Context layer [What It Changes]

TL;DR

• A context layer is a live map of how staff actually move work across apps, built from observation rather than from SOPs, so AI agents can follow real exception paths instead of the happy-path manual.

• As of August 12, 2026, Skan AI closed a $63 million Series C to productize that map for enterprise AI agents, mainly in banking, insurance, and healthcare, taking lifetime capital to about $120 million.

• The operational pain is already visible in small shops: a two-truck HVAC dispatcher, a ten-person agency, and a solo clinic all lose hours to the same gap — the work that happens between systems never shows up in the log the agent was trained on.

• Treat vendor dollar claims as identified opportunity until a named process shows a cost-per-transaction change you can audit; the bank case Skan cites is the clearest numbered example in the public record.

Key Takeaways

• Official process docs and backend logs capture committed transactions; they miss the copy-paste, rework, and judgment that sit between those commits.

• Screen-level observation plus aggregation is the mechanism Skan is selling: watch how work moves, abstract intent, then ground agents in that model rather than in a training manual.

• Insurance is already in the blast radius. Skan's press note says the company works with three of the five largest U.S. insurers, and Allianz's COO described quote work that takes tied agents more than half an hour even when the calculator step is three to five minutes.

• Privacy is the binding constraint. Watching screens is useful and contested; Reuters reported in June 2026 that Meta scaled back an internal mouse-and-keystroke collector after staff pushback.

• For an independent agency, the practical move is not a Fortune 50 desktop sensor. Map the human path between Applied Epic, carrier portals, email, and spreadsheets, then plug that map into an existing workflow instead of rebuilding the AMS.

What a context layer is, and why a small shop should care

A Context layer is a living map of how people actually move work across applications, built from watching real cases rather than from manuals, so AI agents can be grounded in operations instead of guessing from SOPs.

A two-truck HVAC shop already lives this problem. The dispatch board says “diagnose, quote, schedule,” but the real job hops from a voicemail to a parts spreadsheet to a manufacturer PDF to a text thread with the helper on the truck. An agent trained on the board will book the wrong return trip the first time a part is backordered.

A ten-person marketing agency hits the same wall when a client asks for “last quarter’s paid-search numbers.” The SOP says pull the report from the ad platform. The actual path is a screenshot, a cleaned CSV, a Slack clarification, and a slide that finance will accept. The agent that only sees the ad platform invents a number that no one will put in front of a client.

A solo-run clinic feels it at intake. The EHR has a clean new-patient form. The real intake is an insurance card photo, a eligibility portal, a sticky note about a prior authorization, and a callback after lunch. If you drop an agent on the form alone, you get a polite chatbot that still leaves the front desk doing the work.

Insurance agencies and carriers scale that same gap. Claims, endorsements, and FNOL packets already bounce across an AMS, a carrier portal, email, and a spreadsheet that never made it into the procedure binder. That is the gap this term is trying to name, and it is why a shop that will never buy a Fortune 50 observation stack still needs the idea.

What shipped on August 12, 2026

According to VentureBeat, Skan AI raised $63 million in Series C funding co-led by Cathay Innovation and Dell Technologies Capital on August 12, 2026. Skan AI closed a $63 million Series C. Citi Ventures, Bloomberg Beta, State Farm Ventures, and Wipro Ventures also participated, and the same report puts lifetime funding at roughly $120 million for the seven-year-old company.

SiliconANGLE dated the same round August 12, 2026, and noted that Dell Technologies Capital also led Skan’s $40 million Series B in March 2022. That earlier story said Skan had raised a total of $54 million since launching in 2019, which lines up with a later ~$120 million lifetime figure once the $63 million Series C is added.

The company did not only raise money. Skan’s Series C page and its August 12 press note say the round lands with general availability of Skan AI Blueprint and Skan AI Agents, sitting beside the existing Skan AI Intelligence product. Blueprint is the satellite view: where AI can create value and what to agentify. Intelligence turns screen-level observation into process flows, exception paths, and bottlenecks. Agents then run against that observed model, with human review, guardrails, and scoped system access defined in configuration.

Skan’s CEO, Avinash Misra, put the pitch in one line in the press note: everyone is obsessed with building a better car; the bigger opportunity is a better navigation system. VentureBeat recorded a close variant of the same line in an interview ahead of the announcement.

