Welcome to The Folio Dispatch again – the newsletter that charts our progress towards solving the professional services productivity problem...and also provides interesting reading on the market that we're operating in.

In this edition we will cover the usual updates on business and product building, share news on what's happening in the AI & SaaS market, and in Founder Notes our CEO shares some insights on disruption in the professional services industry.

Happy October everyone...I can't believe how quickly this year is flying in! I guess that's what you get running a startup, busy times are brief times. However, I have managed to pack a lot into the last month or so, and you can read about some of that in the sections below. But what I want to focus on in this section is a topic that I have been discussing a lot recently: the state of the professional services market.

Professional services are facing a really tough time at the minute. The economy over the last 3 years has created some clear headwinds:

1. Less work, more competition. As cost reduction programmes have rolled out across the world's brands, in response to post-COVID over hiring, spiraling inflation, and tariff uncertainty, clients are spending less. This means fierce competition to win more limited business and huge pressure on rates.
2. No more easy money. Tougher spend controls mean that every external spend needs solid investment case. Brands can no longer get away with using externals to solve every problem. In addition, the post-COVID hiring-and-firing bonanza means that in-house teams are stronger than ever, and have less need to use externals. This means that once projects are completed, you no longer walk straight into an extension, and your sales costs rise dramatically as a result.
3. New competition. The many experienced people who have struggled to get rehired after layoffs means that there is a glut of qualified contractors and micro-business that are stealing business away from professional services firms – especially via contractor marketplaces. Conversely, PE-driven consolidation of complementary businesses is changing the competitive landscape.
4. New approaches. This one is all about AI. Professional services firms are now competing against ChatGPT for basic tasks and specialised SaaS solutions for deeper work. In addition, there is a huge cost and quality gap between firms that are using AI effectively (to support their service delivery) and those that don’t (who now appear antiquated, slow, and expensive as a result).

So these are tough times, with firms popping up and going bust on a regular basis. And embracing AI fully is going to be critical. However, as we can see in the news, even the biggest professional services firms are struggling with how best to deploy AI at scale across the organisation.

We’re hoping Folio will be at the heart of helping agencies shape up!

- Niall.

Investor Portal

We've knocked up a little microsite to share all things Folio with our interested stakeholders. In line with our radical transparency principle, we're uploading more and more to this portal, and also allowing people to ask us questions, suggest new documents to share, and track progress against our growth milestones.

Nb. it's probably not the best name for this little site tbf, as it's not just for investors - it's for everyone!

AI-CON

At the end of September we hit up AICON at the Titanic, with day two at W5. This was a fun and interesting event, and a great opportunity to reconnect with friends, ex-colleagues, and industry thinkers about everything AI.

SaaStock

Hot on the feels of AICON, we headed down to the RDS in Dublin for SaaStock. This is a brilliant event - small enough to let you really connect with the community but big enough to draw big names. We loved catching up with Kieran Flanagan and Manny Medina to talk about everything AI, growth, and funding. We crammed in a few VC meetings also - with positive chats in particular with Nauta and Secways.

Up next: Websummit

The big one is coming up soon, and we head off to Lisbon on 10th November. We have a packed agenda of meetings - mainly stirring up interest in the big professional services firms that will be attending in the hope of finding innovative new solutions that can give them an edge.  We'll be happy to oblige!

Everyone's launching "conversation-first" AI platforms these days. But when you actually open them, you're staring at the same carousels, dropdowns, and endless scrolls that ruled web design in 2015—just with a chat bubble bolted on. Groundbreaking, right?

Here's the uncomfortable truth: most companies building for agentic AI aren't actually rethinking interaction design. They're repurposing it. They're taking a 30-year-old GUI playbook and gluing a chatbot on the side like that's innovation.

Why this matters

Designing for agents isn't the same as designing for apps. Agentic software inverts everything we know about traditional interfaces. The UI becomes a bridge (think Slack or iMessage), with intelligence behind the scenes. You no longer "use" the software—it acts on your behalf. This is a fundamentally different design problem.

The patterns that actually work aren't coming from UI frameworks. They're coming from conversation design and how humans actually collaborate. The best agentic systems emerging today do something radically different: they work with you, not for you.

What real patterns look like

Instead of approval gates, you get collaboration. Before doing something big, the agent asks. It gathers findings, shows you what it discovered, and lets you guide what happens next. You see the evidence, the observations, the reasoning—not a black-box answer. Work keeps happening in the background while you provide feedback. Over time, the agent learns what matters to you and gets smarter about what needs your attention.

None of this is a GUI pattern. It's just how humans actually work together.

The design shift

Here's the core insight that changes everything: you're not designing a system. You're designing a working relationship.

