Where Model ML actually is
Read the position before proposing the move.
Model ML has product-market fit and a distribution problem it has not had to solve yet. Those numbers are inputs. Nobody has turned them into an operating system.
The wedge is real and it is unusual. Rogo and Hebbia answer questions. Model ML produces the deliverable - client-ready Word, PowerPoint and Excel in the firm's exact prior format - and since Captide it can cite the filing behind every number. In a market where the loudest recurring complaint about competitors is output quality and the need to manually verify answers, owning the finished artefact plus its provenance is the sharpest position on the board.
The proof points are enterprise-grade. Two of the Big Four deployed. Three Hills Capital automating monthly portfolio reporting and first-draft investment memos. ISO 27001, SOC 2 Type II, GDPR. A Series A led by FT Partners, which is itself an investment bank - the investor base and the customer base are the same people.
What is not yet built is the machine around it. There is no codified sales or engagement playbook. Expansion is not instrumented. Pricing is seat-shaped in a market moving to outcomes. The engagement team is a forward-deployed motion carrying the gross-margin cost of one, without the productisation that eventually pays it back. Those four gaps are the Strategy and Operations job, and they are the difference between a company that doubles ARR once and one that compounds.
The threat I'd put on the wall
Any strategy that does not name this is not a strategy.
Microsoft has stopped being horizontal
On 25 June 2026 Microsoft published "Copilot in Excel: built for the era of Frontier Finance." This is not generic document generation any more. Microsoft is shipping finance-specific skills - building a DCF, closing the books, refreshing monthly reporting models, preparing variance analyses - deployable through Microsoft Marketplace. Those are Model ML's workflows, named individually, by the owner of the file formats.
Two details make it sharper. First, Rogo is named in Microsoft's first wave of finance partners, alongside LSEG, Ramp, Samaya AI, Velixo and Vena, with partner-built skills landing in Q3 2026 - now. A direct competitor is inside the ecosystem, and it is better capitalised: $160M Series D in April 2026, $300M+ total, roughly $2B valuation, 35,000 professionals at 250+ institutions.
Second, Microsoft has signed the data layer: FactSet, PitchBook, Morningstar, S&P Global / Kensho, Daloopa, CB Insights, LSEG and Moody's are all Excel connectors now. The assumption that a finance AI platform wins by aggregating data sources is closing.
Model ML's answer is narrower than it was, and stronger for it. Copilot can now reach the data and run the calculation. What it cannot do is reproduce a specific firm's prior deck and memo format exactly, carry a citation trail back to the filing behind every number the way Captide enables, or put a forward-deployed engagement pod inside the client to make it stick. The defensible ground is provenance and the finished artefact, not data access. My first strategic act would be to make that distinction measurable in every bake-off, not arguable in a deck.
The strategic corollary: Microsoft is a partner-or-compete decision, not a passive risk, and the window is narrowing rather than theoretical. Rogo chose partner. Model ML should decide deliberately, before the Excel surface becomes the assumed starting point for every bank's AI committee.
Seven levers
The full architecture. I would own two on day one and sequence the rest.
Productise the engagement motion
The engagement team is a forward-deployed model, and forward-deployed models trade margin for moat. ServiceNow was at 63% gross margin at IPO and 79% by 2024; Workday 54% to 75%. The climb comes from productising the repeatable parts of each deployment.
I would decompose the last ten rollouts into what was genuinely bespoke and what was repeated, then turn the repeated 60% into templated modules, reusable Grids, and a staged rollout kit. Same white-glove feel, materially less bespoke build.
Why now: there is no playbook today. Pod ten should not relearn what pod one already paid for.
Instrument land-and-expand
47,000 seats is a headline, not a metric. The questions that matter are how many are activated, which workflows drive week-four retention, which accounts are expansion-ready, and which are quietly idle before the renewal conversation.
Enterprise SaaS above $100K ACV runs a median 118% net revenue retention; 120%+ is the extreme-PMF bar. I would build the expansion dashboard that makes NRR, seat activation and workflow depth visible per account, and route the signals to the engagement pods weekly.