Why agents keep failing without that map

According to VentureBeat, Gartner research cited by the company finds that only 8% of enterprises have AI agents in production, and 95% of early implementations will require a complete redesign. Only 8% of enterprises have production AI agents. Those figures sit in the same paragraph as an MIT report from last year, covered by Fortune, which found that roughly 95% of enterprise generative AI pilots were failing to deliver measurable returns.

According to Fortune, the 95% failure rate for enterprise AI solutions is the clearest line in MIT NANDA’s The GenAI Divide: State of AI in Business 2025. That coverage, dated August 18, 2025, says about 5% of AI pilot programs achieve rapid revenue acceleration, based on 150 interviews, a survey of 350 employees, and 300 public AI deployments. The same piece says purchasing tools from specialized vendors succeeds about 67% of the time, while internal builds succeed only about one-third as often. MIT NANDA is the group behind that research program.

The diagnosis Skan is selling is not “the model is dumb.” Skan’s Context Graph of Work essay, dated March 5, 2026, splits the failure into three gaps: the process gap (the SOP says five steps and the floor runs twelve), the decision-trace gap (the agent has no precedent for the exception a senior adjuster would handle from memory), and the integration gap (the work that lives between systems). In one large health-insurer deployment described in that essay, agents switched between applications an average of 37 times per call.

Misra’s version of the same point, in the VentureBeat interview, is that backend data is a committed state of work, and “eighty percent of what you’re interested in, from an AI point of view, in execution of work, actually lies between those systems.” That 80% figure is his claim, not an independent census, but it is the mechanism in plain language: the agent can see the closed claim in the system of record and still have no idea how the adjuster got there.

How the mechanism works, without the brochure fog

Skan deploys observation on employee desktops. SiliconANGLE describes the capture as screenshots processed on the machine itself; the images never leave, and what goes back is anonymized metadata — which applications were used, in what order, and where decisions got made. Skan’s security page says customers whitelist applications, data is encrypted in transit with TLS 1.2+ and at rest with AES-256, and the company lists SOC 2 Type II, ISO 27001, and ISO/IEC 42001 among its certifications. ISO/IEC 42001:2023 is the first international AI management-system standard, published in December 2023.

Skan’s privacy essay, dated August 24, 2026, says the unit of analysis is the process, not the person: sensitive content is masked before storage, identities are pseudonymized, and only non-identifying aggregated information leaves the company network. That architecture is a direct answer to the objection Misra says he heard from AXA Mexico’s then chief transformation officer before he wrote a line of code: this sounds like surveillance.

The objection is not hypothetical. According to Reuters, Meta in June 2026 scaled back an internal tool that collected employee mouse movements, clicks, and keystrokes for AI training after weeks of staff pushback, adding a pause of up to 30 minutes and an exemption path. A context layer that cannot survive that conversation will not survive a European works council or a U.S. insurance HR review either.

What Skan says it does with the observation is the hard part. Watching screens is copyable; abstracting intent is not, in Misra’s telling. The agent product page describes a four-step loop: observe with a lightweight sensor, model thousands of cases into an Agent Operating Procedure, deploy with human-in-the-loop controls, then run on a durable workflow engine where every action is traceable. Listed insurance use cases include claims adjudication, policy underwriting, compliance, and claims FNOL.

That is a different job from classic process mining. Celonis sells a Context Model built from end-to-end process data, business knowledge, and intelligence, and says more than 1,400 companies use the platform. UiPath Process Mining reconstructs processes from the digital footprints left in business applications. IBM Process Mining connects ERP, CRM, and other core systems into a process digital twin. Those tools see what systems recorded. Skan’s argument, restated in its context-graph essay, is that system logs never see the twelve browser tabs, the email that resolved the ambiguity, or the spreadsheet that bridged a missing API.

The same argument is aimed at platform agents. Salesforce Agentforce says over 18,000 companies already run on the product. ServiceNow says 85% of the Fortune 500 work on its platform and positions an AI Control Tower that senses data, decides with context, and acts across workflows. Microsoft Copilot sits inside Word, Excel, Outlook, and Teams. Misra’s critique, in VentureBeat, is that those agents’ context ends at their own walls, while receiving an email and deciding whether a customer entry has to be created is a process that spans them.