Think about how you'd hire a colleague. You wouldn't hand them a carousel to click through or force them to fill out forms to do their job. A good colleague remembers context, anticipates needs, knows when to ask permission and when to just act. They build trust through competence.

That's the design brief for agents. Not "How do I make this system more efficient?" but "How would I want to work with this person sitting next to me?"

If you wouldn't make a colleague do it, don't make the agent do it. If you'd trust a colleague to decide and report back, design the agent for that autonomy. If you'd want a colleague to show their thinking, build that into the agent.

The companies cracking this first won't be the ones with the fanciest LLMs. They'll be the ones that understood this was a fundamentally different design challenge—and had the discipline to start from zero.

Read the full article →

September and October brought several technical developments worth tracking: memory architectures that actually work in production, Claude Sonnet 4.5's coding capabilities, and a funding landscape that's consolidating around fewer, larger bets. LangChain hit unicorn status while Anthropic raised $13B - both pointing to infrastructure over experimentation.

Memory Developments: Practical Progress

September 19, 2025 - Researchers unveiled a procedural memory architecture enabling AI agents to learn, store, and reuse operational steps, reducing retraining costs. The framework helps agents build cumulative expertise for complex, multi-phase tasks. (Crescendo AI)

September 9-11, 2025 - SK hynix showcased its AiM (Accelerator-in-Memory) solution at AI Infra Summit in Santa Clara. The vLLM framework integration provides stable support for long token generation in reasoning models, addressing cost and power consumption challenges in LLM service management. (TechPowerUp)

August 28, 2025 - Researchers developed a procedural memory architecture for AI agents that incrementally learns and stores operational steps, reducing costly retraining while improving long-term performance. (DrugTargetReview)

These aren't lab experiments. SK hynix's implementation reduces operational costs versus GPU-only systems. The procedural memory framework is targeting enterprise deployments where retraining costs matter.

Implementation note: If you're evaluating AI tools, ask vendors how their systems retain context across sessions and whether they can learn from past failures without retraining. Memory architecture is now a basic requirement, not a differentiator.

Key Resources:

• Procedural Memory Architecture (August 28, 2025)
• SK hynix AiM Solution (September 9-11, 2025)

Model Launches: Claude Sonnet 4.5

September 29, 2025 - Anthropic released Claude Sonnet 4.5, achieving 77.2% on SWE-bench Verified with a 200K token context window (expandable to 1M). The model maintains focus for 30+ hours on multi-step tasks, compared to Claude Opus 4's 7-hour capability from May. Pricing remains at $3/$15 per million tokens. (Anthropic, TechCrunch)

September 29, 2025 - GitHub integrated Claude Sonnet 4.5 into Copilot for Pro, Pro+, Business, and Enterprise users. Early testing shows improvements in tool orchestration and context editing. (GitHub Changelog)

Cursor CEO Michael Truell and Windsurf CEO Jeff Wang both cited state-of-the-art performance on longer-horizon tasks. Anthropic researcher David Hershey reported watching the model autonomously build applications, stand up databases, purchase domains, and perform SOC 2 audits during trials. (TechCrunch)

The model also shows improvements on OSWorld (computer use benchmark), scoring 61.4% versus Claude Sonnet 4's 42.2% from four months prior. (VentureBeat)

What matters: Extended context windows may eliminate RAG complexity for some use cases. For coding applications, Claude Sonnet 4.5's 30-hour autonomous capability represents meaningful progress for sustained work.

Key Resources:

• Claude Sonnet 4.5 Launch (September 29, 2025)
• TechCrunch Coverage (September 29, 2025)
• GitHub Copilot Integration (September 29, 2025)

Funding: Consolidation Around Infrastructure

October 20, 2025 - LangChain raised $125M at $1.25B valuation in Series B led by IVP, with participation from CapitalG, Sapphire Ventures, Sequoia, and Benchmark. The company generates $12-16M ARR as of June 2025. (TechCrunch, Fortune)

LangChain's framework enables developers to build AI agents with ~10 lines of code using prepackaged components. The company's LangSmith platform provides observability and evaluation tools for production deployments. Competition includes LlamaIndex, Haystack, and built-in capabilities from OpenAI and Anthropic. (SiliconANGLE)

September 2025 - Anthropic secured $13B in its largest Q3 funding round, led by Iconiq Capital with Fidelity Management and Lightspeed Venture Partners. (BestBrokers)

Q3 2025 Overview - AI startups raised $64.3B across 2,427 deals, continuing the decline in deal volume while total investment increased. This compares to $55.3B across 2,655 deals in Q2 2025. (BestBrokers)

The pattern is clear: fewer deals, larger checks, emphasis on proven enterprise applications over experiments.

For founders: Demonstrate unit economics and deployed systems with paying customers. The market for pilot projects is contracting while infrastructure companies with integration stories get funded.