Precedent I have run: the AI screening engine at Farnam, and the internal data layer that partners now run deals on.
Move pricing from seats to outcomes
Seat pricing caps you at headcount and invites procurement to cut idle licences - sizing seats to real weekly use typically removes more than half of them. Outcome pricing (per deck, per memo, per diligence pack) grows with value delivered and competes for the labour budget rather than the IT budget.
Model: Snowflake's consumption model produced 158% NRR, the highest of any cloud company at listing. I would build the pricing model and the migration path, not just the argument.
Turn the Big Four into channel
Two of the Big Four are already customers. AlphaSense converted exactly this position into distribution: Accenture Ventures invested and now embeds AlphaSense across its client base of roughly 9,000 enterprises. Systems integrators partner because they cannot build the platform themselves.
The play: a co-delivery motion where the consultancy sells transformation and Model ML is the engine inside it. One partner agreement can outperform a year of direct outbound.
Make provenance the product
Racing Microsoft to sign data providers is now a losing lever - FactSet, PitchBook, Morningstar, S&P Global, Moody's and LSEG are already Excel connectors. Access is commoditising.
What Captide actually bought is harder to copy: a citation trail from any number back to the filing it came from. In a market where the standing complaint about AI tools is that analysts still have to verify every output by hand, and where regulators are moving on AI explainability, provenance is the feature that survives.
The play: productise citability as a headline capability with its own metric - percentage of generated figures traceable to source - rather than leaving it as an architecture detail.
Continue the roll-up, with diligence
Flippr brought PowerPoint generation, Captide brought the data layer, and Chaz has said publicly he is looking at more agent companies. That is a deliberate build-versus-buy engine, and it needs a repeatable evaluation framework rather than opportunism.
Direct fit: this is my day job. Screening, LBO and valuation work, IC-ready materials, and post-deal integration planning across £50M to £1.5B EV targets.
The Gulf, before someone else takes it
Model ML is hiring across London, New York, Hong Kong, Singapore, India and now Madrid and Paris. There is no Gulf presence, while ADIA, Mubadala, PIF and QIA sit on an enormous share of global AUM, run exactly the diligence-and-memo workflows Model ML automates, and buy technology rather than build it. Rogo is pushing EMEA and Asia; the Gulf is currently uncontested by both.
Why I can help: I grew up in Bahrain, my family is in Saudi, and I have the regional network alongside the PE credibility to walk into a sovereign fund and be taken seriously. I would scope it as a thesis with named targets before anyone commits headcount.
Playbooks worth stealing
Four companies solved a version of Model ML's problem. Each hands over one specific mechanic.
Palantir's forward-deployed engineers write production code inside the customer and own the data-to-decision loop. Pairing that with AIP turned a government contractor into a commercial force, and it is now the model Anthropic and OpenAI have both copied - Anthropic's first hundred enterprise contracts were closed by FDEs, not account executives.
Applied here: Model ML's engagement team is already this motion, staffed with finance people rather than engineers, which is arguably better for the buyer. The missing half is Palantir's discipline about converting each deployment into reusable platform capability.
Consumption pricing meant customer spend grew automatically as usage grew, producing 158% net revenue retention - expansion without a sales conversation.
Applied here: every deck, memo and diligence pack Model ML generates is a natural billing unit. The product already counts them. Pricing simply has not caught up to the telemetry.
Bloomberg sustains roughly $30K per user per year because the moat is not the data. It is data gravity plus a network layer - Instant Bloomberg became the default way the market talks to itself, so the switching cost exceeds the subscription cost.
Applied here: the agent harness is the analogue. Skills, memories, context and templates accumulate per firm, and every month of use makes leaving more expensive. A shared module and template library across customers is the closest thing to a network effect in this category, and I would build the strategy for it deliberately rather than letting it emerge.
$600M+ ARR, $7.5B valuation, 90% of the S&P 100. Then it took strategic investment from Accenture Ventures and turned a consultancy into a distribution arm across roughly 9,000 enterprise clients.
Applied here: Model ML has Big Four firms as customers already. Converting one of them from buyer to channel is the single highest-leverage distribution move available, and it costs headcount rather than capital.
