risk team is afraid of."
Click worker 18, billing-core. Wait for the drawer to open.
"So we open one. Not fifteen, one. Billing. The old account format doesn't map cleanly onto the new one, and choosing that mapping is a business decision. The agent could have written the fix. It wasn't allowed to choose it."
Click Inject Shim Diff, then Propagate. The drawer closes itself. Step back and let the grid turn green before you speak.
"Now watch one decision reach the other fourteen. It isn't copy-paste. It only applies where the same problem appears, and each service re-runs its own checks before it turns green. Anything that fails stays stopped. Sixty of sixty. One human decision."
Click Consent Gate Ready in the bar at the bottom. Scroll inside the card to the tamper button before you start talking.
"And this is what the risk team gets. Every service, every change, the approver's name. And zero attempts to send anything outside, under a deny-everything policy. That zero is a signed count anyone can re-check, not a promise."
Do NOT say “mathematical proof that nothing left the network.” A technical judge will ask how you prove a negative. Then click Simulate Tampering.
"Change one character. Refused. Anything can say valid. The value is that it can say invalid."
Switch back to the deck.
Slide 09 · Build with AI 5:00 – 5:25
Bridge straight from the demo — you have just shown both halves. “Build with AI” is a scored criterion, and most teams answer it by naming the model they called. Answer it structurally instead.
"What you just watched is two different kinds of machine. The agents did the understanding: sixty systems built differently over years, which no script can handle. That problem only exists because of AI, and only AI solves it."
"And the approval, the names and the receipt are deliberately not AI. An audit that depends on a model being in a good mood is not an audit. We use AI where judgement is needed, and refuse it where proof is needed."
Slide 10 · Who buys it 5:25 – 6:00
The business slide. Two sentences on who, two on price. Point at each buyer as you name them.
"So who pays? Two people have to say yes to a change like the one you just watched. The head of platform, who can order it but can't vouch for it. And the head of risk, who can approve it but has nothing to approve with today. CoWork gives each of them the thing they're missing."
"We charge per service we govern, per month, because that's the unit a platform budget is already written in. Two hundred services is about two hundred and ninety thousand euro a year. We land on one security fix with a deadline, and expand to every organisation-wide change after it."
If pressed on the numbers: it is a proposed structure with no customers behind it, and the slide says so. Defend the unit, not the price.
Slide 11 · Market 6:00 – 6:10 — optional, cut this first if you are running long
"On sizing — I could quote you a five-billion-dollar market report. Instead here's arithmetic you can argue with. Roughly twenty-five thousand organisations worldwide run a hundred or more services. At a hundred and fifty services average, that's about two hundred and sixteen thousand a year each. Regulated sectors are about a fifth of that and they're where the compliance trigger already exists."
"Every one of those inputs is our assumption, and they're on the slide so you can take one apart."
Slide 12 · Team 6:10 – 6:30
Say the names once, then get straight to the credential. Do not walk the panel through five biographies.
"Five of us. The thing worth knowing is that every one of us has shipped production software in banking, financial services and insurance. That is not a coincidence with the pitch. It is why we picked this problem. Between us we have also worked in health and medical, automotive, and consumer internet at scale."
"So we are not guessing why a security fix sits in a backlog for six months, or what a change-approval board wants before it signs. We have been on both sides of that table."
Then close the loop on your own honesty rule. This is what makes the credential land rather than sound like a boast.
"And the limit, same as everything else I have shown you: having worked in the sector is not the same as having validated this with the people who would buy it. That's the question I want to close on."
Slide 13 · Close 6:30 – 6:50 — leaves ten seconds of slack
Do not thank them and trail off. End on the question, then stop talking. Silence is fine — it hands the room to Q&A on your terms.
"Enterprises will not deploy agents they cannot supervise. CoWork is how one person says yes to sixty at once and can still prove afterwards exactly what they said yes to."
"One ask. We couldn't get in front of a real buyer before today, so the most valuable thing we can walk out with is your answer to one question: would you approve a real change across your organisation on the receipt you just saw?"
