The $800M Credit War: How OpenAI, Anthropic, and Google Are Buying Tomorrow's AI Stack
Frigade's July 14 launch of Skills — an in-app AI assistant that completes tasks for users instead of guiding them through steps — marks the clearest signal yet that action-complete onboarding will replace the show-and-tell model that's kept median activation rates stuck at 37% for four years.
On July 14, 2026, Frigade announced Skills — and the press release was buried in a news cycle dominated by model benchmark announcements and funding rounds. That's a mistake.
Skills is the clearest commercial signal yet that the onboarding software category is about to undergo the same structural shift that AI brought to customer support: a move from assisted self-service to action-complete automation. The implications for how SaaS companies think about activation, retention, and product design are significant — and largely misunderstood.
The Activation Industry Built the Wrong Thing
For ten years, the SaaS onboarding industry has iterated on variations of the same core model: identify the steps users need to take to reach first value, then create progressively more engaging ways to communicate those steps. Product tours became interactive tours. Checklists became animated progress bars. Tooltips became contextual guidance overlays. Walkthroughs became video-embedded tutorials.
The industry spent a decade building better signs, and median activation rates haven't improved. The 2026 SaaS Activation Report from Perspective AI shows median B2B activation at 37%, and that number has been flat for four years despite the explosion in onboarding tooling investment. Top-quartile companies hit 68% or higher — not because their onboarding is better designed, but because their product reaches first value faster. The gap is not a communication problem. It is an execution problem.
The relevant benchmark is the three-day activation cliff: 90% of users who don't engage with core product functionality in their first 72 hours never do. The onboarding tools industry has been building better instructions for a process that users mostly stop attempting before instruction quality matters.
What's actually happening in those 72 hours is this: users understand what the product is supposed to do — that knowledge problem was largely solved by good marketing and better onboarding — but they hit execution friction when trying to set it up. The data import that requires matching column headers to schema fields. The campaign configuration that involves six decision points before anything runs. The dashboard setup that requires knowing which of twelve available widgets represents the metric the user actually cares about. Each of these friction points is not a knowledge problem. It is an effort problem. And effort is what action-complete AI eliminates.
What Frigade Skills Actually Does
The most important thing to understand about Frigade Skills is what it is not: it is not a chatbot embedded in your product. A chatbot answers questions. Skills completes actions.
Here is how Frigade describes the difference: when a user asks for last week's revenue report, a chatbot tells them which menu to navigate to and which report to run. Skills retrieves the report and delivers it. When a user asks to "set up my email integration," a chatbot explains what fields to fill in. Skills initiates the OAuth flow, authenticates the account, and confirms the connection.
The architecture is an AI agent — specifically, a function-calling LLM that can invoke the product's own capabilities as tools. The novel engineering is in the tool discovery: Skills learns what a product can do autonomously by analyzing the product's APIs, component library, and user action logs, and surfaces those capabilities as a tool set the assistant can invoke. As the product changes, the tool set updates without requiring engineering work.
For enterprise customers, Frigade supports self-hosted deployment with the customer's own LLM keys — meaning no product data leaves the enterprise environment. The assistant runs against the customer's LLM infrastructure. Frigade handles the tool discovery, conversation management, and the onboarding context layer. The model and the data stay inside the customer's perimeter.
The Architecture of Action-Complete vs Show-and-Tell
Traditional onboarding architectures share a common structure: detect where the user is in their journey, identify the next knowledge or action gap, and communicate the path to closing that gap. The user closes the gap.
Action-complete architectures detect where the user is, identify the next action required, and close the gap on behalf of the user. The key distinction is who executes.
| Dimension | Show-and-Tell (Traditional) | Action-Complete (Skills) |
|---|---|---|
| Primary mechanism | Instruction and guidance | Task execution on request |
| User required action | Follow steps provided | State the desired outcome |
| Cognitive load | Medium-high (understand, remember, execute) | Low (state intent) |
| Failure mode | User understands but doesn't execute | User never states intent |
| Data required | Funnel drop-off patterns | Successful action logs |
| Iteration cycle | Redesign guidance content | Expand action coverage |
| Retention signal | Activation event completion | Human follow-through rate |
| Enterprise requirements | Low (just content) | High (self-hosting, audit trails) |
The failure mode difference is the most commercially significant row. Traditional onboarding fails when users understand what to do but don't do it — the category failure that produced 37% median activation rates. Action-complete onboarding fails when users don't state intent — which is a much smaller population, because the bar for expressing intent ("set up my integration") is lower than the bar for executing a multi-step setup flow.
