Which Companies Develop Custom AI Agents for Insurance in 2026

Insurance runs on paperwork, and paperwork is what AI agents handle best: forms to read, fields to check, records to update, cases to route. Carriers are moving these agents out of pilots and into claims, underwriting, and policy servicing. Gartner predicts that 40% of enterprise apps will feature task-specific AI agents by the end of 2026, up from less than 5% in 2025.

The companies building these agents for insurance today split into two groups: custom builders such as Folio3 AI, Accenture, Neurons Lab, Intellias, Lasting Dynamics, and Comarch, and platforms such as CoverGo, Shift Technology, Sixfold, and Gradient AI.

For US insurers, the open question is which companies develop custom AI agents for insurance that hold up under state regulators, legacy core systems, and messy claim files. This guide covers six custom builders, the four platforms they compete with, and a scorecard for choosing between them. For the short version, see the Insurance AI Agent page.

Companies that develop custom AI agents for insurance at a glance

The companies building custom AI agents for US insurers in 2026 fall into two camps. Custom builders design agents around your workflows and integrate them into your policy administration and claims systems. Platform vendors ship prebuilt agents you configure. Both have a place, and most insurers end up with some mix of the two.

Custom builders worth shortlisting:

• Folio3 AI
• Accenture
• Neurons Lab
• Intellias
• Lasting Dynamics
• Comarch

Platforms to compare against:

• CoverGo
• Shift Technology
• Sixfold
• Gradient AI

What counts as a custom insurance AI agent

An AI agent is software that can assess a situation, reason about it, and take action within the permissions and guardrails it has been given, without a person driving each step. In insurance, that usually means reading a claim submission, checking it against the policy, pulling data from three or four other systems, and deciding what happens next. "Custom" matters because insurance workflows are rarely generic enough to run on a template.

Custom build vs configurable platform

A configurable platform gives you a working agent on day one and asks you to fit your process to its model. A custom build starts from your process and constructs the agent around it. Platforms are faster to start. The trade-off is covered in more detail in the AI implementation cost explained. Tell me if you want that swapped into the final blog above, or send the correct URL if the original post has moved. 

AI agent vs chatbot vs RPA

A chatbot answers questions. RPA follows a fixed script across screens. An AI agent decides what to do based on the situation in front of it, then acts across systems. The difference matters because insurance leadership sometimes buys a chatbot and expects it to adjudicate claims. It will not.

Why insurance workflows need custom logic

Two carriers writing the same line of business will have different rating factors, different claims triage rules, and different reinsurance treaties. An agent tuned to one carrier's decisions will get the other carrier's wrong. That mismatch is a common reason generic insurance AI demos stumble once a carrier tests them against its own rating and triage rules.

Where off-the-shelf platforms stop

Off-the-shelf platforms handle the common cases well. They break on the long tail: unusual endorsements, state-specific compliance quirks, brokered submissions in odd formats, and anything tied to your specific reinsurance arrangements. That long tail is where custom work pays back.

Why insurance AI agent projects stall before production

Most insurance AI projects do not fail because the model was bad. They fail on data, integration, governance, and ownership, and all four are decided long before anyone judges the model's accuracy. Each of these failure points is predictable, which means each one can be planned for.

Pilots built on clean demo data

A pilot that runs on a curated dataset shows what the agent can do in ideal conditions. Production looks different:

• Scanned forms arrive at odd angles.
• Broker emails bury the key details in attachments.
• Loss descriptions contradict the adjuster's notes.
• Policy records carry decades of legacy coding.

An agent that scored well on a hand-picked sample of claims can stumble in its first week of real FNOL traffic. Insurers who skip this messy-data stage often end up rebuilding extraction logic and retesting from scratch. The fix is simple but unglamorous: pilot on a random, unfiltered sample of recent production data, ugly cases included, and measure accuracy there.