Who already runs it, with numbers attached

According to SiliconANGLE, the platform has logged upward of 25 billion work signals. The same story says customers include a quarter of the Fortune 50, seven of the ten largest U.S. banks, and three of the five largest U.S. insurers, with revenue growth of more than 300% year over year and net dollar retention averaging 150%. Skan’s homepage repeats the 1-in-4 Fortune 50, 7-in-10 top U.S. banks, and 3-in-5 top U.S. insurers footprint, plus a 30%+ average productivity lift, $500 million delivered in value, and “6 weeks or less” from observations to agent.

Skan’s press note names the numbered bank case in full: 11.2 million context switches across 1,500 finance professionals, $37 million in operational friction, cost per transaction down 32%, throughput up 41%, and $18 million in annualized savings. Observed 11.2 million context switches across 1,500 staff. VentureBeat carries the same bank figures and adds two more company claims: an AML operation where 60% of cases are now run by AI agents, and a typical ~25% productivity uplift in core claims among insurers, with one customer doubling case volume without adding a claims specialist.

Publicly named customers in that VentureBeat piece include Unum, described there as a $13.8 billion employee-benefits provider, and Mitie, the U.K. facilities group. Unum’s own site says it protects 38 million+ people worldwide. Mitie’s homepage lists £5.6 billion in FY26 revenue, up 10.5%, and 3,000+ customers. Mitie CTO and digital officer Cijo Joseph is quoted in Skan’s press note as saying the technology gave Mitie operational visibility that accelerated its AI transformation.

On the insurance side, Skan’s Allianz recap of a March 2026 on-stage conversation with Jan Malmendier, COO of Allianz Versicherung AG, says Allianz handles more than 10 million customer interactions a year and currently operates at 4.75 stars across touchpoints. Malmendier’s example is the one small agencies will recognize: calculating a quote takes three, four, or five minutes, but tied agents spend more than half an hour, and that extra work is nowhere in the rule book.

Simon Wu of Cathay Innovation — which lists Skan AI among its San Francisco portfolio companies — is quoted in both SiliconANGLE and the Skan press note arguing that enterprise work context is becoming foundational infrastructure the way CRM became the system of record for customer relationships. State Farm Ventures executive Kate Strubhar, in the same press note, frames the layer as a way to connect the information behind how work gets done. Skan also says the platform is powered by NVIDIA AI Enterprise and NVIDIA NIM microservices, with Aser Blanco, NVIDIA’s global head of banking, quoted on observing thousands of real cases and running agents on infrastructure the institution owns.

The $500 million “customer value” number needs the caveat VentureBeat already printed. Pressed on whether it is realized savings or projections, Misra said it is an envelope of quantified savings customers expect to recoup through process redesign, technology changes, and agents — identified opportunity, some portion of which has been captured. Do not treat that figure as cash in the bank.

Sources: SiliconANGLE Series B; SiliconANGLE Series C; VentureBeat.

Sources: Skan AI press note, August 12, 2026; SiliconANGLE.

Sources: VentureBeat; Fortune / MIT NANDA; SiliconANGLE.

Source: NAIC Artificial Intelligence topic page, last updated April 3, 2026, summarizing Big Data and Artificial Intelligence (H) Working Group surveys.

USTA analysis

USTA analysis: using only figures already cited above, the named U.S. bank case converts $18 million of annualized savings against $37 million of identified operational friction, which is 18 ÷ 37 ≈ 48.6% of identified friction captured in the annualized-savings line. Split across the 1,500 finance professionals Skan says it observed, $18 million ÷ 1,500 = $12,000 of annualized savings per observed person. On the capital side, $63 million of Series C against ~$120 million lifetime capital is 63 ÷ 120 = 52.5% of all disclosed funding arriving in this round. Combining the two operating deltas from the same bank case — cost per transaction at 68% of the prior level (a 32% cut) and throughput at 141% of the prior level — implies output per dollar of remaining transaction cost of 1.41 ÷ 0.68 ≈ 2.07× versus the pre-agent baseline. Those ratios are arithmetic on Skan’s disclosed figures, not an independent audit, and they do not transfer to a ten-person agency until that shop has its own friction number and its own captured-savings number.