Why the Ceiling Exists at 37%
The 37% median activation rate is not evidence of bad onboarding design. It is evidence of a correct understanding of the wrong problem.
Consider the anatomy of a typical SaaS product's user cohort at activation checkpoint: - Segment 1: Engaged activators (~37%) — these users have already found their way to first value through their own initiative, prior experience with similar tools, or luck. - Segment 2: Guided completers (~15-25% additional in top-quartile products) — users who would activate with clear, friction-reduced guidance that meets them at their specific knowledge gap. - Segment 3: Execution-blocked users (~20-30%) — users who understand the product's value and know what they want to do, but are stopped by effort friction: too many steps, too many decisions, setup complexity that feels disproportionate to the value it unlocks. - Segment 4: Wrong-fit users (~15-20%) — users who will not activate regardless of onboarding quality, because the product does not fit their actual use case or org context.
Traditional onboarding improves conversion in Segment 2. It cannot move Segment 3, because Segment 3 is not blocked by knowledge — they are blocked by effort. Action-complete AI moves Segment 3 by removing the effort requirement entirely.
The activation revenue relationship matters here: AI-powered apps in 2026 generate 41% more ARPU than conventional SaaS, but also churn 30% faster among users who don't deeply engage with the core product. The failure mode for high-ARPU AI products is exactly Segment 3: users who understand the value, pay for it, but churn because the effort to extract that value exceeds their patience. Action-complete activation converts Segment 3 into retained customers who experience value immediately, then build the product habit needed for long-term retention.
The Action-Complete Activation Playbook
Implementing action-complete activation is not a drop-in product change. The following five-step sequence is how teams in production are approaching the transition.
1. Map your first-value task inventory. Before deploying an action-complete assistant, identify the 3-7 distinct paths through which users can reach first value in your product. Each path should have a clear starting state (what the user brings to the product) and a clear success state (the first moment where the product delivers tangible value). These paths become the action coverage target for your assistant.
2. Audit execution friction per path. For each first-value path, map the individual steps required and score each step on execution friction: how many product UI interactions does it require? How many decisions does the user need to make? How many external systems does it touch? Steps with high friction scores are the highest-priority targets for action-complete automation.
3. Define intent recognition triggers. Action-complete assistants don't proactively complete tasks — they respond to user intent. Define the natural language patterns that indicate user intent for each high-friction step. "Set up my [X]," "help me connect [Y]," "create my first [Z]" are common patterns. The assistant's effectiveness depends on how well it recognizes the full range of intent expressions for each action.
4. Build the human follow-through measurement. Before launch, instrument your product to track what happens after the assistant completes an action. Does the user interact with the completed configuration? Do they build on it? Do they return to the product in the following 48 hours? This data — human follow-through rate — is the leading indicator of whether action-complete activation is creating genuine product engagement.
5. Iterate on action coverage, not conversation quality. The most common failure mode in early action-complete deployments is treating the assistant as a conversation product and iterating on response quality. The right iteration metric is action coverage: what percentage of activation-blocking interactions can the assistant complete? A 60% action coverage rate with good conversation quality will outperform 90% conversation quality with 30% action coverage every time.
Which Products Win and Which Don't
Action-complete activation is not universally applicable. The mechanism works best when the gap between user intent and product execution is large and when first-value tasks are well-defined.
Data platforms, analytics tools, and BI software are the highest-value targets. The setup sequences for these products — connecting data sources, defining metrics, configuring refresh schedules, building initial dashboards — are execution-intensive and well-defined. Users know they want "a dashboard showing this week's revenue broken down by region." The path to that outcome involves fifteen product interactions. Action-complete AI reduces that to one intent statement.
CRM and sales tools are strong candidates for similar reasons. The first-value experience in CRM is importing contacts, configuring pipeline stages, and creating the first deal. These are rote tasks — nobody needs to think hard about how to import a CSV, they just need the CSV to be in the system.