Integration left until the end

Policy administration systems like Guidewire, Duck Creek, and Majesco are not casual integrations. When integration starts after the model is built, teams often find the data they assumed was available is missing, arrives in overnight batches instead of real time, or sits behind a vendor API that needs its own licensing and approval.

Write access is harder still. An agent that can read a claim but cannot update it just adds a re-keying step for the adjuster, and the efficiency gain disappears. Integration mapping belongs in week one, alongside the workflow design, with read and write paths confirmed before the build begins.

Compliance added after the build

The NAIC Model Bulletin expects governance across the full AI lifecycle, from design through retirement. Retrofitting is hard because the requirements shape the architecture itself:

• Decision logs that capture each step of the agent's reasoning
• Explanations a policyholder can understand
• Human review points for consequential decisions
• Testing for unfairly discriminatory outcomes

Adding these to an agent that was never designed for them often means rebuilding core parts of it, such as decision logging and review gates. Bringing compliance into the first design review costs far less than a rebuild triggered by the first regulator question.

No owner once the agent goes live

Agents drift. Claim patterns shift, a product gets refiled, a core system upgrade changes a field format, or a state adopts new AI guidance. Without a named owner, performance degrades quietly. Accuracy slips a little each month until an adjuster, an auditor, or a regulator notices something wrong. Production agents need the same operating discipline as any other critical system:

• A business owner accountable for outcomes
• Monitoring for drift and error rates
• A retraining schedule
• A clear process for rolling back a bad update

Decide who owns the agent after go-live before the contract is signed, not after the first incident.

Where insurers deploy AI agents today

These are the use cases where agents are doing real work in insurance operations today.

Claims intake and FNOL agents

These agents take a first notice of loss from any channel (phone, app, email, broker portal) and structure it into the claims system. They extract key facts, classify the claim, and route it. The good ones handle photos and free text without human help on straightforward losses, and send anything unusual to an adjuster.

Underwriting and submission triage agents

Commercial submissions arrive as PDFs, spreadsheets, and email chains. Triage agents read them, normalize the data, check appetite, and hand the underwriter a workable file. Platform vendors like Sixfold have built entire products around this step, which tells you where carriers are spending.

Fraud detection and SIU referral agents

These agents watch claims in real time, score them against known fraud patterns, and build a referral package for the SIU when the score crosses a threshold. The Coalition Against Insurance Fraud estimates that fraud costs Americans at least $308.6 billion a year. The business case is not subtle.

Policy servicing and customer support agents

Servicing is where most policyholder contact happens, and most of it is routine. A servicing agent answers billing questions, reissues ID cards and certificates of insurance, confirms coverage, and processes simple endorsements such as adding a driver or changing a mailing address. What separates it from a chatbot is write access. The agent reads the live policy record and writes the change back to the policy administration system, instead of opening a ticket for someone else to finish.

A well-built servicing agent typically handles:

• Billing and payments: Explains a premium change, confirms payment status, and sets up installment reminders.
• Policy documents: Issues ID cards, certificates of insurance, and declarations pages from the current policy record.
• Coverage questions: Answers from the actual policy wording, with the relevant clause cited, rather than from a generic FAQ.
• Simple endorsements: Collects the details, checks them against underwriting rules, and applies changes that fall inside pre-set authority limits.

Requests outside the agent's authority, like coverage disputes or complaints, go to a licensed representative with the full context attached, and every coverage answer comes from the policy wording itself.

Distribution, quoting, and onboarding agents

Distribution runs on paperwork that nobody enjoys, and everybody depends on. Before a producer can sell, they need to be appointed with the carrier and licensed in the right states. Before a quote converts, someone has to chase missing information and follow up. Agents take on the repetitive parts of each step.

• Producer onboarding and appointments: Collect appointment documents, check license status against state records such as the NIPR database, and flag expiring licenses before they become a compliance issue.
• Quote intake: Pre-fill applications from submitted documents and third-party data, so prospects answer fewer questions.
• Quote follow-up: Track open quotes, send reminders, and alert the producer when a prospect goes quiet or asks something that needs a person.
• Broker submissions: Read emailed submissions, check them against appetite, and tell the broker quickly what is missing or out of appetite.

