Adoption is real but narrow: 23% of firms are scaling an agentic system somewhere and no single business function exceeds 10%. The portfolio question is not which use cases exist, but which ones a firm will refuse, and who is allowed to retire the rest.
Why this matters
The tempering number comes first. About 23% of 1,993 organizations report scaling an agentic system somewhere and another 39% are experimenting, which reads like a wave until the ceiling from the same survey lands: in any given business function, no more than 10% are scaling. Most scalers are in one or two functions. Financial institutions sit outside the leading sectors, which are technology, media, telecom and healthcare rather than anyone a fund competes with.
Where agents actually land, in the one survey that asked: data analytics and business intelligence 28%, then software engineering, customer service and IT/cybersecurity at 23% each, with legal, risk and compliance last at 6% (n=623, AWS-sponsored, ~10% financial services). For a fund that ordering is the useful part: demand originates in analysis, not in compliance, so governance arrives after the use case, never before it.
The portfolio failure mode is never deciding, rather than picking wrong. Use cases accumulate bottom-up because each one is individually reasonable, and nobody owns retirement. That is the road to agent sprawl, measured as an ownership ratio rather than a count. The single cheapest portfolio discipline is a named owner with authority to kill things, applied at registry review.
Two selection heuristics worth stealing, both cheap and neither validated:
- Only a process a firm could already have automated is ready to be agentified (a CIO’s rule of thumb). Automated workflow plus clean data predicts readiness; nothing in the sources tests this, and it costs nothing to apply.
- Tolerance for error varies enormously by baseline. Anti-money-laundering screening is the classic first mover because the incumbent process is mostly waste: across 19 institutions in 2017, roughly 16 million alerts produced just over 640,000 filings, about 4% converting. An agent that is wrong a third of the time can still beat that. Contrast a process where the baseline is right 99% of the time and the failure is a wire transfer.
Two framings will be sold to firms, and both deserve resistance. The first is the modelled uplift ladder: 5–10% from assistance, 20–40% from agents inside the existing workflow, 60–90% from a reinvented process. It comes from a consultancy’s hypothetical call centre, labelled “estimated impact,” with no fund analogue anywhere in the report. The second is the use-case menu, which lists agents of wildly different risk in one column: every such menu mixes tier 1 and tier 3, and tiering is the first thing applied to it before deciding anything.
The honest counterweight to portfolio discipline comes from the citizen-development literature: accept a failure rate among citizen-built tools in exchange for volume IT could never supply. That is right, and it is the argument for keeping the bottom tier light. It stops being right the moment an artifact acquires production reach, the drift described in the problem.
For value claims, note what the base rates look like when someone finally measures. Firms overwhelmingly report using text generation (41%) over anything agentic, and among the pre-agent citizen-automation programs with published numbers, the currency was hours saved, which the CFO objection disposes of in four words: we cannot eat hours.
Where you stand
| Level | Looks like | Cheapest next move |
|---|---|---|
| Crawl | Use cases arrive as individual requests or appear unannounced. No list, no owner, no retirement. | List what exists and who depends on it. That list is the portfolio, and most firms have never seen theirs. |
| Walk | A list exists. Everything on it is “active.” Nothing has ever been retired. | Retire one thing deliberately, and record why — the muscle matters more than the item. |
| Run | Use cases enter through tiering, carry an owner and a metric, and are reviewed on a cadence with a kill option. | Rank by criticality against usage; the top-right quadrant is where controls go first. |
| Fly | Portfolio managed against a stated appetite, with retirement rates reported alongside launches and pilots designed to be killable. | Report the retirement rate to the operating committee next to the launch count. |
Concerns this dimension covers
- Agent sprawl — the portfolio failure mode: use cases with no owner, inventory, or retirement path.
- Quality debt and orphaned apps — what an unmanaged portfolio accrues while everyone is launching.
- Resource overload and runaway cost — the portfolio’s cost side, which arrives faster than its benefit side.
Controls that answer them
- Agent inventory and registry — a portfolio is unmanageable until it is enumerable.
- Cost controls — per-use-case metering is what makes retirement arguable rather than political.
- Grounded in: risk tiers and trust zones — criticality tiering is how a portfolio gets triaged.
Who sells it
- Domain-specific finance AI platforms — the buy option that removes whole use cases from the firm’s citizen portfolio, and the vendor-viability question that comes with it.
- Enterprise AI assistant platforms — where most of the analytics and drafting demand can be satisfied without a build.
Sequencing and where this is checked
- 90 — the canonical playbook, including how to design a pilot the firm is willing to kill.
- Day 3 sequencing — where pilots sit in the larger build order.
- Registry review cadence — the recurring portfolio triage.
- Risk-tier assignment — every new use case enters through it.
Open questions
- What separates a lighthouse pilot from a demo that never scales? The consultancy answer is process redesign; the only causal evidence nearby is a null result on task composition from individually-provisioned AI. Nobody has measured it for agents.
- There is no published portfolio benchmark for firms of this size: no agents-per-employee, no retirement rate, nothing. A benchmark built internally is more useful than anything cited here.