Before You Sign an AI Partnership, Demand This Scorecard
A strategic AI partnership is not a value-creation plan.
It can remove procurement friction. It can give portfolio companies structured access to cloud infrastructure, models, security tooling and implementation support. It can also create a strong announcement for the fund and the vendor.
None of that proves deployment.
A fund earns alpha when a named management owner changes a production workflow, users adopt it, finance validates the bridge to EBITDA and the company can still exit the vendor relationship without rebuilding its operating system.
That distinction matters now. In August, Clearlake Capital and Google Cloud announced a strategic partnership spanning infrastructure, data, cybersecurity, Gemini Enterprise, Vertex AI, open and third-party models, and hands-on expertise. The strategic logic is clear: aggregate access and delivery capacity at fund level, then make it available across the portfolio.
But the public announcement describes access and intent. It doesn't publish company-level deployment targets, committed implementation capacity, adoption gates, an EBITDA attribution method, portability tests or renewal and exit rules. Those terms may exist privately. The point is simple: before a fund signs, the operating case needs to be as explicit as the commercial case.
I've spent more than 20 years building managed infrastructure businesses. At €240 million ARR, through 15-plus acquisitions and toward a €1.5 billion exit, I learned that vendor access is the easy line in the board deck. Repeatable deployment across different operating companies is the work.
Here's what works: demand a Portfolio Partnership Scorecard before the signature.
Access, deployment and value are different assets
The PE value-creation bar has moved. Deloitte's August 2026 analysis argues that familiar levers such as roll-ups and margin improvement have become table stakes. Its prescription is execution first: plan for exit from day one, monitor assets continuously and intervene when the operating case moves off plan. That is a current point of view, not a law of nature, but it gets the operating discipline right. Read the Deloitte analysis.
Current deployment evidence also shows why entitlement shouldn't be confused with scale. FTI Consulting's 2026 Private Equity AI Radar surveyed 200 senior PE decision-makers at firms with at least $1 billion in AUM. Thirty-six percent reported portfolio AI in production at use-case, functional or enterprise level. Only 7% reported enterprise-level production.
That 7% is a self-reported maturity snapshot, not a 93% failure rate. Still, it exposes the gap between having AI initiatives and running AI as a repeatable portfolio capability.
An investment committee should separate three assets:
- Entitlement: the products, credits, support, regions and commercial terms the agreement makes available.
- Deployment: the production workflows integrated with a portfolio company's systems, controls and operating routines.
- Value realization: the adopted change that moves revenue, margin, working capital, risk or exit readiness, validated by finance.
A full-stack agreement may strengthen entitlement while deployment remains blocked by data boundaries, missing owners, weak integrations or no implementation team. Consumption may rise even when value realization does not.
That's why the scorecard must measure operating conditions, not vendor breadth.
The Portfolio Partnership Scorecard
Score each field from zero to two:
- 0 — absent: no named answer and no evidence.
- 1 — promised: a narrative commitment, roadmap or vendor-owned proof.
- 2 — executable: a named owner, dated deliverable, acceptance test and evidence source; contractual where the risk requires it.
This is a PromptPartner operating framework, not an industry benchmark or a magic predictive score. Its job is to make missing conditions visible before the fund concentrates spend, data and delivery around one provider.
Do not hide a red field inside a green total. A zero in the data boundary, management owner, EBITDA bridge, portability or renewal/exit gate should block a portfolio-wide commitment. It may still justify a ring-fenced proof.
1. Eligible workloads
Name the workflows and systems in scope. Define data class, risk tier, baseline KPI and acceptance test. “AI across the portfolio” is not a workload register. Start with jobs such as quote generation, service-ticket triage, contract review or pricing analysis where the operating event and user are observable.
2. Portfolio coverage
List which companies can use the agreement now and what makes each one deployable. Sequence them by value, data readiness, integration effort and management capacity. A portfolio-wide logo with no PortCo queue is coverage theatre.
3. Data and residency boundary
Map what data may enter which service, model, region, support path and logs. Record subprocessors, retention and deletion, training-use terms, encryption, key ownership and the incident route. “Enterprise-grade security” is a label. A reviewed data-flow map is evidence.
4. Co-investment
State who pays for discovery, migration, integration, cloud consumption, change management and overruns. Credits have expiry dates and conditions. They don't write integrations or change frontline behavior. Put cash, services, credit value and the overrun owner in one schedule.
5. Implementation capacity
How many squads can start? Which skills are committed? What must the PortCo provide? Demand named fund, vendor and delivery-partner staff, committed hours, backlog capacity, decision rights and escalation. “Hands-on support” without people and dates is still marketing.
6. First-value lead time
Define the clock from approved access to a user completing a live workflow with measurable output. Record prerequisites and stop-clock rules. A project that reports a two-week build after three months of data and identity work did not reach first value in two weeks.
7. Management owner
Name the PortCo executive who owns the result and can change the workflow. Add technical, security, finance and frontline owners. Fund-level sponsorship can open the door; only local authority can redesign the work.
8. Adoption evidence
Measure eligible users, active completion, rework, override, cohort retention and support burden. Licences, logins, prompts and demos are not adoption. The useful question is whether the target users repeatedly complete the intended production job.