The cost comes from volume across producers and open quotes, so agents flag licensing and appetite gaps while a person makes the final call.

Subrogation and recovery agents

Subrogation is money the insurer is owed but often never collects. After paying a claim, the carrier can recover costs from the responsible party, whether that is another driver's insurer, a product manufacturer, or a contractor. The recovery signals sit buried in adjuster notes, police reports, and photos, and adjusters working through a heavy caseload do not always catch them before the file closes.

A subrogation agent reviews open and recently closed claim files for those signals:

• Liability indicators: Details in notes and reports that point to a third party, such as another vehicle running a red light or a failed appliance part.
• Third-party details: Other carriers, policy numbers, and contact information already sitting in the file.
• Filing deadlines: Statutes of limitation and inter-company arbitration windows, so recoverable claims do not expire unnoticed.
• Evidence packaging: Photos, statements, and payment records assembled into a file the recovery team can act on immediately.

The agent recommends, and the recovery team decides. Running the review at claim intake as well as closure catches opportunities before evidence is lost.

Regulations that shape insurance AI agent design

Insurance is regulated state by state in the US and by separate regimes overseas. If a vendor does not raise regulation in the first meeting, treat that as a red flag.

NAIC Model Bulletin on AI in insurance

The NAIC Model Bulletin on the Use of AI Systems by Insurers, adopted in December 2023, has since been adopted by roughly half of US states as of September 2026. Because adoption is ongoing, confirm the current count before you cite a specific number to a client. The bulletin expects insurers to maintain a written AI program covering governance, risk controls, testing for model drift and unfair discrimination, and oversight of third-party AI vendors. If your vendor cannot map their delivery to this bulletin, keep looking.

Colorado's rules on algorithms and external consumer data

Colorado went further with Regulation 10-1-1, which requires insurers using external consumer data or algorithms to test for unfairly discriminatory outcomes and keep board-level oversight. The regulation started with life insurers in 2023. The Colorado Division of Insurance has since amended it to also cover private passenger auto and health benefit plan insurers, per its notice of adoption.

EU AI Act high-risk scope for life and health insurance

The EU AI Act classifies AI systems used for risk assessment and pricing of individuals in life and health insurance as high-risk. Under the 2026 Digital Omnibus, the compliance deadline for these Annex III obligations moved from 2 August 2026 to 2 December 2027, but the classification itself has not changed: if you write those lines in Europe, your agents will need documentation, human oversight, and ongoing monitoring by that date.

Solvency II and model risk governance

For European insurers, Solvency II's governance and risk management requirements extend to the models they rely on, AI models included. That means the same documentation, validation, and board accountability expected of actuarial models.

For a practical starting point on governance work, see AI governance and compliance readiness.

How these companies were evaluated

Each vendor was compared on five criteria that insurance buyers ask about in procurement:

• Insurance domain depth: Have they built for insurers before, and can they name lines of business and use cases?
• Production deployment evidence: Case studies with named carriers, not "a leading insurance client."
• Regulatory and audit readiness: Do they design for the NAIC bulletin from day one?
• Core system integration ability: Do they have working experience with Guidewire, Duck Creek, Majesco, Applied Epic, or your specific PAS?
• Ownership and exit terms: Who owns the model, the training data, and the prompts when the contract ends?

Best AI agent development companies for insurance: custom builders

1. Folio3 AI

Folio3 AI builds custom AI agents for insurance carriers, MGAs, and brokers, with a focus on production delivery rather than proof-of-concept work.

How Folio3 AI structures custom insurance builds

Engagements start with workflow mapping and a data and compliance audit, then move through agent architecture, core-system integration, accuracy testing against historical cases, and a human-in-the-loop pilot before full rollout. Integration with policy, claims, billing, and CRM systems is scoped in week one.

