Enterprise AI adoption: are ‘forward deployed engineers’ the whole solution?

Something strange is happening in the state of enterprise AI. Almost 90% of companies report using AI in at least one business function, yet only 5% are actually achieving value at scale. The remaining 95% are stuck in the infamous “pilot trap”: endlessly experimenting, rarely deploying, never scaling.

Meanwhile, a parallel reality is unfolding on the ground. Menlo Security documents a 68% year-over-year surge in shadow AI usage.

This striking disconnect raises a critical question: are organisations piloting the wrong initiatives while their employees have already discovered what actually works in what’s called ‘shadow AI’?

The hidden reality, according to BCG, is that 54% of employees would use AI tools even if explicitly unauthorised by their company. Employees aren’t waiting for permission; they’ve discovered that consumer-grade AI tools solve some real problems and they’re using them anyway, which means individual AI adoption is racing ahead even as organisational adoption stalls.

This gap is creating a massive, fragmented opportunity in the enterprise AI adoption layer: for professional services firms managing ‘integration hell’, for training platforms teaching workforce literacy and for software providers embedding AI into workflows.

AI promises dramatic efficiency gains. In theory, it should lower costs, increase output and reshape how organisations operate. But in practice, enterprise‑wide impact remains elusive.

So now we’re past the hype cycle and into the messy middle, where real money is made by companies that can actually deliver ROI at scale.

Read on for:

• A breakdown of the AI adoption category
• Market opportunities where capital is flowing
• A market map, drawn from analysis of +2.5k companies in this space
• What founders should build to win
• Our bets for how this plays out through 2030

* * *

“You can see the computer age everywhere but in the productivity statistics.”

Nobel laureate Robert Solow’s quip from the 1980s gets cited in every AI adoption think-piece. The implication: be patient. It took decades for productivity gains to materialise after electrification and computerisation. Even traditional ERP implementations took years to go live, with value realisation another 5–10 years after that.

AI will follow the same curve. The gains are coming — just wait.

We think this framing is a trap.

The ‘Productivity J-Curve’ suggests that companies initially see productivity dip during early adoption before a sharp rise in later stages as the tech matures and people get over the learning curve associated with new tools.

The Solow Paradox isn’t a timing problem. It’s a measurement problem. Productivity statistics measure output per worker. But AI adoption isn’t primarily about doing the same work faster — it’s about doing different work better.

So, the gains won’t arrive automatically as the technology matures. They’ll only arrive when organisations are forced to change how they work, just as electrification required physical reorganisation (new factory layouts, new machinery, new buildings) and computerisation required administrative reorganisation (new departments, new workflows, new job categories).

For AI, that’s happening now. Two catalysts are compressing what might otherwise have been a decade-long transition:

First, the EU AI Act, specifically Article 4, which mandates “AI literacy” for staff operating AI systems, effectively converting adoption software from a discretionary L&D expense into a non-discretionary compliance requirement.

Second, the technological shift from chat-based Generative AI to Agentic AI — autonomous systems capable of executing complex workflows — is forcing a transition in the enabling layer from simple prompt training to sophisticated governance and orchestration.

These forces are why, at Emerge, we believe 2026 marks a turning point:

• Phase 1–2023–25 (Investment & disruption): Companies spend heavily on software and consulting. Employees are distracted by learning new tools. Productivity dips.
• Phase 2–2026 (The turning point): ‘Integration Hell’ is resolved. Governance frameworks are established. Pilot programs identify high-value use cases.
• Phase 3–2027-beyond (Exponential payoff): AI agents move from ‘assistants’ to ‘autonomous actors’. Marginal cost of complex tasks collapses. Productivity spikes.

We’re in phase 2 now.

So… won’t this adoption gap just close on its own?

As models get smarter, faster and cheaper, surely adoption will follow? After all, if AI becomes easier to use — requiring less prompting skill, less technical knowledge, less trial and error — won’t employees just figure it out?

That assumption underpins a lot of enterprise AI strategy today — and it’s wrong. The blockers are structural, rooted in incentives, processes and skills. Better models don’t automatically fix broken workflows or retrain workforces for Phase 3. Easier-to-use AI still requires humans to change how they work.