Developer tools, infrastructure software, and highly customizable platforms are weaker candidates. The first-value experience in these products requires product-specific judgment that is unique to each user's context. A developer deploying infrastructure knows what they want but the right answer depends on their technical environment in ways the AI cannot fully anticipate. Action-complete assistance helps at the margins here — it can complete well-defined subtasks — but it cannot replace the user's judgment about the overall setup.
The Enterprise Implications
Frigade's pricing page reveals something important about the enterprise trajectory of action-complete onboarding: self-hosting is a first-class feature, not an afterthought. Enterprise software vendors evaluating activation tooling have two primary concerns beyond functionality: data residency and audit trails.
Data residency is resolved by the self-hosted architecture. No user action data transits Frigade's infrastructure when the assistant runs against the customer's own LLM keys. This satisfies the most common enterprise data governance requirement.
Audit trails are more complex. When an AI assistant completes an action on behalf of a user, the audit record needs to capture not just what was done but who requested it, what authorization was in place, and what the assistant's decision process was. These requirements are still being standardized across the enterprise AI deployment space — but they are the questions procurement teams are starting to ask, and vendors who design for them from the start have a durable advantage.
What to Measure When AI Does the Work
Traditional activation metrics — time-to-first-value, percentage of users completing the activation sequence, feature adoption breadth — all still apply, but they require a new dimension: attribution.
An activation event completed by the AI assistant is not equivalent to an activation event completed by the user independently. The AI-completed event demonstrates that the user knew what they wanted and could communicate it. It does not demonstrate that the user can navigate to that outcome again, build on it, or troubleshoot when it fails. Retention depends on the latter three capabilities.
The metric that distinguishes genuine activation from AI-assisted first impressions is human follow-through rate: after the AI completes the first-value action, does the user interact with the output? Do they modify it? Build on it? Return to the product the next day?
High human follow-through rates indicate that AI-assisted activation is creating genuine product engagement — the user experienced the value, understood it, and wants more of it. Low follow-through rates indicate that the AI is completing tasks without the user developing the product familiarity needed to become a retained customer.
Separately, products need to track the underlying retention dynamics that action-complete activation is designed to improve. The leading indicator is the ratio of assisted first-value completions to subsequent unassisted product actions in the following seven days. If users require AI assistance for every product interaction indefinitely, the AI is masking a product complexity problem rather than solving an onboarding one.
The Market Position
Frigade is not the only company working on action-complete onboarding. Appcues, Pendo, UserGuiding, and WalkMe have all begun integrating AI assistance into their onboarding platforms. But the architectural approach varies significantly.
Most incumbents are adding AI to existing show-and-tell infrastructure: the AI writes better tooltip copy, personalizes which checklist items to show, or answers questions in a chatbot sidebar. These are optimizations to the wrong model.
The distinction is architectural. Show-and-tell with AI assistance is still show-and-tell. Skills is built from the ground up as an action-complete system — the AI is the primary onboarding mechanism, not a supplement to conventional guidance. That architectural difference creates a different product and a different performance ceiling.
The market will sort out which approach wins, and the aggregate activation data from 2026 PLG deployments will make that answer clear. But the trajectory is unmistakable: the SaaS products that own activation in 2027 and beyond will not be the ones with the most polished product tours. They will be the ones whose AI assistant closes the gap between user intent and first value in the fewest interactions possible.
Takeaway: Frigade Skills is the first commercially available implementation of a principle that will define the next decade of SaaS activation: the AI does the work, the user experiences the value. The 37% median activation rate is not a product design problem — it is a model problem. Show-and-tell onboarding solves the knowledge gap. Action-complete onboarding solves the execution gap. The execution gap is where activation goes to die, and the products that eliminate it will compound a structural retention advantage that compounds every month they lead.
Frequently Asked Questions
What is Frigade Skills and how does it work?
Frigade Skills, launched July 14, 2026, is an in-app AI assistant layer that product teams can add to their software without writing individual action handlers. Unlike a chatbot that replies with step-by-step instructions, Frigade Skills performs the action inside the product on the user's behalf. When a user asks for 'last week's revenue report,' the assistant retrieves and formats it. When a user asks to 'set up my first campaign,' the assistant completes the setup flow. The system learns a product's capabilities autonomously — mapping what the product can do into a set of callable tools — and keeps those tools current as the product evolves. For enterprise customers, Frigade can run self-hosted with the customer's own LLM keys, which means no product data leaves their environment. The platform is SOC 2 Type II certified and GDPR compliant with EU data residency available. Skills builds on Frigade's broader Assist product, which also handles traditional onboarding elements like checklists, product tours, and announcements.