Insurance agents in Folio3 AI's build catalog

Builds cover FNOL and claims processing, underwriting decision support, fraud investigation, policy servicing and endorsements, renewal and churn prevention, KYC and AML compliance, actuarial data preparation, claims status, and subrogation recovery. They span P&C, life and annuities, health, specialty, reinsurance, MGAs, and brokers. Multi-agent builds use an orchestrator that assigns work to specialist agents. More detail is on the Insurance AI Agent page.

Controls, fit, and trade-offs with Folio3 AI

Sensitive policyholder data is tokenized or masked before it reaches the model, agent actions are recorded in write-once (WORM) audit logs, and each agent gets only the system permissions its role needs. Folio3 holds ISO 27001 certification. Best fit for carriers, MGAs, and brokers that want agents built around their own policy logic and systems, and that want to own the result. Not the right choice if you want a productized platform you can switch on next week.

Expert Insight

"The insurers who get real value from custom agents are the ones who treat the build like a claims transformation, not a technology project. That means the ops team owns the workflow, engineering owns the integration, and compliance is in the room from the first design review. When those three do not show up together, the agent ships late and gets pulled back within a year."

Muhammad Nasir

Sr. Project Manager, Folio3 AI

2. Accenture

Accenture runs insurance AI as part of larger consulting and transformation programs, and its scale makes it a common pick for multinational carriers.

Accenture's consulting-led insurance delivery

Engagements typically move from strategy to design, build, and run. That structure suits large insurers whose change management work is as heavy as the technical build.

Underwriting and claims agent work at Accenture

QBE Insurance Group, an Accenture insurance client, is scaling AI underwriting across multiple lines of business. Accenture's underwriting research describes teams of task-specific agents reporting to an orchestrator agent, and its health claims work focuses on pulling information out of legacy systems without replacing them.

Who Accenture suits and what to weigh

Best for large carriers running multi-line programs with executive sponsorship. The trade-off is program overhead and the number of layers between the buyer and the engineers writing the code.

3. Neurons Lab

Neurons Lab is a UK- and Singapore-based AI engineering firm that works only with financial services organizations, insurers included.

Neurons Lab's financial-services-only model

The narrow focus is deliberate. Clients choose between AI training on existing tools and custom-built agents. Neurons Lab also offers reusable components for document processing, claims triage, underwriting support, and fraud detection to shorten build time.

Insurance claims work Neurons Lab has published

For one of the world's largest insurers, Neurons Lab built a tool that reviews medical treatment claims using an LLM and a knowledge graph of clinical guidelines. The tool scores each treatment element and explains the score to the reviewing medical officer. Neurons Lab also built a prior authorization platform with insurtech Treatline, as described in an AWS Partner Network post.

Where Neurons Lab fits and where it may not

A good fit for insurers and financial groups that want a sector specialist. Its published insurance work leans toward health claims, so P&C carriers should ask for references in their own lines.

4. Intellias

Intellias is an engineering services firm working with insurers, reinsurers, and brokers on modernization, claims automation, and underwriting.

Intellias's engineering services for insurers

Its insurance practice covers cloud migration of legacy systems, data modernization, and AI work in underwriting and claims. The delivery model suits insurers that want their own product and engineering leads setting direction.

Agentic delivery Intellias has shown in insurance

For US crop insurer ProAg, Intellias used agentic development to modernize one of its most complex legacy processes in five weeks, against a six-month estimate, and validated the result end to end against the old system. This is Intellias's own published account, not one independently verified by ProAg or a third party. Because crop insurance is regulated, every agent interaction ran on infrastructure ProAg already controlled.

When Intellias is the stronger pick

When AI work is tied to core modernization, and you want steady engineering capacity. The ProAg case is AI-assisted engineering rather than a policyholder-facing agent, so ask for claims or underwriting agent references if that is your use case.