Well, won’t agentic AI make adoption irrelevant anyway?

This is a different argument. There’s a seductive version of an agentic future which isn’t about AI becoming easier to interact with; it’s about humans not needing to interact with it at all, where adoption work simply disappears. AI agents handle tasks end-to-end; humans don’t need to learn new tools because the tools simply run themselves.

We don’t buy it.

First, agents still require human oversight. The EU AI Act mandates it for high-risk systems, but honestly, most enterprises will demand it regardless. Someone needs to instruct and supervise these systems; someone has to interpret outputs and handle exceptions. The skills are different. The need for enablement is not.

Second, agents don’t skip the adoption layer — they depend on it. They can’t operate on siloed data they can’t access. They can’t execute workflows that haven’t been redesigned. They can’t navigate governance frameworks that don’t exist.

Agentic AI raises the ceiling on what’s possible. It doesn’t lower the floor on what’s required to get there.

Four barriers to enterprise AI adoption

This isn’t a story about AI failing to catch on. At the individual level, adoption has been unprecedented: ChatGPT infamously became the fastest consumer application to reach 100 million users in history. Almost a billion people now use AI tools weekly. By any historical benchmark, this pace is remarkable.

The stall we’re describing is specifically inside organisations: behind the firewall, within corporate governance structures, where pilot programs proliferate but scaled deployment remains elusive. Anecdotally, we hear this even from companies at the centre of the AI ecosystem; firms with significant stakes in AI platforms struggle to drive adoption of their own tools internally.

What’s going on?

1. Organisation: strategy gap and organisational barriers leading to superficial implementations

Many AI initiatives are doomed to failure because the underlying business processes have not been optimised to leverage their potential.

In many organisations, C‑level leaders do not yet have a clear, shared understanding of what AI can realistically do today, where it can drive material efficiency and what needs to change to unlock that value.

There is also the structural fact that often no one inside organisations is truly accountable for AI-driven productivity outcomes, nor are those outcomes measured in a way that forces behavioural change:

• AI budgets sit in IT or “innovation”
• Productivity gains are diffuse, lagging and hard to attribute
• Managers are not rewarded for eliminating work, only for maintaining delivery
• Cost savings often disappear into the organisation rather than flowing to a P&L owner

As a result, AI can be used everywhere while owned nowhere.

Even when strategy is sound, execution often stalls in the middle of the organisation, for a variety of reasons:

• AI framed as an innovation side project rather than a core operating model shift, with middle management protecting existing teams, budgets and processes
• Focus on tools rather than outcomes and redesigned workflows, with KPIs focused on stability nor transformation
• Incremental automation layered onto broken workflows, with no explicit mandate to remove or re‑architect legacy processes

AI adoption requires re‑designing processes from the ground up. Once processes are streamlined, organisations can strategically introduce AI where it will create maximum value.

“When we look at where businesses are with AI, it’s clear they’re eager to accelerate but are still navigating the complex realities of adopting it successfully and responsibly.

My advice is simple: start experimenting. You may not yet know what the ‘big’ AI opportunity will be, but you won’t learn to surf by standing on the beach. Get curious, get moving, get ready because when the right moment arrives, experience will be your advantage. At the same time, be clear about what your business needs and how AI can support it, so your strategy has direction.

Ultimately AI will reshape business models, and leaders will need to meet that moment with focus, clarity and speed — alongside the same fundamentals that have always defined strong businesses.”

Christine Foster, GM AI and Automation at Experian

2. Workforce: training gap leading to low engagement

The big problem here is engagement: while 88% of employees use AI at work, only 5% of employees qualify as advanced AI users that increase their efficiency by >20%, according to EY.

The same report also finds that employees with 81+ hours of AI training/year save 14 hours/week as opposed to savings of 3 hours/week with <4 hours of AI training. But, crucially, only 12% of employees received 81+ hours of AI training in the past 12 months.

A staggering 82% of companies in the early stages of AI maturity have not implemented a talent reinvention strategy, planned to meet workforce needs, or acquired new talent or training to prepare workers for generative AI-led workflows.