Appendix A · Q&A defence
Three more, kept off the slide because they are rarer:
"Who is liable if it breaks production?" — The named approver in the receipt, which is deliberate. The receipt exists so that liability has somewhere to land instead of evaporating into "the AI did it."
"Why would a CISO ever allow this?" — They already allow it, badly. Engineers are running agents on production code today with no gate and no record. We are not introducing the risk, we are introducing the control. That reframe matters: the counterfactual is not "no agents", it is "ungoverned agents."
"What is your unfair advantage?" — Honest answer: none yet, we are three weeks old. The structural answer: consent, identity and signature are awkward to bolt on afterwards, so whoever builds the governance substrate first has a real head start over anyone who built the worker first.
Appendix B · The governance layer
Use if a judge asks what the product actually consists of, underneath the demo.
Read the four left-hand labels as questions, not features. They are the four questions a compliance officer asks.
"Four questions decide whether an agent change is shippable. Who may act — the agent gets credentials scoped to the task, so it physically cannot reach the database or production keys. Who approved — a policy gate with a named human at the boundary. What happened — a signed record you can replay, not a log you hope is complete. And what does it look like — because you can't approve what you can't see."
"Compute and storage under that are commodity. This layer is the part nobody can bolt on afterwards."
Appendix C · Propagating safely
Backup still for the cascade, and the slide to go to if a judge asks “isn’t that just copy-paste?”
The drawer now closes itself on propagate and scrolls the matrix into view, so just click and step back. The sweep to green runs for about a second and a half — let them watch it before you speak.
The silent objection in the room right now is "so it's a for-loop." Answer it while the cascade is running, not in Q&A.
"Now watch what one decision does. And I want to be precise about this, because it looks like copy-paste and it isn't. What propagates is a rule with a match condition, not a patch. It fires only where that exact failure signature appears, inside a bounded set of paths."
"And every one of those fourteen runs its own test suite in its own sandbox before it turns green. It earns the green, it doesn't inherit it. Anything that fails stays blocked. Bad code cannot ride the cascade."
"Sixty of sixty. One human decision. That's the leverage."
Appendix D · The receipt
Backup still for the receipt, and the slide to go to if a judge asks how the signature works.
Open the consent gate. Scroll to the tamper button before you start talking so you are not hunting for it mid-sentence.
"Now the other side of the building. The compliance lead gets this. Sixty repositories, sixty pre-verified pull requests, zero egress calls under a deny-all policy, and a named human's identity bound into the signature."
Say this exactly. Do NOT say "mathematical proof nothing left the network" — a technical judge will ask how you prove a negative and you'll lose the room.
"And be careful how you read that zero. We are not claiming proof that nothing left the network. We're claiming the policy was deny-all, the receipt counts the attempts, the count is zero, and it's signed — so anyone can re-derive it."
Click Simulate Tampering.
"Flip one byte. Refused. And that's the point — the value isn't that it can say valid. Anything can say valid. The value is that it can say invalid."
Appendix E · What we built, and what we did not
Go here the moment a judge asks “is the demo real?” or “what have you actually built?” Answer from the slide, without apologising.
Most teams hide this slide. Put it on stage deliberately and make eye contact. It reframes every claim you already made as trustworthy. Do not apologise while saying it.
"Let me be exact about what's real. The coordination engine, the blast-radius fence and the signing path are built and verified — the tamper detection you just saw is real code, not a staged animation. The cloud mounts, live inference and Slack sign-off are modelled deterministically, which is also why the demo can't go down."
"And what we are not claiming: live model accuracy, generality across real external repos, and most importantly, that enterprises will actually accept this receipt. That's the assumption that kills us if it's wrong, and we couldn't test it before walking into this room. A system that sells auditability should be audited about itself first."
Appendix F · Evidence
If a judge challenges a number, do not defend it. Turn to this slide and read them your own caveat before they finish the question. It converts an attack into a credit.
Appendix G · Scenarios
Use scenario 3 only if a judge asks "what's next" or "how big can this get". It is the strongest expansion story you have — it makes CoWork the adjudication layer, not just the execution layer — but it is not built, and saying so is mandatory.