How does action-complete onboarding differ from product tours or guided tours?
Product tours and guided tours operate on the 'show and tell' model: they highlight UI elements, explain what each feature does, and walk users through a scripted sequence of steps. The user still has to execute each step themselves. Action-complete onboarding inverts this: the AI assistant executes the steps on the user's behalf when asked. The cognitive load difference is significant. Traditional onboarding requires the user to understand what to do, remember where the relevant UI control is, and perform the correct sequence of interactions. Action-complete onboarding requires only that the user know what outcome they want. For activation purposes, this matters because the gap between 'knowing how to use a product' and 'using it well enough to reach first value' is almost entirely about execution friction, not knowledge. Product tours were solving the wrong problem. Action-complete onboarding solves the right one.
What activation rate improvement can products expect from AI-assisted onboarding?
There is no published benchmark for action-complete onboarding specifically, because the category is new. But the relevant comparison is to the existing gap between median and top-quartile activation rates: median B2B SaaS activation sits at 37% as of 2026 (per the Perspective AI SaaS Activation Report), while top-quartile companies reach 68% or higher. The gap — roughly 31 percentage points — represents the population of users who get through onboarding with knowledge but fail to execute first value. Action-complete AI addresses this execution gap directly. Early adopters of AI-assisted onboarding flows (pre-Frigade Skills, using bespoke implementations) have reported first-value task completion rates 40-60% higher than traditional guided flows for similar user cohorts. Frigade does not yet publish activation benchmark data from Skills customers, but the mechanism — removing execution friction entirely — suggests gains concentrated in the 'understands but doesn't do' user segment, which tends to be the largest driver of the gap between median and top-quartile rates.
Is Frigade Skills safe for enterprise products with sensitive data?
Frigade designed Skills with enterprise data security as a first-class requirement. The self-hosted deployment option lets enterprise customers run the assistant against their own LLM keys, which means product data is processed within the customer's own infrastructure and never transmitted to Frigade's servers or external LLM providers. This is the architecture required for SOC 2 Type II compliance and satisfies most enterprise data governance policies. The platform also supports GDPR compliance and offers EU data residency for European deployments. For enterprise procurement teams, the relevant questions are: Is user action data (what the assistant was asked to do, and what it did) retained, and for how long? What audit trail is available for compliance review? Frigade's SOC 2 Type II certification means these questions have documented answers that can be produced during vendor evaluation. The self-hosted option resolves the LLM data transit question entirely.
Which product types benefit most from action-complete activation?
Action-complete activation produces the highest lift in products where first-value tasks are well-defined, repeatable, and execution-intensive rather than judgment-intensive. Data and analytics tools (where setup involves connecting data sources, configuring views, and building initial reports) are the strongest candidates — the user knows what outcome they want but the path to that outcome involves many sequential product interactions. CRM and sales tools (where activation typically requires data import, pipeline configuration, and initial deal setup) are similar. Developer tools and APIs are weaker candidates because the first-value task often requires product-specific judgment that the user must supply. Highly customizable platforms (where 'first value' means something different for every customer) present a harder challenge for action-complete automation because the AI cannot know which of many possible starting states is right for a given user. The sweet spot is products with 3-7 well-defined activation paths that map cleanly to common user intent statements.
How do you measure activation success when AI completes tasks for users?
Standard activation metrics (time-to-first-value, activation rate by cohort, feature adoption breadth) still apply but require one critical segmentation: human-initiated vs AI-assisted actions must be tracked separately. An activation event completed by the AI assistant rather than the user is a weaker signal of genuine product understanding and habit formation. The metric that matters most for action-complete activation is reactivation rate: after the AI completes the first-value task for the user, does the user return to the product under their own initiative and perform subsequent actions? High reactivation rates indicate that AI-assisted activation is creating genuine product engagement. Low reactivation rates indicate that the AI is completing tasks without the user developing the product familiarity needed for retention. Some teams track 'human follow-through rate' — the percentage of AI-completed tasks where the user subsequently interacts with the output. This is a leading indicator of whether AI-assisted onboarding is building real retention or just inflating activation metrics.