Without training, there are strong human barriers to AI adoption:

• Fear of automation and job displacement
• Low confidence in how to use AI tools effectively
• Lack of understanding around how to connect tools to daily work

3. Data: legacy systems not built for AI agents

Most enterprise data infrastructure was not designed for agentic use:

• Siloed systems of record
• Inconsistent data schemas and permissions
• High friction in accessing and combining data

One recent CIO survey highlights that 85% of companies struggle to integrate AI due to poor data quality and legacy architecture. A single AI initiative often requires custom integrations with 15–30 distinct systems.

These issues surface upstream as usability problems, limiting what AI systems can safely and reliably do.

Crucially, McKinsey analysis highlights that 90% of the opportunity here lies in unstructured data — not structured systems-of-record but the messy corpus where real value sits: emails, chats, call recordings, PDFs, video, images…

Many AI adoption programs fail because they start where data is clean (structured) rather than where value is large (unstructured).

4. Governance: trust and compliance control scaling

The adoption gap in large enterprises is also inextricable from permissioning, auditability, model risk and accountability: legal/compliance sign-off cycles, procurement, third-party risk, IP leakage fears and regulated data boundaries.

Scaling requires controls…controls slow deployment…slow deployment increases shadow AI.

The full implementation of the EU AI Act in 2026 has transitioned this from a theoretical risk to a mandatory compliance market. The Act imposes strict obligations on “high-risk” AI systems, including human oversight, data governance and transparency. Because re-engineering a product for one region is costly, the EU AI Act effectively sets the global standard; US-based companies are adopting EU compliance standards globally to streamline development. Non-compliance carries penalties of up to 7% of global turnover, a risk profile that forces AI adoption out of the shadow IT laboratories and into the boardroom.

Key takeaways:

The adoption gap is structural — it is produced by incentives, not tech immaturity.

This has created a status quo of over-experimentation, under-deployment which is ripe for disruption.

“Companies that get real value from AI narrow the scope, integrate deeply and ship at ‘good enough’ rather than chasing perfection. They bring risk and compliance in early, give one person real accountability for outcomes, and — most importantly — design AI to replace default human decisions, not just sit alongside them as advice.

So startups underestimate how explicit decision-making needs to be. Enterprises don’t buy AI for insight — they buy it to take actions — and if it’s not clear who is supposed to act on the output, when it is binding or what happens if it makes a mistake, it will never get past a pilot.”

Shay David, chief data solutions at SolarEdge and Emerge Venture Partner

Market opportunities created by the adoption gap

The AI adoption problem is creating a large and fragmented opportunity space. What is emerging looks like a spectrum, from services‑led on the left to technology‑led on the right.

Left: Services‑led solutions

1. a) Consulting Process re‑engineering and organisational change consulting focused on overcoming structural and cultural barriers to AI adoption.

1. b) Implementation Hands‑on technical execution: deploying tools, integrating systems and operationalising new AI‑driven workflows.

These approaches are effective but labour‑intensive and hard to scale.

Middle: training

1. c) Training and enablement This spans a wide range:

• High‑end, charismatic thought leaders delivering $50k+ executive workshops to shift C‑suite understanding
• Dedicated upskilling platforms with scaled, asynchronous, interactive courses teaching the broader workforce practical AI techniques and tools

Training addresses skills gaps, confidence issues and some organisational resistance, but on its own rarely changes systems.

Right: Technology‑led solutions

1. d) Digital adoption platforms Software overlays that guide users through enterprise applications via walkthroughs and pop-ups.

These products have traditionally optimised for click-through compliance — teaching employees to fill in the right forms, not to change how they think.

1. e) In-workflow AI enablement Real-time guidance, co-pilots and agents embedded inside AI interfaces (Copilot, ChatGPT, etc) that help users understand what to do, when to do it and how to do it — inside their existing tools and workflows. These products aim to lower friction, build confidence and embed AI usage into daily work.