Key Resources:

• LangChain $125M Raise (October 20, 2025)
• Fortune Coverage (October 20, 2025)
• Q3 2025 VC Trends (October 2025)

Investment Climate: Structural Shifts

Q3 2025 - Applied AI investment reached $17.4B, up 47% year-over-year. Projections suggest agentic AI spending could reach $155B by 2030. The focus has shifted from developing LLMs to integrating AI into workflows. (Lexology)

2025 YTD - AI startups have raised $192.7B, representing 52.5% of total venture capital investment. This exceeds the entire 2021 AI boom ($168.1B). However, deal count dropped 44% since Q1 2022, from 16,821 deals to 9,358 in Q3 2025. (BestBrokers)

October 2, 2025 - Global AI investment is projected to exceed $2 trillion by 2026 (nearly 2% of global GDP), despite current returns lagging expenditures. Sustaining this pace may require $2 trillion in annual revenue by 2030, with an estimated $800B industry deficit under optimistic assumptions. (TechXplore)

OpenAI pulled in ~$4.3B in revenue during H1 2025, serving 700M users. The company raised ~$40B in March 2025, bringing valuation to ~$300B, and is valued at $500B in employee share sales. (TechXplore)

Investors are prioritizing startups demonstrating enterprise adoption traction. Deal terms emphasize integration over innovation, with acqui-hires resurging and talent retention clauses becoming standard. (Lexology)

Context: The math matters. Data center investments through 2030 will cost $500B annually, requiring $2T in annual revenues to justify expenses. Current trajectory suggests an $800B deficit even under optimistic assumptions.

Key Resources:

• Applied AI Investment Trends (Q3 2025)
• AI VC State (October 2025)
• Investment Economics (October 2, 2025)

Agentic Development: Infrastructure Over Theory

The technical work in September-October focused on production deployment rather than capabilities research. LangChain's unicorn status reflects demand for orchestration tools that actually ship. Anthropic's Claude Agent SDK exposes production patterns for memory management and sub-agent coordination. (TechCrunch)

The shift from "can agents work?" to "how do we deploy them safely?" continues. Organizations need orchestration platforms to manage multiple AI agents and applications. Competitors to LangChain include Microsoft's Magentic and LlamaIndex. (Fortune)

Implementation patterns: Start with bounded workflows, build evaluation systems before deployment, accept architectural messiness while optimizing for shipping. Microsoft VP of AI Agents Ray Smith noted that breaking processes into multiple agents improves maintainability and reliability versus squeezing everything into one agent.

Reality check: 77% of prompts through Claude's API involve task automation rather than advice, indicating actual production deployment. The challenge remains changing human behavior - employees choosing manual methods over AI tools despite deployment.

Key Resources:

• LangChain Series B Details (October 20, 2025)
• Fortune Analysis (October 20, 2025)

We have a diverse group of 5 professional services business in our early access programme, with another one to be added shortly. These are all deliberately small organisations, ranging between only 5-30 staff. However, they cover the majority of our customer types: consultancy, architecture, tech services, and finance advisory.  They also have a range of priority business challenges that mean we are not only testing and refining Payment and Benefits, we're also working with them on our 'next stage' Team and Project capabilities.

So what have we learned so far? Well, for a start we are now on version three of our self-service onboarding flows. Our previous testing phases were obviously a bit too forgiving on the signup, setup and activation of new company and employee accounts...and the good news is that we have significantly streamlined and automated this area.

The second focus area is using the new payments data to enhance Folio's expense management intelligence, improving the % of receipts that are accurately processed without human input. The target is 100%, but we're currently floating in and around 50%. So consider this productivity improvement half-complete.

If you want to participate in the early-access beta and get hands-on experience of Folio - just let us know!

Get early access here →

Market intelligence

We've launched our new, proprietary company enrichment model. This allows Folio to independently build insights into customers clients, suppliers, and merchants - supporting Folio to make better judgements on everything from expense approvals and invoice payments through to client suitability for new professional services offerings.

The previous service that we had been using for merchant intelligence was too limited for our needs, and as a third-party service it introduced a dependency for Folio. This means it was prioritised for replacement with a proprietary model as part of our 'dig the moat' initiative. For anybody who wants to nerd out about how this works behind the scenes, please just reach out.

It's ALIVE!

We've added Google integration to Folio - giving Workspace users the ability to sign in with Google, and for Folio to help manage their diary, inbox, and information. This is another big labour save from the Folio co-worker, and opens up access to a huge potential customer base.

Slippery onboarding flows

We have gone through two refinement iterations on our self-service onboarding flows this month...all based on feedback from our early access programme. This has been a labour of love for the Folio team: it's the first interaction that customers will have with Folio, and we want to make it as friction-free as possible. Hopefully we're there now!