The first 90 days
Concrete, sequenced, and deliberately narrow. Two levers shipped properly beats seven started.
- Ride along on ten engagements. Sit with the pods across investment banking, private equity and consulting accounts. Document what actually happens between contract signature and the first deliverable a client trusts.
- Decompose the last ten rollouts into bespoke versus repeated work. That ratio is the single most important number nobody has calculated yet.
- Map the data. What does the product already emit - seats, sessions, workflows run, documents generated - and what is missing to compute activation and NRR per account.
- Interview the two Big Four accounts as partners, not customers. Test appetite for co-delivery before building a case for it.
- Ship v1 of the expansion dashboard. Seat activation, workflow depth, week-four retention and expansion-readiness per account, reviewed weekly with the engagement and GTM leads. Not a report - an operating rhythm.
- Ship v1 of the engagement playbook. The champion-identification stage Milan described, the compliance and security gate, the staged rollout, the first-value milestone, with templated Grids and modules for the repeated 60%.
- Build the compliance asset. Package ISO 27001, SOC 2 Type II, GDPR and the Captide citation story into a reusable diligence kit. Enterprise AI procurement runs four to nine months; shortening it is revenue, not admin.
- Model outcome-based pricing against the current book. What does per-deliverable pricing do to the top twenty accounts, and where does it break.
- Run the playbook on two live deployments and measure time-to-first-value against the pre-playbook baseline. If it does not move, the playbook is wrong and I will say so.
- Publish the competitive answer to Copilot's finance skills. A structured, evidence-backed comparison the whole GTM team uses - exact-format fidelity, citation trails, deployment depth - built from real side-by-side outputs against Copilot in Excel, not claims.
- Bring one channel proposal to the table. A Big Four co-delivery structure with named sponsors, commercial shape and a pilot account.
- Scope the Gulf thesis. Named sovereign and regional targets, the workflows they run, entry sequence and what it would cost to test. A decision memo, not a hiring request.
The metrics I'd own
A strategy you cannot measure is a memo. These are the numbers I would put my name against.
| Metric | Why it is the right one | Benchmark |
|---|---|---|
| Net revenue retention | The single truest measure of whether land-and-expand is working | 118% median · 120%+ elite |
| Seat activation rate | 47K seats only count if they are used; idle seats are next year's churn | >50% typically idle |
| Time to first trusted deliverable | The moment a client stops checking the output is the moment they are retained | baseline, then reduce |
| Bespoke vs repeatable build ratio | The margin story; it is how forward-deployed models eventually pay back | 63% → 79% (ServiceNow) |
| Procurement cycle length | Four to nine months of enterprise diligence is compressible with a reusable kit | 4 to 9 months |
| Gross dollar retention | Mission-critical status shows up here before it shows up anywhere else | 95% to 97% |
Every benchmark above is drawn from 2026 enterprise-SaaS and vertical-AI data rather than invented, and I would expect to revise them against Model ML's own book in the first month.
Why me for this seat
I am the customer, the builder, and the operator - which is an unusual combination for this role.
I am the buyer. I run buy-side and sell-side diligence on £50M to £1.5B EV targets. I write the IC memos and build the LBO models Model ML generates. I know precisely which outputs a partner will trust and which ones get quietly rebuilt in Excel at 11pm.
I have built this product, badly, by hand. At Farnam I built an AI-scored screening engine that expanded a fund's qualified pipeline 4× to 1,200+ targets, and Companies House pipelines across 3,000+ managed-service providers. I learned which objections are real - trust, verification, the one wrong score that discredits the system for a month - and which are theatre. That is also why I understand the agent harness and Grids as architecture rather than as marketing.
I have built operating layers before. A 100-day value-creation plan with owners, KPI targets and weekly cadence at Nootra. Operating-model design and KPI governance across a $100M healthcare group. Founded and exited a SaaS company. This is what Strategy and Operations is: the internal machine that lets everyone else move faster.
Chaz's stated ambition is to replace the Office suite for financial services. I think that is winnable, and I think it is won on operating architecture rather than model quality - because the models are commoditising and the workflow is not. I would like to build that architecture.