Today, this means teaching people to work with chatbots. Tomorrow, it means dynamically generating content in response to newly deployed agents — guidance humans couldn’t write fast enough on their own.

The market map, unpacked

Here is our market map of the AI adoption category.

Note: this map does not focus on AI or data infrastructure (e.g. model training, governance/observability or dev tools) nor vertical applications for workflow automation.

On the left-hand side, we have services-led solutions, split into consulting (advisory/strategy) and implementation (delivery/execution).

In the middle are training solutions, split into legacy aggregators (MOOCs), enterprise L&D platforms (including skills intelligence) — both subcategories here are learning platforms that teach AI skills outside the workflow. Beneath this are dedicated AI upskilling platforms.

On the right-hand side, we find tech-led solutions…

Services‑led solutions

a) Consulting

• Definition: Management consulting firms that define where and why an enterprise should adopt AI. They focus on change management, organisational design and data architecture, risk frameworks and high-level use case identification.
• Revenue Model: Retainer-based or Project-based fees (High Margin/Low Volume).
• Key Value: Risk mitigation, board-level assurance and agentic roadmap design.

Big Players: McKinsey, BCG, Bain & Co, PwC, Accenture, Capgemini, Deloitte, Infosys, KPMG, EY, Tata Consultancy Services, Genpact.

Analysis: The professional services sector faces a ‘cannibalisation paradox’. AI threatens the traditional model (a pyramid of junior staff billing hourly for routine work), yet the demand for implementing these technologies is driving record bookings. The sector is in a race against time: it must cannibalise its own low-value revenue streams before competitors or automation do, while simultaneously building high-margin, asset-based revenue streams.

Fun facts:

• Accenture: Reported substantial gains in its AI business, with GenAI revenues tripling to $2.7 billion and bookings nearly doubling to $5.9 billion in FY2025. Total bookings reached $80.6 billion, signalling robust demand.
• Capgemini: GenAI and Agentic AI accounted for more than 7% of group bookings in Q2 2025, driven by “Intelligent Operations”.
• BCG: 20% of BCG’s revenue in 2024 came from AI and GenAI.

b) Implementation

• Definition: Large-scale technology services firms that handle the physical integration of AI models into legacy tech stacks, data sanitisation and cloud infrastructure.
• Revenue Model: Mixed model: Project-based fees for implementation + recurring revenue for managed services.
• Key Value: Scale, speed of deployment and management of ‘data gravity’.

Big players: Palantir, Accenture, Deloitte, Wipro, Infosys, Capgemini.

Analysis: To counter the erosion of billable hours, many consulting firms are aggressively pivoting toward managed services. In this model, the consultancy builds the AI solution and then also runs it for the client, charging a recurring fee based on outcomes (e.g., claims processed, customers served) rather than hours worked. This shifts the risk of efficiency to the provider; if they can use AI to do the work with fewer humans, they keep the margin upside. This creates a ‘data gravity’ effect where they manage the entire lifecycle by owning the ‘plumbing’ of the AI rollout: the enterprise data estate.

As enterprise clients increasingly move from ‘Build in-house’ to ‘Buy/Configure’, some players such as Infosys’ ‘Topaz’ and Capgemini’s ‘RAISE’ are creating ‘AI factories’ — a kind of industrialised delivery platform — by deploying pre-built ‘capability pathways’ that solve 80% of the problem out of the box. These are libraries of 12,000+ pre-built AI assets, connectors and prompts. Instead of building a customer service agent from zero, they deploy a pre-configured module that only needs tuning for client data.

Fun fact:

• Accenture’s managed services revenue grew 9% in FY2025, outpacing consulting growth of 5%.

Training-led solutions

• Definition: Specialised software tools focusing on ‘just-in-time’ learning, prompt engineering literacy and human-behavioural change.
• Revenue Model: SaaS (Per User / Enterprise License).
• Key Value: Skill acquisition and meeting regulatory literacy mandates.

Big Players: Coursera, Udemy, Skillsoft, Pluralsight.

Analysis: The competitive landscape here is fragmented, with many players approaching the problem from different angles.