What this changes for insurance work you already run

Independent agencies do not need a 1,500-person desktop sensor to use the idea. They need an honest map of the path a CSR actually walks between Applied Epic — which Applied says is used by 7 of 10 top agencies on the Business Insurance Top 100 list — and the carrier portal, the email thread, and the spreadsheet that holds the exception. That comparison is the same one we already walk in Applied Epic versus Salesforce Financial Services Cloud: the system of record is not the whole job.

Teams already routing FNOL packets and ACORD PDFs through US Tech Automations workflows can treat a context layer as a model input, not a rebuild: keep the document extract and the routing, then swap the guessed SOP for observed exception paths. The same pattern shows up in the state of insurance automation work we have already published: the bottleneck is rarely the next chatbot; it is the undocumented variant that dumps the file back on a person.

Retention marketing inside an agency has the same shape. A nurture sequence that assumes every book-roll follows the playbook will miss the six extra screens a producer actually hits before the account is safe, which is why agency retention workflow comparisons keep coming back to handoffs rather than slogans. A midsize shop that already moves endorsements in US Tech Automations between the AMS and a carrier portal still needs that map of the human work between those systems, or the agent will quote the happy path and stall on the first surplus-lines exception.

Regulators are not waiting for the vendor category to settle. According to the NAIC, out of the 193 auto insurers responding, 88% reported they use, plan to use, or plan to explore AI/ML models. The same page says the NAIC Model Bulletin on the Use of Artificial Intelligence by Insurers was adopted in December 2023, and that as of March 2026 an AI Systems Evaluation Tool was being piloted by 12 states. NAIC’s own homepage is the current front door for that work. On the federal side, NIST’s AI Risk Management Framework remains the voluntary map-measure-manage backbone; the AI RMF 1.0 PDF is dated January 2023 and frames AI systems as socio-technical, which is another way of saying the model plus the people plus the context of use.

Honest limits

A context layer that faithfully encodes how work is done will also encode shortcuts, rework, and non-compliant paths. Misra’s answer, in VentureBeat, is that the model should learn the full distribution of paths and then be constrained along the axes the firm cares about — speed, cost, or compliance — because the longest path may be the most compliant one. That is a design choice, not a guarantee. If you constrain on throughput alone, you can automate the wrong habit.

Surveillance risk does not disappear because the vendor aggregates. The same VentureBeat interview notes that the technology has led some customers to reduce headcount in certain processes. Aggregation can still surface an underperforming team. Works-council approval, opt-in scoping, and on-prem processing are the mitigations Skan cites; they are also the procurement gates that will keep this out of many small agencies for years.

Platform lock-in is the other limit. If the context lives only inside one vendor’s graph, you have replaced a missing SOP with a new system of record you do not own. Skan’s counter is that the context stays behind the enterprise firewall and that a three-tier architecture sends only anonymized metadata to the cloud. Ask for the data-export path before you treat the graph as infrastructure.

Cost and fit are the SMB limits. A product aimed at seven of the ten largest U.S. banks is not a $49/month add-on for a two-producer shop. The transferable piece is the method: pick one painful process, write down the real clicks, time the between-system work, and only then decide whether an agent, a workflow, or a better SOP is the fix.

Signal vs Speculation

Demonstrated fact (sourced): as of August 12, 2026, Skan AI raised $63 million Series C and announced Blueprint plus Agents as generally available, with lifetime capital around $120 million, a disclosed bank case of 11.2 million context switches / $18 million annualized savings, and named customers including Unum and Mitie. Gartner figures cited by the company put production-agent adoption at 8% with 95% of early implementations needing redesign. Fortune’s coverage of MIT NANDA puts GenAI-pilot failure near 95%. NAIC surveys show most responding auto and health insurers already use or plan to explore AI/ML. Meta’s June 2026 rollback shows employee-observation tools can stall on privacy even inside AI-forward firms.

