Well-positioned for the agentic AI opportunity
Hg’s CMD strengthened the case that AI is becoming an operating capability across
the portfolio rather than a thematic overlay. The most attractive opportunity is not
generic AI adoption, but the conversion of trusted, workflow-embedded vertical software
into AI-enabled systems of action. As highlighted in our previous update note, we believe that HgT is well-placed to benefit from the AI augmentation of software
products across its portfolio. Hg has developed an efficient engine allowing for end-to-end
AI product rollout within three months. Hg’s agenda, fuelled by more than $80m of
investments from its own balance sheet per year, is underpinned by Hg Catalyst, a
dedicated AI product incubator with more than 100 engineers, designers and product
managers (20 in-house and more than 80 contractors), which is part of a broader team
of more than 150 AI value-creation experts. Hg’s capabilities are aided by its multiple
partnerships with AI providers (eg Anthropic, Cognition, Forethought) in agentic coding
and support, customer success, answer engine optimisation, sales and chat. Hg’s collaboration
with Anthropic (initiated two years ago) now allows any of its portfolio companies
to access Anthropic’s products (Hg was also beta-testing some of them).
We believe Hg’s portfolio companies are well-placed to become AI software winners
in many of their vertical niches. They are trusted providers of mission-critical software
that typically represents a low proportion of customer spend, but is deeply embedded
in customers’ systems and underpinned by proprietary datasets, deterministic workflows
and extensive domain expertise. These features should matter in agentic AI, because
enterprise customers require accuracy, auditability, governance and workflow integration,
not just access to a capable model. This is particularly relevant where AI agents
are deployed in areas such as finance, payroll, compliance, healthcare IT or legal
and regulatory workflows, where error tolerance is low.
We are encouraged that, according to Hg, c 50% of revenues across its portfolio companies,
and 70–85% for its top companies, are already outcome-based rather than derived from
traditional per-seat pricing. This should position the portfolio well for a shift
towards AI-driven pricing models. Outcomes are defined differently by company and
may include, for instance, the number of invoices processed, issues resolved or certifications
confirmed or may be linked to specific workflows that contribute to the final customer
output. In principle, this model can allow customers to achieve cost savings versus
manual processing, while giving Hg’s portfolio companies scope to capture a greater
share of value per customer. Hg aims to reinforce this positioning through product
personalisation at the user and organisational level, combined with a structured approach
to evaluating AI models as they improve over time (see Exhibit 2).
Some customers may still decide to build their own agentic workflows, use system integrators
to build on top of existing systems or connect general-purpose enterprise agents to
their software platforms. Horizontal workflow platforms such as ServiceNow also represent
a potential competitive risk at the orchestration layer, particularly where larger
customers want a single enterprise-wide control plane for AI agents. We believe this
risk is lower, though not absent, for small and medium-sized enterprises, which form
the core customer segment for many Hg portfolio companies and are less likely to build
bespoke AI workflows themselves. Hg’s head of research also noted that there is, so
far, little evidence that AI-native companies are capturing market share in deterministic
workflows. That said, Hg’s companies will need to become the trusted system of action
within their vertical niches before horizontal platforms commoditise parts of the
workflow.
Finally, it is important to stress that AI products developed across Hg’s portfolio
are model-agnostic and are designed to match the model to the complexity of task in
each process, a practice known as ‘model routing’. As a result, these products rely
mostly on open-source models and small language models, which allows users to optimise
the consumption of tokens, that is units of text that the model processes (80% of
which is now attributable to these models across Hg’s portfolio). The CEO of Cognition,
the provider of the coding agent Devin (which Hg also uses), recently highlighted
that for boilerplate work, companies can obtain a five to tenfold improvement in cost
efficiency by using models other than those offered by frontier AI labs. We consider
model routing a prudent cost optimisation approach, especially given that frontier
labs will, at some stage, prioritise generating a satisfactory return on their massive
infrastructure investment. Models from frontier labs are also used across Hg’s portfolio
and are likely to remain relevant as long as they are at the forefront of agentic
AI progress.