For the first time, training budgets are being tied directly to productivity outcomes rather than just completion rates. Studies show that AI training leads to a 29.6% skill gain and a 35.5% rise in worker productivity. This hard data allows L&D heads to defend budgets even in a cost-cutting environment. As a result, corporate training budgets are expanding specifically for AI: executives plan to allocate 5% of annual budgets to AI, with a heavy focus on workforce upskilling.

We see the outlook here as market bifurcation: the market will split between “commodity content” (cheap video libraries, likely AI-generated) and “high-value cohorts” (expensive, interactive, outcome-driven programs led by human experts). The middle market of generic training will collapse.

Legacy players (such as Coursera and Udemy) offer deep AI curricula — structured pathways, hands-on projects, certifications — but rely on vast libraries of pre-recorded content which will become obsolete in weeks as models update and as content is delivered outside the workflow, engagement suffers. These solutions get considered when budgets are tight, but they struggle to prove ROI on real AI productivity.

Another cluster revolves around behaviour change. Cognician delivers science-backed microlearning through chat-based challenges and behavioural nudges — such as email and Teams reminders that prompt employees to use AI tools. This approach builds habits and tracks behaviour shifts, but it’s still push-based: nudging people toward the tools rather than guiding them inside the tools.

The AI era demands cohort-based and applied learning, which will be met by a new breed of AI training platforms. AI agents will deliver training inside the workflow (e.g., a Salesforce agent teaching a salesperson how to prompt effectively during a sale). This performance support will kill off traditional course-based learning.

Fun facts:

• Coursera: Pivoted to “AI-native learning” with ChatGPT integration. Q3 2025 revenue grew 10% driven by consumer demand for AI skills. They are focusing on “Generative AI Academies” to provide structured learning paths.
• Udemy: Reported an 859% year-over-year growth in GenAI for productivity learning. However, their marketplace model faces quality control challenges with the flood of low-quality AI content.

Technology‑led solutions

d) Digital adoption platforms

• Definition: Software infrastructure that layers directly over existing applications to guide user behaviour through automated walkthroughs, pop-ups and field validations delivered directly in the workflow, with robust behaviour analytics, adoption funnels and friction alerts.
• Revenue Model: SaaS (Per Seat, MAU, or Application Tier).
• Key Value: Process compliance, forced adoption and shadow AI visibility.

Big Players: WalkMe, WhatFix, Pendo

Analysis: DAPs emerged to solve a legitimate problem: expensive software sitting unused because employees couldn’t figure out how to navigate it. SAP’s 2024 acquisition of WalkMe, the pioneer in this category, for ~$1.5B signals that DAPs are becoming a feature of the ERP stack rather than a standalone.

But there’s a structural limitation. DAPs are designed around guiding clicks, not guiding thinking. They work well for deterministic workflows: fill this field, click this button, submit this form. AI tools don’t work that way. A Copilot prompt isn’t a click sequence; it’s a reasoning task. DAPs can track whether someone opened Copilot, but not whether they prompted it well. DAPs can tell you where to type but not what to think. Pendo has moved furthest toward AI-specific tracking, including visibility into agent usage and proactive recommendations for agent maintenance, but it remains analytics-first — strong on measurement, thin on enablement.

Right now, we think the market is mispricing the DAP category by viewing it as an AI-enabled evolution of legacy tech. In reality, DAPs are built for the attention economy, managing where a user looks and clicks within a fixed UI.

And the implementation model is heavy: clients typically build their own content and maintain their own integrations, a maintenance burden that scales poorly with AI because the ‘guides’ frequently break when application UIs update, requiring manual intervention. All this produces a high total cost of ownership; enterprise implementations can take 3–6 months and often require dedicated administrators with coding knowledge to maintain overlays.

Fun fact:

• WalkMe’s growth had begun to slow significantly (to ~2%) before its acquisition, which suggests that the DAP model may have hit a saturation point.

e) In-workflow AI enablement

• Definition: A GenAI-native category that provides real-time guidance embedded directly inside AI interfaces which adapts to what employees are actually doing — teaching them how to prompt, iterate and think with AI tools in the flow of work.
• Revenue Model: SaaS (Per User / Enterprise License), often usage-based.
• Key Value: Skill acquisition in context, measurable productivity gains and AI-specific ROI tracking.