9. Realized EBITDA bridge
Require a CFO-approved baseline and a monthly ledger. Separate observed value, annualized run rate and forecast. Include one-time implementation cost, ongoing cloud and model cost, support cost, displacement and leakage. Hours released are not cash removed unless management changes the cost base or turns capacity into additional profitable output.
10. Vendor concentration
Map where data, identity, models, workflow logic and operating skills accumulate. Set outage and substitution plans, commercial exposure and a second-source decision. A full stack may reduce integration friction while increasing concentration. Both can be true.
11. Portability
Test whether the company can export data, prompts, evaluations, embeddings, workflows, logs and model interfaces in usable formats. A contractual right to export is not a technical exit. Run an export-and-restore or provider-substitution test and record the time, cost and missing artifacts.
12. Renewal and exit gate
Define the evidence required to expand, renew, remediate or exit. Include the review date, minimum evidence pack, pricing reset, termination assistance, deletion certificate, transition period and approver. Consumption is not a renewal case.
The scorecard should be a heat map by company and workload. One relationship may be strong on platform scope but weak on delivery capacity. Another may be expensive but highly portable. The pattern matters more than the total.
Make the EBITDA bridge hard to game
The bridge should be boring enough for finance to own:
Realized monthly EBITDA = validated revenue contribution + removed cash operating cost − incremental run cost − leakage or displacement
Every component needs a baseline period, source system, finance owner, realization date and confidence note.
Keep pipeline influenced separate from recognized gross profit. Keep capacity released separate from cash cost removed. Show the cost of cloud, models, integration, human review, support and governance below the benefit line. Report one-time build cost separately so the investment committee can see payback without pretending forecast value is already realized.
The same rule applied across the acquisitions I worked on: if a synergy bridge had no baseline, owner, date and finance sign-off, it wasn't a bridge. It was a story. AI doesn't get a special exemption because the demo looks impressive.
Bain's 2026 private equity midyear report reaches a compatible conclusion from a wider market view: constrained resources and longer holding periods increase the premium on repeatable underwriting and value-creation systems. The fund needs a method it can run across companies, not isolated proofs owned by enthusiastic individuals.
Thirty days to proof
Thirty days will not prove portfolio-wide EBITDA. It can prove whether the partnership has a working deployment mechanism, whether first operational value can be observed and whether the evidence deserves more capital.
Days 1–3: Select and baseline
Choose two portfolio companies with different systems or data constraints. Select one narrow, high-frequency workflow per company. Freeze baseline cycle time, completion volume, error and rework, unit economics, current run cost and user pain.
Name the P&L owner, technical owner, security or privacy reviewer, frontline process owner and finance validator. Agree the success, stop and safety criteria before anyone builds.
Days 4–7: Contract the boundary
Complete the data-flow and residency map. Approve the model and service list, retention, access controls, logging and incident route. Confirm who funds integration and consumption during proof and scale.
Commit named implementation staff. Unblock environments, APIs, identity and test data. Design the export route now, before dependency forms.
Days 8–14: Build the thin production slice
Connect the smallest end-to-end path that completes a real workflow inside the systems of record. Add human approval where consequence demands it, plus quality tests, fallback behavior, telemetry and cost tags.
Record first-value lead time from the agreed starting event, not from the moment the engineering team finally received clean access.
Days 15–21: Run a controlled live cohort
Put the workflow in the hands of named users. Track eligible users, active completion, overrides, rework, defects, latency, unit cost, support load and incidents.
Hold two short operating reviews with the PortCo owner. Change the workflow, training or controls based on evidence. This is operating work, not a showcase.
Days 22–26: Validate economics and portability
Finance compares the live cohort with the frozen baseline. Label value as observed, annualized run rate or forecast. Run one export-and-restore or provider-substitution test for the critical artifacts.
Estimate the capacity and prerequisites needed for the next five portfolio companies. Don't extrapolate from two nearly identical businesses.
Days 27–30: Run the evidence gate
Bring the completed scorecard, telemetry, exceptions, cost ledger, finance bridge, user evidence, portability result and deployment backlog to the investment committee.
Decide scale, remediate, narrow or stop by workload and company. Convert successful proof conditions into contractual milestones for expansion and renewal.
NIST's voluntary AI Risk Management Framework and Playbook provide a useful cross-check: govern accountability, map context and risk, measure performance, then manage the production system. The Playbook explicitly isn't a universal checklist, and this proof isn't “NIST certified.” Use the functions to test whether the operating loop is complete.
Sign the deployment system, not the access story
A fund-level AI partnership can be a real advantage. It can aggregate purchasing power, standardize parts of the stack, bring scarce expertise into the portfolio and shorten the path to production.
The wider the agreement, though, the easier it is for access, consumption and dependency to grow faster than verified value.
Before the next signature, ask for the scorecard with names, dates, baselines, evidence sources and exit tests. If the answers fit only in a press release, the fund is not underwriting deployment yet.
Here's what works: 30 days to proof, then scale what survives the evidence gate. Build the operating capability, keep the artifacts portable and transfer ownership into the portfolio. Systems win when the companies can run them without the announcement.