5. Lasting Dynamics

Lasting Dynamics is a custom software and AI firm founded in 2015, headquartered in Naples, Italy, with an office in Las Palmas, Spain.

Lasting Dynamics's custom software practice

It builds custom software, AI solutions, and SaaS platforms, and holds ISO 9001 certification. SectorPunk, a vendor-ranking site rather than an industry analyst, places it second on its 2026 list of AI agent developers for insurance.

Insurance evidence for Lasting Dynamics

No named insurance deployment turned up in the research for this guide. Treat the insurance fit as unproven until the firm shows one.

What to confirm before shortlisting Lasting Dynamics

Ask for named case studies in your line of business, integration examples with your core systems, and a plan for supporting the agent after go-live.

6. Comarch

Comarch is a Kraków-based IT company that sells insurance software and delivers AI work on top of it.

Comarch's platform-plus-services approach

Its main insurance product, Comarch Digital Insurance, is an omnichannel front-office platform for agents, brokers, and customers. Comarch also builds AI models for insurers on their own data and can add AI functions to systems it did not build.

Where AI sits in Comarch's insurance stack

Comarch positions AI as an add-on for specific functions such as forecasting, risk assessment, and claims handling, supported by a separate Comarch AI Framework for machine learning work. Agentic AI is part of the offer rather than the core product.

Comarch's best-fit buyers and limits

Best for insurers already running or adopting Comarch platforms, particularly in Europe, where published clients such as UNIQA are based. Buyers wanting an agent-first build partner should look elsewhere on this list.

Top AI agent solutions for insurance: platforms to compare against

If a platform can do most of what you need, the math often favors the platform. These are the ones worth benchmarking custom proposals against.

7. CoverGo

No-code core insurance platform for health, life, and P&C. CoverGo launched insurance AI agents in February 2026, with document processing, customer support, and quotation agents already in production with insurers and brokers.

8. Shift Technology

Fraud detection and claims platform used by hundreds of insurers. Shift Claims, launched in September 2025, adds agentic AI on top of existing claims and core systems, with every agent action logged in the claim audit trail. If fraud is your primary use case, this is the benchmark to beat.

9. Sixfold

Underwriting platform for P&C and life and health insurers. Its AI Underwriter, launched in June 2026, reads submissions, checks them against each carrier's appetite, and recommends the next action. Sixfold states that one carrier's data is never used to train another carrier's system.

10. Gradient AI

SaaS underwriting and claims prediction built on an industry data lake of tens of millions of policies and claims. Its SAIL platform is used in group health and medical stop loss underwriting.

Core-system compatibility and ownership matrix

This matrix lists only what each company states publicly. Where a cell says "Not publicly stated," ask for the answer in writing.

Platform, custom build, or core-suite AI: choosing the right model

When a platform is the faster route

If your use case is fraud, standard FNOL, or common underwriting triage, and your core system is one the platform already works with, buy the platform. You will usually go live faster than with a custom build.

When a custom build pays back

If your workflows are unusual, your core system is old or heavily modified, or you need to own the intellectual property for competitive or regulatory reasons, custom is worth the extra time and cost.

When AI embedded in a core suite makes sense

If you are already replacing your PAS and the new suite has agent capabilities you find acceptable, use those first. Buying a separate agent to bolt onto a fresh Guidewire install is usually premature. For teams building more than one agent, companies that build multi-agent systems is a useful companion read.

Which partner fits which insurer

Large multinational carriers running multi-line programs usually end up with Accenture or a similar global firm. Carriers, MGAs, and brokers that want agents built into their own core systems, and want to own them, should look at Folio3 AI, Neurons Lab, or Intellias: Neurons Lab for a financial services specialist, Intellias when AI is paired with legacy modernization.

Insurers already on Comarch platforms should start with Comarch. On the platform side, fraud-focused claims teams should benchmark Shift Technology, underwriting teams should look at Sixfold, and group health and stop loss writers should look at Gradient AI. There is no single right answer, and most insurers end up with two or three vendors covering different parts of the value chain.