Big Players: Mendo

Analysis: This category starts from a different premise. Employees don’t struggle to find the chat box — they struggle to know what to ask, how to refine outputs and when to trust the response.

In-workflow AI enablement represents a paradigm shift, from interface guidance to cognitive orchestration. Unlike DAPs, which reside on top of tools, in-workflow enablement platforms integrate directly into the AI stack to trigger the right habits at the right moment without introducing a new tool.

Few players have cracked this, which makes it an emerging (and exciting!) category. The barrier to entry is integration depth: actually embedding inside AI chat interfaces, not just pointing employees toward them. For example, Mindstone offers a browser extension that provides inline coaching inside Copilot, Gemini, ChatGPT and Claude, suggesting prompt improvements as users type.

The future of this category will look like platforms that generate guidance dynamically by parsing agent documentation, surface edge cases as agent fleets scale and enabling humans as supervisors rather than just users.

We are right at the start of a transition of the human worker from ‘Operator’ to ‘Apprentice Master’. As agentic AI begins to execute multi-step workflows autonomously, the skills gap is not learning how to use the tool, but learning how to supervise the agent. We’re calling this the ‘Apprenticeship Loop’.

Legacy DAPs are structurally incapable of building this Apprenticeship Loop because they are static; they cannot read an agent’s real-time reasoning or its shifting documentation. The in-workflow AI enablement category is the only one building the infrastructure to provide dynamic enablement, automatically generating guidance for humans to audit and steer agents that may only exist for a single week or project.

“One problem Microsoft have identified is that 70% of their GenAI projects fail at user training — not at technical setup. That matches what we were seeing. Companies deploy the tools, but employees don’t really use them. Or they use them badly.

Even one of the most advanced GenAI companies we work with, who’ve invested a lot in training, found that 80% of their people were just using AI to make emails sound better. That’s not changing how you work. That’s a spell-checker! The productivity impact is so low.

So we built Mendo to sit inside the AI tools enterprises already have — Copilot, ChatGPT, whatever — and push people toward real meaningful use cases for their job. Our training tabs all spawn inside a user’s live workflow, with knowledge diagnostics leading to custom learning pathways and content tailored to roles so we can boost usage by giving people specific AI skills that are smart for their jobs. And then the next step is we give them prompts at the end of these use cases that they could use to create their own assistants.

We then track whether people are using AI multiple days a week, whether they’ve built something recurring, whether they feel confident. That confidence metric is interesting: we typically move it from around 40% to 95% of users within three months.

On adoption, we usually triple daily usage in that same period. About half our users actually end up creating their own assistants, and more start putting in place recurring use cases so they will gain somewhere between 3 to 10 hours per week.

We work with EY, PwC, Danone, Crédit Agricole — more than 100 enterprises now. But the thing we try not to do is sell time savings. Everyone claims that. We focus on getting companies ready to rethink their processes. That’s where the real ROI is.”

Quentin Amaudry, cofounder and CEO, Mendo

How startups can win

This section offers concrete, pragmatic advice for founders looking to build models that are highly scalable and investable.

Services-led: or, Workflow-Transformation-as-a-Service

There is real money to be made in AI adoption services. But most services businesses are not venture‑scale by default.

How do you transform from an unscalable consulting shop into a multi‑billion‑dollar, trusted brand delivering a repeatable, turnkey AI adoption solution?

The model that’s emerging — we call it ‘Workflow-Transformation-as-a-Service’ — bundles three things into a single subscription: process redesign (consulting), system integration (implementation) and in-workflow tooling (software). Companies don’t buy a strategy deck and then separately buy implementation and separately buy software. They buy outcomes.

This is where the big incumbents are already moving. Every major consultancy/service implementer has aligned with an AI lab, with differentiated strategies ranging from broad partner networks and ecosystems to more exclusive model alliances. And when partnership isn’t enough, they’re simply buying capability: Accenture’s acquisition of Faculty AI signals that AI-native implementation talent and productised decision-intelligence are now viewed as strategic assets to collapse the gap between strategy and delivery. This follows Accenture’s acquisition of Udacity, vertically integrating the ‘skilling’ supply chain to productise its workforce transformation.

