2027 research agenda

Three questions worth following.

The opportunity is a hypothesis. Each theme needs a buyer, a measurable advantage and a condition that would invalidate it.

01

Memory & context

Can memory earn a separate budget?

Explore companies in this theme →

Customer and buyer

Product teams building assistants that work across people, changing facts and permission boundaries.

Why now

Longer-running workflows make stale context and fragmented state visible to users. More stored text is not automatically better context.

Where value may accrue

Temporal updates, deletion, provenance and access-aware retrieval can become durable value when measured against a simpler baseline.

Strongest counterargument

Application state, ordinary retrieval and native model features may solve enough of the task for a lower total cost.

Stage-appropriate evidence

At pre-seed, test a narrow user problem and retrieval correctness. At seed, establish repeated use, deployment effort and willingness to pay.

What invalidates the thesis

A simpler system matches task success and permission safety at lower operating and integration cost.

Related screens

Qontext

02

Reliable deployment

Can evaluation become an operating advantage?

Explore companies in this theme →

Customer and buyer

Engineering and QA teams shipping AI into customer-facing or commercially consequential work.

Why now

Agent behavior spans tools, state and retries. A plausible final answer does not establish that the requested action succeeded.

Where value may accrue

Evaluations tied to human judgments, incident prevention and the cost per successful task can earn an ongoing budget.

Strongest counterargument

Model providers, open-source tools and in-house test suites may absorb the product. Dashboards without changed outcomes are vulnerable.

Stage-appropriate evidence

At pre-seed, compare a defined failure class with blinded human labels. At seed, establish production correlation and an owner who renews the budget.

What invalidates the thesis

The product cannot predict costly failures, or its value is reproducible using the customer’s existing stack.

Related screens

SuperBryn

03

Business workflows

Can software own a painful, repeated job?

Explore companies in this theme →

Customer and buyer

The budget owner responsible for a specific process: banking operations, procurement or customer operations.

Why now

Models can perform more individual tasks; adoption still depends on trust, integration and the economics of the complete process.

Where value may accrue

Domain context, distribution, audit trails and repeated workflow use may support pricing and make replacement costly.

Strongest counterargument

An incumbent can add the feature, or a service-heavy implementation consumes the supposed software margin.

Stage-appropriate evidence

At pre-seed, identify the buyer and observe the current work. At seed, inspect repeated use, payment, exception handling and deployment burden.

What invalidates the thesis

Repeat demand disappears after a pilot, or manual intervention and support erase customer ROI and contribution margin.

Related screens

Neurofin

Framework sources

These investor and builder perspectives inform the questions; the theme conclusions above are analytical hypotheses. Model-provider APIs remain a valid implementation choice. Differentiation, data control, economics and portability need to be tested in the customer’s workflow.

  1. Seedcamp investment FAQ · investor perspective
  2. Bessemer: Vertical AI principles · investor perspective
  3. Anthropic: Demystifying evals for AI agents · builder research
  4. Sequoia: Own Your Intelligence · investor perspective