Capabilities a production insurance agent needs

Beyond the model itself, a production-ready agent needs auditable decision logs, human-in-the-loop escalation paths, versioning that lets you roll back a bad update, monitoring that catches drift, and a documented governance program that maps to the NAIC bulletin. Anything short of that will get pulled by compliance the first time a state examiner asks a hard question.

10-question vendor scorecard for insurance AI procurement

Use this in your next vendor meeting. If they cannot answer, that is your answer.

To test your own AI governance readiness before going to market, use the AI governance maturity model as a self-check.

How Folio3 AI can help

Folio3 AI builds custom AI agents for US insurers who want to own what they buy. Engagements typically start with a two- to four-week proof-of-concept sprint on one workflow (FNOL, underwriting triage, fraud referral, or servicing), followed by a production build and expansion once the first agent is producing measurable results. Integration with Guidewire, Duck Creek, or custom policy systems is scoped from the start, and governance work runs alongside the build. The Insurance AI Agent and AI Agent Development pages cover how engagements are structured.

Conclusion

The vendor list will keep shifting. New firms will enter, some will exit, and platforms will absorb features that are custom builds today. The underlying calculus will not change. Pick a partner who understands your line of business, will integrate honestly with your core systems, and will still be around when the NAIC updates its bulletin again. If you are also evaluating AI agents outside insurance, top custom legal AI agent providers and platforms cover a similar market in adjacent territory.

Frequently asked questions

Which companies develop custom AI agents for insurance?

The main custom builders in 2026 are Accenture, Folio3 AI, Neurons Lab, Intellias, Lasting Dynamics, and Comarch. Platforms like CoverGo, Shift Technology, Sixfold, and Gradient AI compete for the same budgets.

What is the difference between an AI agent and an insurance chatbot?

A chatbot answers questions from a script or a knowledge base. An AI agent decides what to do based on the situation, takes action across systems, and completes multi-step workflows like FNOL intake or submission triage.

How long does a custom insurance AI agent take to build?

A proof-of-concept sprint on one workflow takes two to four weeks, and a production agent with integrations takes two to four months. Rolling agents out across teams and systems takes six to twelve months, with data readiness and integration access driving the timeline.

What drives the cost of custom insurance AI agent development?

Data readiness, integration complexity, regulatory requirements, and the number of use cases in scope. The AI implementation cost guide breaks each driver down.

Can AI agents integrate with Guidewire, Duck Creek, or Applied Epic?

Yes. All three offer integration paths, though the specifics vary by version and by which modules you run. Ask vendors for named integration experience with your exact configuration.

Are AI agents allowed in regulated insurance underwriting?

Yes, subject to the NAIC Model Bulletin in adopting states, state rules like Colorado's Regulation 10-1-1, and the EU AI Act for life and health business in Europe. The agent has to be governed, tested, and auditable.

Should insurers build custom agents or buy a platform?

Buy the platform if your use case is common, and it works cleanly with your systems. Build custom when your workflows are unusual, or you need to own the intellectual property.

Will AI agents replace adjusters and underwriters?

Not in 2026, and probably not soon. Agents take over the routine work, and adjusters and underwriters spend more of their time on complicated cases and final decisions.

What happens when an insurance AI agent gets a decision wrong?

A well-designed agent logs its reasoning, flags low-confidence decisions for human review, and has a rollback path. A badly designed one causes a regulatory problem.

Who owns a custom insurance AI agent after it goes live?

It depends on the contract. With custom builders like Folio3 AI, the client usually owns the model, the prompts, and the integration code, while platform vendors own the platform and the client owns their configuration.

How should insurers measure ROI from AI agents?

Track cycle time on the target workflow, straight-through processing rate, adjuster or underwriter time freed up, and error or leakage rates before and after launch. Skip vanity metrics like "queries handled."