Smaller players can’t out-scale Accenture. But they can out-specialise.

What founders should build:

Pick a vertical and own the full stack. Generic ‘AI transformation’ is a losing game against incumbents with infinite bench depth. Verticalised offerings — AI adoption for insurance claims, for pharma R&D, for logistics ops — can charge premium rates and build defensible expertise. The narrower the wedge, the faster the sale.

Productise early. Turn repeatable delivery into software. Every engagement should generate reusable assets: workflow templates, integration connectors, training modules. These become the product layer that eventually scales without headcount.

Scale through M&A. Organic growth in services is slow. Inorganic growth — acquiring small consultancies with complementary verticals or geographies — can compress timelines. The playbook: buy boutiques, integrate their delivery onto your productised backbone, expand coverage without proportionally expanding headcount.

Shift the buyer. CIOs control tech budgets but often aren’t accountable for productivity outcomes. COOs and CHROs are. Position Workflow-Transformation-as-a-Service as an operating model shift, not a technology purchase. The prize is a seat at the leadership table, not a line item in IT spend.

Technology‑led paths

Technology‑led approaches to AI adoption are more scalable, but face a different tension: the underlying AI capabilities are moving too fast for point solutions that cannot adapt in parallel.

The solution is an approach designed to evolve alongside AI itself. That means three things: in-flow enablement (teaching employees to work with AI where they already are), ROI and usage analytics (proving value back to buyers), and agent adoption infrastructure (onboarding humans to new agents as fast as agents ship).

What founders should build:

Embed in the AI interface, not alongside it. The gap in the market is integration depth. Browser extensions and nudge systems point employees toward AI tools. The platforms that win will live inside Copilot, ChatGPT, Gemini and Claude, triggering guidance at the moment of need, not before or after.

Instrument everything. Adoption without measurement is a cost centre; adoption with measurement is a strategic investment. Track prompts, not just logins. Track reasoning patterns, not just feature clicks. Build dashboards that tie AI usage to business outcomes: time saved, tasks completed, revenue influenced. This is what gets contracts renewed.

Build for agent velocity. Chatbot adoption is table stakes. The next wave is agent adoption — and agents ship fast. A new internal agent can spin up in days, soon to be hours. A new vendor agent can launch with a week’s notice. Static training content can’t keep pace. The platforms that win will generate enablement dynamically, parsing agent documentation, inferring capabilities, surfacing edge cases and updating guidance as agents evolve.

Own the Apprenticeship Loop. As agents handle execution, humans shift from operators to supervisors. This is a skills transition most organisations, let alone employees, aren’t ready for. The platforms that parse a newly deployed agent’s documentation in real-time to teach employees how to audit, steer and exception-handle agent output will capture structural demand for tomorrow’s autonomous AI fleets.

Design for enterprise privacy from day one. Many adoption platforms route data through centralised cloud servers. For regulated industries, that’s a non-starter. In-browser classification, on-premise deployment options and data residency controls aren’t nice-to-haves — they’re deal-breakers. Build the architecture now; retrofitting is expensive.

“Behaviour changes when you see someone get an advantage and you want it for yourself. If you see a colleague avoid a parking fine by generating a complaint letter in three seconds — next time you get a fine, you know what to do. That’s not training. That’s social learning. It’s fear of missing out.

I’ve seen enormous lists of AI-generated use cases. The problem is they don’t ring true; they don’t have the frustration/delight/pride, the emotion built in. Compare that to someone saying ‘I was pissed off with my boss for 15 years before I could do this.’ Feelings travel. That’s what changes behaviour.

The opportunity is technology that accelerates peer-to-peer. Make best practice visible. Cross-pollinate what’s working. ‘Jane did this, Victor did that.’ It’s mostly bottom-up, with some top-down stimulus.”