Our read: if the 8% production-agent figure holds, the next 12–36 months for small and mid-size insurance shops will not be “buy a context-graph appliance.” It will be a split. Carriers and large MGAs will trial observation stacks in claims and underwriting, because that is where the 25% productivity claim, if it survives audit, pays for itself. Independent agencies will get the idea second-hand, through AMS vendors and workflow tools that start asking for exception paths instead of happy-path SOPs. The shops that win are the ones that already know their real click path for FNOL, endorsement, and renewal, and can drop that path into an agentic workflow without a consulting army. The shops that lose are the ones that paste a chatbot on the portal and call the undocumented half-hour of quote work “user error.”

Our read, continued: the CRM analogy from Cathay’s Simon Wu is the speculative piece, not the fundraise. CRM became a system of record because every seller had to log the customer. A context layer becomes infrastructure only if every operator’s between-system work is cheaper to observe than to interview. That is true in a 1,500-person finance factory. It is not yet true in a five-desk agency. Expect the term to show up in AMS roadmaps and in RFP language long before it shows up as a line item on a small-shop budget. If you already run document routing on US Tech Automations, the near-term move is to treat observed exception paths as just another input to the workflow you already own.

Frequently asked questions

What is a context layer in plain English?

A context layer is a live map of how work actually moves across the apps your people already use, so an AI agent can follow real exceptions instead of the version in the manual. It is built from observation of cases, not from a workshop that produces a swimlane nobody follows on Monday.

How is a context layer different from process mining?

Process mining, as UiPath, Celonis, and IBM describe it, reconstructs flows from system event logs and digital footprints inside applications. A context layer, in Skan’s telling, also watches the human work between those committed states — the copy-paste, the extra tabs, the email that resolved the ambiguity — and turns that into a model an agent can execute.

Do small insurance agencies need this, or is it a Fortune 50 toy?

The product Skan is staffing is a Fortune 50 toy. The problem is not. If your CSRs spend a hidden half hour around a five-minute quote, as Allianz’s COO described, you already have a context problem. Map one process on paper first; buy software only when that map is too big to maintain by hand.

Is watching employee screens legal or even wise?

It is contested. Skan says it aggregates, masks, and pseudonymizes, and that European works councils have approved deployments. Reuters documented Meta scaling back a mouse-and-keystroke collector in June 2026 after staff pushback. In insurance, pair any observation plan with the NAIC AI bulletin and with NIST AI RMF controls for privacy-enhanced, accountable systems.

What did the $63 million actually fund?

Skan’s press note says the money accompanies general availability of Blueprint and Agents, and SiliconANGLE says the plan is product work plus harder selling into financial services, insurance, healthcare, and technology. It is not a completed nationwide rollout; it is fuel for a category the investors want to own.

How should a claims team start without a desktop sensor?

Pick one FNOL or endorsement path. Time the between-system work. Write the real clicks, including the spreadsheet. Compare that path to the SOP. Then decide whether the next step is a better checklist, a document-extract workflow, or an agent grounded in those exceptions. Ground that agent in a workflow you control rather than in a one-off chatbot.

Glossary

• Context layer: a living map of how work actually runs across applications, used to ground AI agents in real operations instead of in manuals.

• Context graph of work: Skan’s name for a property graph of people, apps, tasks, decisions, and policies, populated from desktop observation.

• Process mining: reconstructing workflows from event logs left in ERP, CRM, and similar systems.

• Task mining: desktop-level capture of clicks and screens, often stitched back onto a system-log backbone.

• Agent Operating Procedure: Skan’s structured model of steps, decisions, rules, and exception paths an agent should follow.

• FNOL: first notice of loss, the intake path that starts a claim and usually spans phone, email, portal, and AMS.

• Net dollar retention: expansion minus churn on existing accounts; Skan cites about 150%.

• AI RMF: NIST’s voluntary Artificial Intelligence Risk Management Framework (Govern, Map, Measure, Manage).

A context layer will not make a bad claims process good, and it will not replace the judgment a senior adjuster brings to a contested file. What it changes is the input: agents stop being asked to invent the path from a binder that was last updated after the previous AMS conversion.

If you already know the real clicks, the next step is to put them in a workflow you can audit. See how agentic workflows take a mapped process into production, or start from the US Tech Automations homepage and pick the document-routing path you already run. Get the map on paper first. Then decide what the agent is allowed to touch.