Marc Zao-Sanders, author of ‘How People Are Really Using Gen AI in 2025’ (Harvard Business Review’s #2 top read of 2025) and Timeboxing (Penguin); founder, AI in the Wild.

Our bets for the AI adoption category through 2030

We see three plausible futures, but they collapse into a simple question: does AI adopt itself, or does it need help?

1. Base case (probability: 65%): AI adoption follows the historical J-curve pattern. Bottom-up adoption outpaces top-down strategy. Employees experiment faster than IT can govern. Companies scramble to catch up with what’s already happening.

• Winners: A new ‘mid-market McKinsey’ for AI, plus in-workflow enablement platforms that meet employees inside the tools they’re already using. Consulting helps organisations catch up strategically; in-workflow enablement helps them govern and optimise the adoption that’s already underway.
• Leading indicator to watch for: Macro productivity data. A tangible uptick in non-farm labour productivity starting late 2027.

2. Upside case (probability: 15%): Technological breakthroughs in reasoning models combined with plummeting inference costs allow AI agents to solve integration challenges autonomously.

• Winners: Massive winner-take-all dynamics for platforms that own the data.
• Leading indicator to watch for: Inference cost deflation. >80% drop in reasoning model costs within 12 months.

3. Downside case (probability: 30%): A major agentic failure (flash crash, data breach, critical infrastructure collapse caused by hallucination) triggers a compliance freeze. Adoption spend shifts from productivity to governance and insurance.

• Winners: Governance, compliance and cybersecurity firms. Adoption platforms that built deep auditability and privacy-first architectures (e.g. on-premise deployment).
• Leading indicators to watch for: Shelfware metrics. >20% of sold AI seats inactive after 6 months. Liability lawsuits: High-profile court cases holding enterprises liable for agentic errors.

So what does this mean for AI adoption companies?

The base case is the only scenario where the current AI adoption category thrives. In the upside case, AI implements itself so the AI adoption category collapses. In the downside case, spending shifts to compliance and risk control; AI spending collapses so the AI adoption category collapses. Either way, today’s adoption platforms face pressure.

This raises two questions every AI adoption company needs to answer:

1. Are you building for a one-off need or an ongoing one?

The bull case for adoption platforms assumes AI creates permanent change management demand: new agents shipping continuously, requiring fresh onboarding cycles.

The bear case: once employees learn to work with AI, the training need evaporates.

History suggests the truth is somewhere between. Database implementation consultancies of the 1990s didn’t disappear; they evolved into ERP practices, then cloud migration, then data engineering. The work changed, but the need for human-to-system translation persisted.

2. What’s your strategy to do that?

Adoption platforms will face a choice:

• Lean into tech: Become infrastructure for agent governance, monitoring and maintenance. The play here is tooling that audits agent behaviour, manages permissions and tracks ROI across an expanding agent fleet. This is defensible if you own the data layer.
• Lean into services: Become the ‘mid-market McKinsey’ for AI — verticalised services firms that combine software with human delivery. The play here is productised transformation — industry-specific playbooks, packaged change management, ongoing optimisation retainers — growing through the aggressive acquisition of niche boutiques.
• Hybrid: Build software that enables a services layer without requiring it. Platforms that can serve both self-directed enterprises and consulting partners will capture more of the market.

What’s our bet at Emerge?

We’re positioned for the base case with upside optionality.

AI adoption is an ongoing need, but the type of work shifts.

Early-stage adoption is heavy on training and workflow redesign. The base case describes bottom-up adoption racing ahead of organisational strategy, thus:

• Consulting helps organisations catch up: redesigning workflows, building governance, identifying high-value use cases.
• In-workflow AI enablement is the natural solution for what’s already happening: rather than fighting shadow AI or waiting for top-down programmes, these platforms embed into tools and help companies harness what employees are already doing.

Mature-stage adoption, however, is heavy on governance, agent maintenance and continuous optimisation. This actually holds true regardless of which scenario plays out.

That means we’re backing founders who can survive the grind — real revenue, clear ROI metrics, enterprise-grade governance — while building vertical AI and/or agent infrastructure in case the breakout arrives faster than expected.